pog-mcp 0.9.2 → 0.9.4

package/skill/reference/measurements.md

@@ -1228,6 +1228,480 @@ N=2000 WIDE_N=1000 npx tsx …   # faster; the bar is fixed, the detectable EFFE
 
 ---
 
+## Is a per-player scene weight a decision? (`appr-role.mts`)
+
+The section above swept engine constants and found nothing that re-prices a
+squad choice — not because the effects were small, but structurally: every
+constant it can reach sits in front of BOTH `ofPoint` and `dfPoint`, so it moves
+the two squads together. Per-player selection weights are the one remaining
+candidate that creates a **difference between two squads** rather than moving
+both sides, and they are the axis the tactical-sheet track would ship.
+
+**The lever measured here does not exist in the engine.** `*_APPR` is a table
+per POSITION, so two forwards in the same squad are drawn with the same weight
+and a manager cannot tell them apart. The probe patches `getPlayer` to multiply
+each candidate's weight by a per-player, per-scene number carried on the player,
+defaulting to 1 — nothing else — and sweeps that number. So the answer is about
+the axis, not about one proposed interface.
+
+### The statistic, and the controls
+
+Rows are in **win-equivalent share**: a win is 1, a draw is 0.5, every match
+counts. Deltas are **paired** — every sheet replays one seed column, so the
+default cell and the swept cell are matched match for match — and the seed
+depends on the squads and the side and never on the sheet.
+
+**Bonferroni over m = 13,314 pre-registered comparisons, two-sided: |z| >= 4.62.**
+That m prices the SEARCHES as well as the tests. A ladder read against the
+FIXED shipped default costs only its rungs, because an argmax over the ladder
+could only ever have surfaced one of the contrasts already counted. A comparison
+whose BOTH sides are data-chosen — "is the optimum interior", which pits a
+selected rung against a selected neighbour, and "which sheet is the best
+response", which pits two selected sheets — is charged K(K−1)/2, the number of
+pairs the search could have surfaced.
+
+The noise floor is the whole run's, not one section's: **the largest standard
+error anywhere in this run is 0.59 points of share**, which puts the
+significance threshold at **2.74 points** and the 80%-power effect at **3.24**.
+A paired standard error depends on the outcome variance and on the covariance
+between the two conditions, so it is not constant across sections merely because
+`N` is — quoting the concentration ladders' own worst cell and extending it to
+every later null was a review finding on this section. So every effect called
+large below is an order of magnitude above the run's own floor, and every "no
+effect" is a claim about a band 2.74 points wide.
+
+| Control | Result |
+| --- | --- |
+| Mirror, both modes (one squad against itself; true z = 0) | z **+0.4** / **−0.5** |
+| POLICY mirror, both modes (same squad AND same sheet on both sides; true z = 0) | z **+1.0** / **+0.8** |
+| Patch identity — the patched engine with no sheet vs the stock engine | 1,000 matches identical, **whole `MatchResult`** |
+| Permutation null — the same ×32 on either of `role-433`'s two byte-identical midfielders | 1,000 matches identical, scoreline and winner |
+| Dead-scene null — ×1024 and ×0 on the offence table's ZP/PP1/FK1/FK2 columns | 1,000 matches identical, **whole `MatchResult`** |
+| Zero-weight null — ×1024 on the keeper, whose offence weight is 0 everywhere | 1,000 matches identical, **whole `MatchResult`** |
+| Paired alignment — the default column, replayed | 10,000 matches, largest per-match difference **0** |
+| Anti-focus calibration — funnelling `stars-433`'s 12-point defender must LOSE | **−11.6pp** (z −28.9) / **−14.0pp** (z −26.6) |
+
+<!-- not a payoff table -->
+
+**Every one of those aborts the run.** The comparison DEPTH in the right-hand
+column is part of the control rather than an implementation detail: where the
+claim is that nothing happened at all, the whole `MatchResult` is compared —
+every event in order, every player stat, every shootout round — because a
+rewrite that changed which player was drawn while landing on the same score
+would pass a scoreline check and leave every table below measuring the rewrite.
+The permutation null is the one exception and deliberately so: there the two
+players ARE interchanged, so the event names and per-player stats move by
+construction and only the outcome can be asserted.
+
+Three are worth naming for what they separate. The **permutation null**
+separates "the weight moved the ball to a better player" from "the weight
+perturbed the RNG": moving the same multiplier between two players who are
+identical to the byte must change nothing, and it does not. The **anti-focus
+calibration** is the directional negative control —
+a harness that reported a gain from concentrating on the squad's worst player
+would be measuring the act of concentrating rather than who is concentrated on,
+and every row below would be that artefact. The **dead-scene null** doubles as a
+re-derivation: the probe enumerates the offensive `getPlayer` call sites FROM
+THE SOURCE, finds `A0 A1 A2 CA1 CA2 PP2` and nothing else, and confirms that a
+multiplier on the four remaining offence columns is inert — the dead table the
+plan describes, checked rather than quoted.
+
+### The concentration curve
+
+`stars-433` is the focal squad because it is the only published archetype whose
+players differ inside a position group — three forwards at 28-29 points and four
+defenders at 12. `role-433` rides along as the other half of the question: its
+three forwards are identical to the byte, so any movement there is cross-GROUP
+reallocation with the within-group term pinned at exactly zero.
+
+| Focal slot (`stars-433`) | ×0 | ×1 | ×1024 | League | Knockout |
+| --- | --- | --- | --- | --- | --- |
+| slot 8 — FW, 29 pts, pass 6 shoot 9 | 33.1 | 36.7 | **68.0** | **+31.3pp** | **+35.8pp** |
+| slot 10 — FW, 28 pts, pass 4 shoot 10 | 33.1 | 36.7 | 64.8 | +28.1pp | +32.3pp |
+| slot 7 — OMF, 26 pts, alone in his group | 37.5 | 36.7 | 29.2 | −7.5pp | −8.8pp |
+| slot 1 — DF, 12 pts (anti-focus control) | 38.0 | 36.7 | 25.1 | −11.6pp | −14.0pp |
+| `role-433` slot 8 — FW, identical to its two peers | 46.9 | 48.1 | 58.9 | +10.8pp | +13.2pp |
+
+> **Measured on** — `stars-433` (4-3-3, `STARS` slot weights, `ATTACKING` mix
+> from `squad-lib.mts`: keeper 19, four defenders at 12, midfield 17/16/26,
+> forwards 29/29/28) against `def-433` (4-3-3, `EVEN` weights, `DEFENSIVE` mix),
+> which plays the shipped weights in every cell. The two squads never change:
+> all 212 points, both formations, both sets of kicker flags and the whole
+> defensive table are byte-identical across every row. The only thing that moves
+> is one number in front of one player's offensive selection weight, in all six
+> reachable scenes at once. The last row swaps the focal squad for `role-433`
+> (`EVEN` weights, `ATTACKING` mix), whose forwards are identical to each other.
+> **Sample** — 10,000 matches per cell, home and away, SHA-256 seeds; 20,668,000
+> matches in the run, about four minutes.
+> **Mode** — both, one column each. League is draws-allowed (`gameFlg 0`),
+> Knockout is winner-guaranteed. The League and Knockout columns are the paired
+> delta from the ×1 cell; the ×0/×1/×1024 columns are League shares.
+> **Effect** — 31.3 percentage points of win-equivalent share in League and 35.8
+> in Knockout, between two squads that are the same eleven players. That is
+> nearly three times the largest thing the engine-constant sweep could move
+> (11.8pp) and larger than any squad-construction axis measured anywhere in this
+> file. Concentrating on the WRONG player is worth −11.6pp, so the axis is
+> double-edged rather than free. Even with the within-group term removed by
+> construction — `role-433`'s identical forwards — pure cross-group
+> reallocation is still worth 10.8pp.
+> **Source** — `appr-role.mts`, section (1).
+
+**The best focal player is not the best finisher.** Slot 8 beats slot 10 by
+3.2pp in League and 3.5pp in Knockout, and slot 10 is the one with `shoot` 10.
+That much is measured. **Why is not** — and it cannot be, on a fixed points
+budget. The two forwards differ in five things at once: `pass` 6 against 4,
+`dribble` 9 against 10, `shoot` 9 against 10, `defense` 5 against 4, and one
+point of total; slot 10 also carries the penalty-kicker flag. Moving a single
+attribute in isolation is impossible when every point spent has to come from
+another attribute of the same player, so no cell in this probe separates them.
+
+The candidate mechanism, offered as a hypothesis rather than a result, is the
+`style = attrs.pass` branch: `playArea2` and `playArea1` both decide
+pass-versus-dribble-versus-long-shot from `pass` alone, so a low-pass carrier
+ends possessions with long shots instead of advancing them. What supports it
+without a cross-player comparison is the scene table below — the same player's
+multiplier reverses sign between the midfield scene and the box scenes — and
+what would settle it is an engine change, not another squad.
+
+### Does the anti-concentration term bite?
+
+`atkAttrSum` is the selected player's POSITION GROUP total minus his own, so the
+better the player selected, the less team-aggregate support stands behind him —
+and `atkRate` varies by scene, so the size of that penalty does too. The
+question the plan asks is whether that produces a non-flat optimum.
+
+It is real, and it is not enough. One point of a selected player's total is
+worth **+1/3** on `ofPoint` directly and costs at most **rate × 0.1 ×
+teamPow/100** of group support. The largest `O_*_RATE` anywhere in the shipped
+tables is 1.00, so the loss coefficient never exceeds 0.100 against a gain
+coefficient of 0.333. The probe prints that arithmetic from the source tables
+and then measures it against a COUNTERFACTUAL engine with the subtraction
+deleted:
+
+| `stars-433` slot 8, all scenes | ×1 | ×8 | ×1024 |
+| --- | --- | --- | --- |
+| shipped, League | 36.7 | 49.2 | 68.0 |
+| subtraction deleted, League | 36.9 | 52.3 | 73.4 |
+| shipped, Knockout | 34.6 | 49.6 | 70.4 |
+| subtraction deleted, Knockout | 35.2 | 52.9 | 75.7 |
+
+> **Measured on** — the `stars-433` vs `def-433` fixture above, played twice:
+> once on the shipped engine and once on an engine identical except that all
+> five `atkAttrSum = <group> - <player>.total` sites — including the home-team
+> DMF path that reads the ball carrier rather than the selected player — are
+> replaced by `atkAttrSum = <group>`. Nothing else differs; the ×1 row is the
+> same squads with no sheet at all, so the gap between the two ×1 cells is what
+> the term is worth to a game nobody has concentrated.
+> **Sample** — 10,000 matches per cell, home and away, SHA-256 seeds.
+> **Mode** — both, two rows each.
+> **Effect** — 5.4 points of share in League (z 13.4) and 5.3 in Knockout
+> (z 11.2) is what the term is worth AT MAXIMUM CONCENTRATION; at the shipped ×1
+> it is worth 0.2 (z 0.5) and 0.7 (z 1.4), both inside the sample's own error.
+> So it takes 5.2 points off a concentration payoff that would otherwise reach
+> 36.5 — about a seventh — and changes nothing at all about who you should
+> concentrate on. **0 of the 50 ladders in the one-opponent sections have a
+> significant interior optimum: 48 put their argmax at a ladder END and the
+> other 2 peak on an interior rung the sample cannot separate from its
+> neighbours,** which is undecided rather than monotone and is counted
+> separately for that reason. Monotonicity is a different question from where
+> the argmax is — a ladder can dip significantly in the middle and still peak at
+> the top — so every adjacent rung is tested too: **134 of the 134 ladders swept
+> anywhere in this run have no significant step against their own direction.**
+> The curve is steeper without the term and the same shape.
+> **Source** — `appr-role.mts`, section (2).
+
+The place where the term is not small is the **lone player in his group**. A
+4-3-3's single OMF gets `omfAttr − his own total` = exactly zero support, while
+a forward in a three-man group carries 4.56 points of it into every A1
+selection. That is most of why slot 7 loses in the table above despite holding
+26 points: funnelling the attack through him costs the whole group term. It is
+also a fact about FORMATION rather than about roles — the same player in a 4-4-2
+would have a partner to be supported by.
+
+### Scene-dependence: the same player, opposite instructions
+
+| `stars-433` sheet on slot 7 (the 26-point OMF) | League | Knockout |
+| --- | --- | --- |
+| default — no sheet | 36.7 | 34.6 |
+| ×32 in A2 only | 39.0 (+2.3pp) | 37.5 (+2.9pp) |
+| ×0 in A0 and A1 only | 38.4 (+1.7pp) | 36.5 (+2.0pp) |
+| **both — ×32 in A2, ×0 in A0/A1** | **41.4 (+4.7pp)** | **40.1 (+5.5pp)** |
+| the same ×32 in EVERY scene | 30.7 (−6.0pp) | 27.6 (−7.0pp) |
+
+> **Measured on** — the `stars-433` vs `def-433` fixture above, with every sheet
+> applied to ONE slot: the 26-point attacking midfielder, pass 8 dribble 8 shoot
+> 7, the only OMF in the formation. Every other slot on both squads is
+> untouched, and the opponent plays the shipped weights throughout.
+> **Sample** — 10,000 matches per cell, home and away, SHA-256 seeds.
+> **Mode** — both, one column each; every row clears the corrected bar (|z| 5.8
+> to 14.4).
+> **Effect** — 10.7 percentage points of share in League and 12.5 in Knockout
+> separate the best and worst way of spending the SAME multiplier on the SAME
+> player: +4.7pp when its sign follows the scene and −6.0pp when it does not.
+> A per-player slider would have produced the losing row. The scene signs are
+> stable across modes: A2 positive, A0/A1/CA1/PP2 negative, CA2 inside the noise.
+> **Source** — `appr-role.mts`, section (3).
+
+This is the strongest evidence for the `style = attrs.pass` hypothesis above,
+and it is stronger for being WITHIN one player: the OMF is the squad's best
+passer, and his multiplier pays in midfield (A2) and costs in the box (A0/A1),
+where the forwards' `shoot` and their group support both beat him. No squad
+attribute changes between those rows — only which scene the weight applies to —
+so the confound that blocks the cross-player comparison is absent here. It is
+still not proof: A0/A1 differ from A2 in what they reward as well as in who
+carries, and separating those needs an engine change. **No scene of the
+forward's own ladder points the other way**, so
+scene-dependence is a property of the player rather than of the axis: slot 8
+gives one distinct argmax in both modes, six positive signs in League and five
+plus one below the bar in Knockout, and not a single negative one.
+
+### Opponent-dependence, and the round robin that decides it
+
+| Sheet, `stars-433` | vs `role` | vs `def-433` | vs `def-532` | vs `shootonly` | vs `passonly` | vs `flat` |
+| --- | --- | --- | --- | --- | --- | --- |
+| default | 37.2 | 36.7 | 34.5 | 42.5 | 36.1 | 47.4 |
+| **carrier** — slot 8 ×32, all scenes | **60.7** | **60.6** | **52.7** | **71.4** | **62.6** | **75.6** |
+| finisher — slot 10 ×32, all scenes | 58.6 | 58.4 | 52.3 | 68.1 | 59.1 | 72.3 |
+| playmaker — slot 7 ×32 A2, ×0 A0/A1 | 42.8 | 41.4 | 37.1 | 50.9 | 44.1 | 55.6 |
+| sheet — playmaker plus slot 10 ×32 in A0/A1 | 54.6 | 51.3 | 44.0 | 67.6 | 57.7 | 71.8 |
+| tier — slots 7-10 ×8, all scenes | 50.7 | 49.3 | 43.8 | 62.2 | 52.7 | 68.2 |
+| spread — the four 12-point defenders ×4 | 28.6 | 30.3 | 30.3 | 30.4 | 26.4 | 33.9 |
+
+> **Measured on** — `stars-433` against six published archetypes built by
+> `squad-lib.mts` — `role-433`, `def-433`, `def-532`, `shootonly-433`,
+> `passonly-433`, `flat-433` — every one of them playing the shipped weights.
+> The focal squad's 212 points never change across the whole grid; only its
+> sheet does. `tier` is the coordinated policy: four players lifted together at
+> a smaller multiplier rather than one player at a large one.
+> **Sample** — 10,000 matches per cell, home and away, SHA-256 seeds; 84 cells
+> across the two modes.
+> **Mode** — League shown; the Knockout grid is in the probe output, where
+> `carrier` runs 54.0 to 78.6 and every margin over `default` is wider. Its
+> ranking matches this one in five of its six rows — the sixth is the `def-532`
+> swap in Effect below.
+> **Effect** — 41.7 percentage points of share separate the best sheet from the
+> worst against the same opponent (`flat-433`: 75.6 vs 33.9), and the ordering
+> of the seven sheets is identical in eleven of the twelve opponent-by-mode
+> cells. The twelfth is `def-532` in Knockout, where `carrier` and `finisher`
+> swap by 0.1 points of share — inside the sample's own error, which is why the
+> band there names both. `carrier` sits in every cell's leader band; `def-532`,
+> the deepest defensive shape in the grid, is the only opponent that pulls
+> anything level with it. The distinct raw argmax count is **2**, and one of the
+> two is never beaten anywhere.
+> **Source** — `appr-role.mts`, section (4).
+
+That grid holds the opponent's sheet at the default. The next one lets the
+opponent choose too, on the SAME squad, so the sheet is the only difference:
+
+| Round robin, `stars-433` both sides | beats significantly (League / KO) | in the unbeaten band |
+| --- | --- | --- |
+| carrier | 5 of 6 / 5 of 6 | yes |
+| sheet | 4 of 6 / 3 of 6 | yes |
+| finisher | 3 of 6 / 3 of 6 | no |
+| tier | 3 of 6 / 3 of 6 | no |
+| playmaker | 2 of 6 / 2 of 6 | no |
+| default | 1 of 6 / 1 of 6 | no |
+| spread | 0 of 6 / 0 of 6 | no |
+
+> **Measured on** — `stars-433` against `stars-433`: the same eleven players,
+> the same 212 points and the same kicker flags on both sides of every fixture,
+> with the sheet as the only difference. Twenty-one fixtures per mode.
+> **Sample** — 10,000 matches per fixture, home and away, SHA-256 seeds.
+> **Mode** — both; the two modes agree row for row.
+> **Effect** — 76.0% of the share for `carrier` against `default` in League and
+> 78.3% in Knockout — a 52-to-57 point win-rate advantage from the weights
+> alone. **Significant 3-cycles: 0 in both modes.** The unbeaten set — the
+> sheets nothing significantly beats — holds `carrier` and `sheet`, 1.4 points
+> of share apart in League and 1.5 in Knockout, under the corrected bar. Read
+> that as a PARTIAL order and not a total one: 3 of the 21 pairs in League and 4
+> in Knockout are inside the bar, so "no cycles and a two-member unbeaten set"
+> leaves some pairs unresolved rather than ranking all seven. What it does
+> settle is the direction: nothing here beats `carrier`, and no sheet outside
+> the unbeaten pair is unbeaten.
+> **Source** — `appr-role.mts`, section (5).
+
+### Crossing the two — the grid the verdict actually rests on
+
+The two grids above are one-dimensional slices: one varies the opponent's SQUAD
+with its sheet pinned to the default, the other varies its SHEET with its squad
+pinned. A best response can be conditional on the product while both slices
+share a leader, so neither settles the roadmap question on its own. Multiplying
+them does:
+
+Each cell is my best-response BAND against that opponent squad playing that
+sheet, and — after the bar — the argmax rung of the full concentration ladder
+swept inside the same cell, so "more is better up to the corner" is checked
+where the best response is rather than only against one fixed opponent.
+
+| My best response \| ladder argmax | vs `default` | vs `carrier` | vs `finisher` | vs `playmaker` | vs `sheet` | vs `tier` | vs `spread` |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| `role-433` | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 |
+| `def-433` | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C/F \| ×1024 | C \| ×1024 | C \| ×1024 |
+| **`def-532`** | **C/F** \| ×1024 | **C/F** \| ×1024 | **C/F** \| ×1024 | **C/F** \| ×1024 | **C/F** \| ×1024 | **C/F** \| ×1024 | **C/F** \| ×1024 |
+| `shootonly-433` | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 |
+| `passonly-433` | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 |
+| `flat-433` | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 | C \| ×1024 |
+
+> **Measured on** — `stars-433` against all six opponent squads, each of them
+> playing each of the seven sheets, with all seven of my own sheets AND the full
+> nine-rung concentration ladder tried against every combination. C is
+> `carrier`, F is `finisher`; a cell names every sheet the sample cannot
+> separate from that cell's leader. A second focal squad runs the same grid:
+> `role-433`, the hard case, since its three forwards are identical to the byte
+> and only cross-group reallocation is left — against the other five squads over
+> a reduced sheet set (`default`, `carrier`, `spread`), 30 further cells.
+> **Sample** — 10,000 matches per cell, home and away, SHA-256 seeds: 114
+> squad-by-sheet-by-mode cells, 14,700,000 matches in this section alone.
+> **Mode** — League shown. The Knockout grid differs in three `def-433` cells
+> and nowhere else: it ties `carrier` with `finisher` against `carrier`,
+> `finisher` and `tier`, where League resolves, and resolves against `sheet`,
+> where League ties. `def-532` is C/F in all fourteen of its cells in both modes.
+> **Effect** — 0 cells out of 114 exclude `carrier` from the best-response band,
+> and the ladder's argmax is the TOP rung in all 84 cells where it was swept: 84
+> ends, 0 interior optima, 0 unresolved, and 84 with no significant step
+> downward at any rung along the way. The distinct raw best-response count is
+> 2, and the three cells where `finisher` edges ahead on the raw number are all
+> `def-532` in Knockout, by 0.1 to 0.2 points of share against a 2.74-point
+> threshold. Every cell of the second focal squad's grid reads C/F, which is the
+> permutation control showing through rather than a conditional answer:
+> `role-433`'s slot 8 and slot 10 are the same player twice, differing only in
+> the penalty-kicker flag. So one sheet is a best response against every
+> squad-and-sheet combination measured, and concentrating harder is better in
+> every one of them.
+> **Source** — `appr-role.mts`, section (6).
+
+### The other half of the table — defensive selection weights
+
+`D_*_APPR` picks who contests each scene, and the same multiplier applies there.
+It is a comparably large lever and it points somewhere surprising.
+
+| Defensive focal slot | ×1 | ×1024 | League | Knockout |
+| --- | --- | --- | --- | --- |
+| `stars-433` slot 8 — FW, defense 5, **29 pts** | 37.2 | **52.0** | **+14.8pp** | **+18.7pp** |
+| `stars-433` slot 1 — DF, defense 5, 12 pts | 37.2 | 37.8 | +0.6pp | +0.9pp |
+| `def-433` slot 2 — DF, **defense 9**, 20 pts | 52.4 | 57.1 | +4.7pp | +5.3pp |
+| `def-433` slot 8 — FW, defense 2, 19 pts | 52.4 | 17.7 | −34.7pp | −35.5pp |
+
+> **Measured on** — `stars-433` and `def-433`, each against `role-433` playing
+> the shipped weights. The multiplier is applied to the DEFENSIVE table in every
+> scene; the keeper's A0 weight is the one entry left alone, and the offensive
+> table is untouched in all four rows. Both focal squads keep their 212 points
+> and their formations across every cell.
+> **Sample** — 10,000 matches per cell, home and away, SHA-256 seeds.
+> **Mode** — both, one column each. As in the concentration table, the ×1 and
+> ×1024 columns are League shares and the League/Knockout columns are the paired
+> delta from that mode's own ×1 cell.
+> **Effect** — 14.8 percentage points of share in League and 18.7 in Knockout
+> for funnelling defensive duty onto `stars-433`'s biggest FORWARD, against
+> +0.6pp for its nominal defender. Every defensive point is
+> `defense + cond/2 + total/3`, so a 29-point forward with defense 5 outscores a
+> 12-point defender with defense 5 by more than five points before any group
+> term. On `def-433`, whose totals are flat, the answer reverts to the defender
+> with the highest `defense` (+4.7pp) and the forward becomes catastrophic
+> (−34.7pp). **0 of the 8 defensive ladders have a significant interior
+> optimum; 7 peak at an end and 1 peaks inside without separating from a
+> neighbour.**
+> **Source** — `appr-role.mts`, section (7).
+
+### Degrees of freedom this probe LEFT free
+
+The keeper section below was wrong three times in one day and the cause was a
+different uncontrolled degree of freedom each time. So, explicitly, and knowing
+that the verdict under this list is a NEGATIVE one — the things left free are
+the ways it could be wrong:
+
+- **`cond`.** Fixed at 5 for all 22 players in every cell, as production does
+  today. **This is the load-bearing one.** Concentration here has no cost
+  because involvement has no cost; an involvement-driven form system is exactly
+  the change that would put a price on the ×1024 column, and it would invalidate
+  every row above rather than shift it.
+- **Growth buffs.** Every squad is tenure 0, so `teamPow` is exactly 100.
+  Production raises effective `total`, which scales BOTH the individual `total/3`
+  and the group term the anti-concentration section measures — so the 5.4pp that
+  section reports is a tenure-0 figure.
+- **Manager response.** The opponent grid is fixed and nobody re-solves a squad
+  against a sheet. A squad built KNOWING it will funnel one player is a
+  different 212-point problem, and it is not measured here: the co-optimisation
+  of points and weights would only make the effect larger, which is not the
+  direction that would rescue the axis.
+- **The sheet space.** Seven hand-built sheets and one nine-rung ladder per
+  slot. That is a sample of a space with eleven slots × six scenes of free
+  parameters, not an argmax over it. "No cycle among these seven" and "one best
+  response over the crossed grid" are both claims about these seven — a
+  conditional best response living outside them would read as an answer here.
+- **The focal squads.** Two, and only one of them (`stars-433`) has players who
+  differ inside a position group at all, so the within-group half of this axis
+  is measured on one roster. The crossed grid runs `role-433` as a second focal
+  squad, but over a reduced opponent-sheet set and with a formation and a set of
+  kicker flags it shares with the first. **Which** player to funnel is already
+  known to be squad-conditional — the defensive table above reverses its answer
+  between `stars-433` and `def-433` — and nothing here says the best focal SLOT
+  is the same on a roster neither squad resembles. What the grid does support is
+  narrower and is what the verdict uses: whether to concentrate at all does not
+  depend on the opponent.
+- **Set-piece slots.** `squad-lib` pins the FK taker to slot 9 and the penalty
+  taker to slot 10. Slot 10 appears above as a focal player, so its rows carry
+  kicker duty as well as its weight — slot 8, the headline row, carries neither.
+- **Formation.** Every focal squad is a 4-3-3. The lone-OMF result is a fact
+  about a group of one, and a 4-4-2 would not reproduce it.
+- **Interactions.** One sheet at a time against the default. Nothing here says
+  what a role sheet does while an engine constant is also moved.
+- **The defensive keeper slot.** Section (7) leaves the keeper's A0 weight at
+  100 in every cell. Whether a manager should be able to keep his keeper OUT of
+  the 1v1 is a question this probe does not ask.
+
+### The verdict, and what it means for tactical sheets
+
+**Detectable: overwhelmingly.** Up to 35.8 points of win-equivalent share on one
+squad whose 212 points never change, and 76.0% head to head against the same
+eleven players on the default sheet. Every other section of this file measures a
+lever by changing the players; this one changes nothing except which of them the
+ball reaches, and moves more than any of them. The engine-constant sweep — the
+only other section that holds a squad fixed — moved 11.8pp at its widest and
+re-priced nothing.
+
+**A decision: no.** Every diagnostic points the same way. 0 of 50 ladders in the
+one-opponent sections have a significant interior optimum — 48 peak at an end,
+the remaining 2 peak inside without separating from a neighbour — 84 of 84
+ladders swept INSIDE a crossed cell peak at the top rung with nothing
+unresolved, and all 134 ladders in the run are free of any significant step
+against their own direction. The anti-concentration term is real but worth about a seventh of the
+payoff, and it changes the slope without changing the answer. The round robin
+finds zero 3-cycles in either mode and an unbeaten set of two sheets 1.4 points
+of share apart. And the crossed grid — every opponent squad against every
+opponent sheet, on two focal squads, 114 cells — never once excludes `carrier`
+from the best-response band. **Concentration is monotone: more is better, up to
+the corner, against every squad and every sheet measured.**
+
+So the axis is not a decision *yet*, and the word doing the work is `cond`. The
+reason concentration is free is that involvement costs nothing — the eleventh
+selection of a player is exactly as good as his first, because `cond` is a
+constant. That is what a form system changes, and it is the one change that
+could put an interior optimum on these ladders. **Roles before form would ship
+the largest monotone answer in the game.**
+
+Two things in the tables above survive that verdict and are worth keeping:
+
+- **Scene shape is where the interesting half lives.** The same multiplier on
+  the same player is worth +4.7pp or −6.0pp depending only on which scenes it
+  applies to, and the best focal player is the best PASSER rather than the best
+  finisher. Whatever a sheet eventually exposes, a single "involvement" slider
+  would throw both away.
+- **The defensive table is a second lever of the same size and nobody has looked
+  at it.** On a squad with uneven totals the best defender to funnel duty to is
+  the biggest FORWARD, because `total/3` is in every defensive point. That is
+  either a feature to expose or a formula to reconsider, and it is decided by
+  the same track.
+
+Re-run after any engine change:
+
+```bash
+npx tsx packages/mcp/skill/reference/probes/appr-role.mts   # ~21M matches, ~4 min
+N=2000 npx tsx …   # ~50s; the bar is fixed, the detectable EFFECT moves
+```
+
+---
+
 ## Note for the maintainers
 
 > **2026-08-19: this note was rewritten twice in one day.** First it said the
@@ -1317,3 +1791,1703 @@ If a trigger fires, re-run the probes above BEFORE touching anything, then
 decide between the options in #604. Whatever is chosen, the engine change and
 both documents move in the SAME change — `SKILL.md` is what agents believe, and
 a stale sentence there makes them exclude squads that are perfectly legal.
+
+
+
+
+## How much depth does rotation need, and what does depth buy? (`depth-cap.mts`)
+
+Squad order is a lever nobody priced. That is the headline; the depth numbers are below it.
+
+Reordering the SAME eleven — no reserves, no purchase — is worth **+5.7pp** on a squad
+whose outfield players differ, and nothing (+0.1…+0.9pp) on three squads whose outfield
+weights are near-uniform. The array is sorted by `slotIndex`, a required field on the team
+payload, so the manager sets it. It is **not** an API-only lever: dropping one card onto
+another is classified as a `swap` in `FormationPitch`, which calls `handleDrop` →
+`swapSlots(prev, df, i)`, and submission serializes that array straight back as
+`slotIndex: i`. **Create-team only, though**: the owner edit view renders the pitch
+`readOnly` (`interactive = !readOnly`) and its own file header says it has no save button,
+so the drop handler wired there never fires. One drag reaches this at team creation; an
+existing team cannot be reordered that way.
+
+**The mechanism is unidentified, and both candidates are back open.**
+
+| candidate | test | result |
+| --- | --- | --- |
+| home-DMF `this.kiten` bug | split ONE reordering by side | both sides large — **not rejected** |
+| zone-press fallback | clone ablation | tight interval, but at fixed POSITION — **withdrawn** |
+
+Seeing both sides move does not clear the home-DMF path. It shows the effect is not
+home-EXCLUSIVE, which is compatible with the defect contributing to the home number while
+something else produces the away one — and permuting changes the whole slot order, leaving
+every index-sensitive path live. Rejecting it needs the defect disabled and the same
+permutations rerun, which is an engine change and out of scope here.
+
+The zone-press story looked settled and was not. Every defensive appearance weight at
+`SCENE.ZP` is `0.0`, so `weightedPick`'s empty-index fallback fires on every zone press and
+hands the job to `players[first non-GK]` — all true, and none of it is the effect. The
+ablation: make the four DFs **three identical clones plus one marked player** of the same
+total but a different split. Position and weight sequences are then identical whatever the
+arrangement, so every non-ZP draw selects the same way, while ZP deterministically takes
+the first slot. Moving the marked player in and out of that slot is worth −0.1…+0.3pp, and the Bonferroni-corrected simultaneous interval across
+the three placements is **[−1.8, +2.0]pp** — inside ±5pp on both sides. A near-zero point
+estimate is not a rejection on its own, and neither is an upper bound alone: three
+placements all at −10pp would still have an upper bound under 5pp. Equivalence needs the
+whole interval.
+
+The first version of that ablation built the probe squad uniform at 19 apiece — the same
+trap this note accuses §0.3 of, reproduced in the experimental design. A uniform squad is
+where ordering effects vanish by construction, so it could not have distinguished anything.
+Rebuilt heterogeneous (totals 16–26), with clones only inside the DF group.
+
+What IS established is narrower: the effect appears only when the **position sequence**
+changes. Swapping two same-position players moves nothing (−0.2…+0.2pp); swapping across
+positions produces the whole effect (−1.5…**+5.6pp**). So what matters is not which player
+sits at an index but which position does — which points at the arrangement of unequal
+weights in `weightedPick`'s cumulative scan, and no further.
+
+Two lessons are worth carrying out of this. **An index-dependent path is invisible in a
+homogeneous squad** — which is why `01-engine-expansion.md` §0.3 demoted this lever as a
+"weak-seed-era ghost" after 30,000 matches across ONE squad family. And **a clean interval
+proves nothing if the dimension you manipulated is not the one that moves**: the clone
+ablation held position fixed while the real permutation changes it, so its tight interval
+rejected zone-press-via-attributes and nothing else. That withdrawal is recorded below.
+
+The search is 120 random reorderings with no dedup, not an exhaustive pass over 11!, so
++5.7pp is what the search **found** — a floor, not a ceiling.
+
+Every number in this section is measured against the same league basket. The first version
+ranked each ordering against the ORIGINAL ordering — the same objective error already fixed
+on the purchase axis, still living in this function. Fixing one and not the other happened
+twice in this probe, with common random numbers as well as with the objective.
+
+### Depth — how many reserves, and what they buy
+
+Pools of core-11 plus k generated reserves; every candidate eleven is CONSTRUCTED (I-a's
+assignment rules) and run through `validateTeam`, because sums alone do not decide
+legality. Reserves come from the production generator (`progressive-free-agent.ts`, sums
+uniform on [16,28]) and are **unminted** — §3-13 makes mutability a function of
+`innateFixed`, and freezing name/position rejects combinations that are legal.
+
+Four cores, because one is not enough to see the shape. The first version of this
+measurement used `EVEN/FLAT` alone and concluded "one sum-matched reserve is enough for
+everybody". That squad is the WORST template by `band-worstcase`, and it sits with both
+scarcity caps empty (0 tens and 0 eights against a 3/5 budget), so every reserve fits.
+Cores that sit AT the caps behave nothing like it.
+
+Two walls decide whether a reserve can ever enter:
+
+| core | totals | outside FA's [16,28] | tens | eights |
+| --- | --- | --- | --- | --- |
+| `flat` | 19–20 | **0** | 0 | 0 |
+| `shaped` | 12–29 | **6 of 11** | 3 | 5 |
+| `gkheavy` | 18–29 | 1 | 3 | 5 |
+| `gkmin` | 10–21 | 1 | 2 | 5 |
+
+Scarcity headroom is the first wall — a core already holding 3 tens and 5 eights blocks
+any reserve carrying one, whatever its sum. The reachable-sum range is the second: the
+generator cannot produce a sum outside [16,28], so six of `shaped`'s eleven players have
+no matchable reserve that could exist.
+
+Reserves needed for a legal alternative eleven in 90% of pools (Wilson 95% lower bound,
+300 pools per N — a point estimate moves the threshold between runs):
+
+| core | random FA sums | sum-matched |
+| --- | --- | --- |
+| `flat` | 7 | **1** |
+| `shaped` | 6 | **5** |
+| `gkheavy` | >7 | **2** |
+| `gkmin` | 6 | **1** |
+
+Sum-matching still beats random sums everywhere. What changed is that the count it buys
+runs from 1 to 5 depending on the core.
+
+### What this axis does NOT measure
+
+Every match here runs at `cond: 5` with no workload carried between fixtures. So this
+measures **how much a new attribute vector improves one fresh XI** — market quality — and
+not **what depth buys by resting tired players**, which is the mechanism H exists to cap.
+Simulating that needs a recovery rate and a neutral point, and those are C-2's to decide.
+**H's maximum stays open.** What follows is a market-quality result, and it is worth having
+on its own.
+
+**Two errors point opposite ways, so this is not an upper bound.** The buyer is fully
+informed — it sees all 30 listings' vectors, while production hides attributes and even
+`total` until discovery stage 4 (`player-visibility-policy.ts`) and charges globally visible
+trials to reveal one, with no purchase reservation. That pushes the estimate up. But the
+buyer also does not search the attainable optimum: enumeration stops at `ENUM_CAP`, at most
+`ALT_CAP` subsets and `CAND_BUDGET` orderings survive, and the published run hit the cap 423
+times. That pushes it down. The number is conditional on both, not a bound in either
+direction [Codex P2, 2026-08-31].
+
+### The grant bounds the market — rotation quality does not follow it
+
+Two corrections turned this section around. The grant sizes were hardcoded
+(`shaped: 5, gkheavy: 2`) while the proposed rule says `headroom < 4 → 6` — the document
+was recommending one policy and measuring another. And the opponent basket held four bare,
+default-order cores, which are far weaker than the post-grant baselines they were meant to
+represent. Grants now come from the classifier, and every basket opponent gets its own
+grant and optimizes its ordering.
+
+30 pools, selection on 30 seeds, validation 1500, reporting 1500 (untouched). `buy 0` does
+not search, so it is structurally 0.0.
+
+| core | headroom | free | baseline vs bare core | rotation quality (within 5pp, lower) | buy 0 | 1 | 2 | 3 |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| `flat` | 8 | 1 | +3.5 | 87% (70%) | 0.0 | **+17.5** | **+23.0** | **+23.9** |
+| `shaped` | 0 | 6 | +27.5 | 77% (59%) | 0.0 | **+6.5** | **+11.1** | **+12.8** |
+| `gkheavy` | 0 | 6 | +7.7 | 57% (39%) | 0.0 | +0.8 | +2.2 | +3.6 |
+| `gkmin` | 1 | 6 | +10.8 | 80% (63%) | 0.0 | 0.0 | +3.1 | +3.8 |
+
+### Vary the grant, hold everything else — the controlled version
+
+An earlier version read the policy table sideways and claimed two things: that the grant
+size bounds the purchase advantage (`gkmin` on six buys nothing, `flat` on one pays
++14.6pp), and that a bigger grant costs rotation quality (`gkheavy` 87% → 57%). **In both
+comparisons something else moved with the grant** — the core in the first, the whole
+experimental setup in the second, since 87% came from an earlier run with a different
+basket and procedure. So the grant was swept with the core, basket, pools and seeds held
+fixed.
+
+Grants are drawn **nested** — one max-length list per pool seed, sliced — so a smaller grant
+is a SUBSET of a larger one, and the evaluation column is shared across grant sizes. Drawing
+fresh reserves per grant would move the reserve identity along with the grant, which is the
+same control failure the sweep exists to fix.
+
+| core | buy-1 @1 | @2 | @4 | @6 | quality @1 | @2 | @4 | @6 |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| `flat` | **+15.4** | +5.8 | +6.0 | **+5.8** | 90% | 70% | 67% | 67% |
+| `shaped` | **+17.0** | **+19.1** | **+15.6** | +5.8 | 17% | 40% | 63% | **73%** |
+| `gkheavy` | +4.6 | +5.2 | +0.0 | +0.0 | 63% | 83% | 80% | 57% |
+| `gkmin` | **+9.4** | +2.5 | +0.0 | +0.0 | 73% | 90% | 77% | 83% |
+
+**The first claim is established on two cores — and not in the direction the point
+estimates suggested.** The table above is point estimates, and reading it across is how the
+earlier version reached "it falls on all four, past the gate on three". That was said without
+a test: the second claim had a paired one attached and the first did not, though the nested
+pools make the tool obvious — subtract the two grants' purchase advantage within each pool.
+
+| core | paired mean Δ (6 − 1) | simultaneous interval | verdict |
+| --- | --- | --- | --- |
+| `flat` | **−8.7pp** | [−14.1, −3.3] | **reduces it** |
+| `gkmin` | **−6.4pp** | [−10.9, −1.9] | **reduces it** |
+| `shaped` | −7.2pp | [−14.9, 0.5] | undecided |
+| `gkheavy` | −2.5pp | [−6.5, 1.5] | undecided |
+
+**The critical value is t.** The standard deviation is ESTIMATED from those same 30 paired
+deltas, so a normal critical value is anti-conservative at that sample size — and `shaped`'s
+upper bound sat at a rounded −0.0, meaning the verdict turned on the choice. The
+Bonferroni-corrected t with 29 degrees of freedom (2.663 against the normal 2.498) carries
+that core back across zero: **two cores established, not three**. The quantile comes from a
+Cornish-Fisher expansion, and the probe checks it against three published t-table entries at
+startup and exits on a mismatch — a wrong approximation would give every interval below a
+quietly wrong width.
+
+`flat` is in that list — the core the earlier version singled out as the one that "responds
+least" because of its headroom of 8. Paired, its reduction is −8.7pp, the LARGEST of the
+four. **The endpoints hide it entirely**: `flat` and `shaped` both land on +5.8, which says
+nothing, because they start from 15.4 and 17.0.
+
+**The two results do not stand on the same cores.** `shaped` is where the rotation-quality
+claim is established (p<0.001 below) and where the purchase-advantage reduction is not;
+`flat` and `gkmin` are the reverse. That split is why the two claims have to be carried
+separately rather than as one "grants fix the market" sentence. What bounds the market is not "how much you may buy" but how much you are
+given.
+
+**But "it falls below the gate" is withdrawn.** A median under 5 is not a gate verdict —
+the verdict this probe uses everywhere else is the fraction of pools above 5pp with a Wilson
+upper bound, and applied here almost every bound covers 50%: `flat` 93/81/76/74%, `shaped`
+88/88/92/76%, `gkheavy` 74/74/**32**/54%, `gkmin` 83/66/54/**43**% across grants
+1/2/4/6 (family = grants × cores = 16). At six reserves the only core established below the
+gate is `gkmin`. Reducing the advantage and pushing it under the gate are two claims, and
+the earlier version put them in one sentence.
+
+Which cores respond changed when the sampling keys were fixed — hashing the ENUMERATION
+INDEX meant each grant size searched a different slice, blurring exactly this.
+
+**The second holds on one core only, and every other direction has now evaporated.** The point estimates
+look erratic, and two versions of this document read them opposite ways without a test. The
+pools are nested, so the same pools' success indicators pair across grant sizes (`b` =
+succeeds only at the small grant, `c` = only at the large one), tested exactly:
+
+| core | b | c | Δ (6 − 1) | corrected p | verdict |
+| --- | --- | --- | --- | --- | --- |
+| `shaped` | 1 | 18 | **+57%p** | **<0.001** | **bigger grant is better** |
+| `flat` | 10 | 3 | −23%p | 0.37 | undecided |
+| `gkheavy` | 8 | 6 | −7%p | 1.00 | undecided |
+| `gkmin` | 3 | 6 | +10%p | 1.00 | undecided |
+
+The test is exact McNemar, Bonferroni-corrected across cores. `flat` looked significant in
+the opposite direction until two things were fixed: a normal approximation excluded zero by
+a hair on ten discordant pairs, and the sampling keys moved with the grant. Exact test plus
+stable keys, and it is gone (b=10, c=3, p=0.37). **Adopting a test is not the same as
+checking it fits the data** — the lesson that point estimates cannot carry a direction had
+just been learned, and the instrument brought in to fix it went unexamined.
+
+`shaped`'s result matches the mechanism cleanly. With zero headroom, a small grant leaves
+its alternatives threadbare (quality 17%), so more grant is what makes rotation usable — 18
+of 30 pools fail at one reserve and succeed at six, and one goes the other way.
+
+So: **for cores without headroom, a larger grant genuinely makes rotation usable, and
+nothing in this data shows it costing anything anywhere.** The "fairness versus experience"
+trade-off two earlier versions of this note described, in opposite directions, is not
+there.
+
+Sampling limit, stated: the enumeration cap (4000) was hit **423 times**, so those calls
+sampled a shuffled DFS prefix rather than the full legal set. **That is the aggregate for
+the whole strength invocation** and cannot be attributed to any one grant size: a single
+counter collects basket construction, every grant's baseline and buy-1 search, and the later
+K sweep. Writing "with six-reserve pools" charged the whole bucket to one experiment
+[Codex P2, 2026-08-31].
+
+### The ordering lever against policy-realistic opponents
+
+| core | best of 120 reorderings | paired upper bound | vs the gate | ZP: best → first slot restored |
+| --- | --- | --- | --- | --- |
+| `flat` | −0.7 | 0.9 | below | lower bound −2.4 — undecidable |
+| `shaped` | **+6.5** | **8.7** | **undecided** | **6.5 → 0.9** (the gain lives in that slot) |
+| `gkheavy` | −0.4 | 1.6 | below | lower bound −2.3 — undecidable |
+| `gkmin` | +1.8 | 3.9 | below | lower bound −0.3 — undecidable |
+
+**That upper bound is now measured; for one version it was assumed.** The standard error was
+hard-coded as `sqrt(2*0.25/n)`, and three cores were classified below the gate on that
+assumption. It is wrong twice over: candidate and baseline share each seed, so the
+observations are PAIRS rather than two independent samples, and `n` counted seeds × 2 for
+home/away when those two share a seed as well — that correlation is exactly why a squad
+against itself scores 0.00. Computing the interval from per-seed paired differences instead
+(weighting ladder and knockout within a seed, since they share it too) narrowed the bounds
+against the assumed version of the same run. The variance pairing removes exceeded the
+inflation from double-counting sides, so the old assumption erred toward permissive. The
+four verdicts are unchanged, but observation carries them now, and the point estimates do
+not move at all — same seeds, same matches.
+
+**"It shrinks against policy opponents" is withdrawn.** That read set an EARLIER run's
+bare-opponent +6.4pp beside this run's policy figure, and the earlier run selected its basket
+on a ladder-only objective — the two are not comparable. Scoring the same permutation against
+the four bare cores inside THIS run gives `flat` −2.5, `shaped` +4.9, `gkheavy` −0.1, `gkmin`
++1.2, against policy figures of −0.7 / +6.5 / −0.4 / +1.8. The direction is the opposite of
+what was written.
+
+It does not license the reverse claim either: the permutation was CHOSEN against the policy
+basket, so doing better there is expected, and the bare column is an out-of-objective
+evaluation. What stands is that moving to policy opponents did not shrink the lever.
+
+Against opponents who also hold their
+grant and optimize their order, `shaped` reorders for **+6.5pp** — and a point estimate near
+the gate settles nothing in either direction, so with a paired upper bound of 8.7 that core
+is **undecided**, not below. The other three are below it (upper bounds 0.9–3.9), and two of
+them have negative point estimates, so reordering is not a gain there at all.
+
+What survives is sharper, and now it is controlled. `shaped`'s gain is **localized to the
+first non-GK slot**:
+
+| manipulation | `shaped` |
+| --- | --- |
+| best reordering, untouched | **+6.5pp** |
+| restore the **first non-GK slot**'s occupant to the baseline's | **+0.9pp** (nine-tenths gone) |
+| leave the first slot alone, swap the other slot with a SAME-POSITION one | **+7.5pp** (intact) |
+
+All three come from **one run on one seed column** — mixing in a value from a small side
+run is how the same quantity ends up with two numbers.
+
+Restoring moves two slots, so on its own it cannot say which one mattered — the second
+slot or the interaction would do just as well. The paired control settles it: leave the
+first slot alone and the gain survives whole.
+
+**The control has to be position-matched.** A first version swapped in an arbitrary non-GK
+slot, which lets the second slot receive a player of a different position than the
+restoration puts there — two different lineups, so a surviving gain could come from that
+choice rather than from the first slot. `getPoint` branches on position, so position is the
+dimension to match; with no same-position partner the localization is left open. In
+`shaped` the restoration puts a FORWARD in that slot, so the control uses a forward slot,
+and the gain holds at +7.5pp. **The localization is decided on one core only**: on the other
+three the ordering gain's own paired lower bound is below zero (−2.4, −2.3, −0.3), so asking
+where it lives is not a question. That door was held by a magic number (`|gain| < 1.5`) even
+after the measured interval existed, and it let `gkmin` through to print "ZP is likely" on a
+1.8pp gain with a 3.9 upper bound. The zone-press fallback selects exactly that
+slot, deterministically. Confirming it still needs the path disabled in the engine, but its
+standing as the leading candidate is now earned rather than asserted.
+
+**Both earlier rejections are withdrawn.** The home-DMF split showed only that the effect is
+not home-EXCLUSIVE. And the clone ablation held POSITION fixed while varying attributes —
+but the real permutation changes the position of that slot, and `getPoint` branches on
+position into different rate tables. **A clean interval proves nothing if the dimension you
+manipulated is not the dimension that moves.** Settling it needs the path disabled in the
+engine and the same permutations replayed.
+
+### Which grant does an arbitrary core need? Headroom, and reachable sums
+
+Four synthetic cores cannot support "the core's state sets the grant" — nothing maps an
+arbitrary legal core to a number. 30 random legal cores (4 formations × 6 mixes × random
+weights, keeping only those whose eleven is the sole legal XI), 600 pools per N, threshold
+Bonferroni-corrected across `cores × N` = 210:
+
+| scarcity headroom = (3 − tens) + (5 − eights) | cores | grant needed (min · median · max) |
+| --- | --- | --- |
+| **0** (caps saturated) | 15 / 30 | 3 · 4 · **6** |
+| 1 | 3 / 30 | 3 · 4 · 5 |
+| 2 | 3 / 30 | 2 · 2 · 3 |
+| 3 | 1 / 30 | 4 · 4 · 4 |
+| 4 · 5 | 3 / 30 | 1 · 1 · 1 |
+| 7 · 8 | 5 / 30 | 1 · 1 · 1 |
+
+All 30 reach the 90% threshold by N ≤ 18, and the need falls as headroom rises — but the
+rule is the CONJUNCTION below, not headroom alone: quoting headroom by itself lets back in
+the very no-rotation case the reachability check exists to prevent.
+
+The rule is `headroom ≥ 4 AND at least two totals inside [16,28] → 1, else 6`. Headroom
+alone is not enough, because it ignores the second wall this same measurement found: a legal
+core of four 29s, five 14s and two 13s has headroom 4 but **no** total the FA generator can
+reach, so no matched reserve exists and one player cannot preserve 212. Writing a finding
+down is not the same as putting it in the rule. On this sample (reachable counts 2–11) the
+rule under-grants **0 of 30** cores; reachable counts of 0–1 are unmeasured and fall to the
+larger grant.
+
+**The boundary is not pinned**: headroom 1, 2 and 3 have 3, 3 and 1 cores, and headroom 3
+(n=1) comes out above headroom 2, so it is not monotone there.
+
+**Half of the sampled cores sit at the caps — inside this sampler.** 15 of the 30 have zero
+headroom, and that frequency must not be read as the league's. The sampler is 4 formations
+crossed with SIX HAND-AUTHORED mix templates and random weights, then deterministically
+repaired until the scarcity caps hold: it excludes most otherwise-legal attribute layouts and
+is structurally prone to manufacturing saturated squads. Establishing the real frequency
+needs production team data or a generator argued to be representative, and neither is here.
+
+What stands is conditional: within this sampler half the cores have zero headroom, and
+`flat` (headroom 8) is the uncommon shape among them. A grant policy must not be justified
+with "half the league looks like this" — but "cores with zero headroom exist and one reserve
+is not enough for them" is fully supported, and that is what the classifier rests on.
+
+### Rotation quality — the other half of the minimum
+
+Existence does not settle whether rotation is usable. The best alternative whose **member
+set actually differs** — same eleven in another order is not rotation, and membership is
+compared by asset id because two generated reserves can share a name — plays the champion.
+**The numbers live in the policy table above**, where grants come from the classifier; an
+earlier version of this section carried a second table built on the pre-classifier grants
+(5/2/1) and reported different results for the same cores.
+
+The verdict is the tail, not the median. An earlier version read the medians (−0.3 to
+−1.7pp) and concluded that resting a player costs a point or two — while the same rows
+carried −19pp at the bottom, and pools where no alternative was found had been dropped from
+the summary entirely.
+
+Dropping the failures to summarize the successes is the third instance of one error shape
+in this probe: rejection-sampling shortfalls, then pools with no alternative, then this.
+
+**The conclusion moved five times, and each time one arm had been modelled weaker than it
+is** — the baseline four times (worst-template core; only alternatives allowed to explore
+ordering, plus dropped no-alternative pools; no free grant; single grant draw with the
+baseline missing from the candidate set), the treatment once (random grants standing in for
+a buyer who chooses). The sixth round found something worse than a weak arm: **the axis was
+measuring a different quantity than the one the unit asked for.** Both failures have the
+same check — ask what each side was not allowed to do, and ask whether the thing you varied
+is the thing the question is about.
+
+Controls, both of which must hold before any row above is read: the core eleven alone
+yields exactly one legal XI (all four cores), and a squad against itself scores exactly
+0.00pp by home/away symmetry (all four cores).
+
+```
+ONLY=existence POOLS=300 N_MAX=18 npx tsx packages/mcp/skill/reference/probes/depth-cap.mts
+ONLY=cores CORE_SAMPLES=30 POOLS=600 N_MAX=18 \
+  npx tsx packages/mcp/skill/reference/probes/depth-cap.mts
+ONLY=strength STRENGTH_POOLS=30 SELECT_SEEDS=30 STRENGTH_SEEDS=1500 ALT_CAP=24 \
+  CAND_BUDGET=48 STRENGTH_NS=0,1,2,3 PERM_TRIES=120 MARKET_SIZE=30 ENUM_CAP=4000 \
+  ZP_SEEDS=20000 VALIDATE_SEEDS=1500 GRANT_SWEEP=1,2,4,6 \
+  npx tsx packages/mcp/skill/reference/probes/depth-cap.mts
+```
+
+The two axes cost two orders of magnitude apart, which is why `ONLY` exists.
+
+## How big can the condition band be? (`band-worstcase.mts`)
+
+§3-2 recommended "start at 0.5, never exceed 1.0" and flagged that those numbers were
+themselves unvalidated: they came from one squad template playing itself, with growth and
+role concentration excluded. This sweep runs the missing comparison, and the
+recommendation does not survive it.
+
+### Method
+
+Six legal squads, all 36 ordered matchups (mirrors included), four growth combinations
+(0/0, max/max, 0/max, max/0), a Δ grid, and both modes. The band is a TEAM quantity, so
+Δ is applied to every player in the treated XI, matching how §2 and `cond-effect.mts`
+measure it.
+
+30,000 seeds, and **each seed's home and away runs are averaged into one observation**
+before any variance is computed — they share an RNG seed, so counting them separately
+understates nothing but invents variance where a control has none. Bonferroni family 1440.
+
+Two controls, kept separate on purpose:
+
+- **Plumbing check** (Δ=0, 288 cells): the treated team is the baseline team, so this is
+  structurally zero. It verifies `prep` and seed alignment, and is NOT a statistical
+  control — calling it one is how a degenerate cell gets mistaken for evidence.
+- **Side-swap accounting** (24 cells): a squad against itself at equal growth must score
+  exactly 0.5 win-equivalent. It does.
+
+**The maximum of many noisy estimates is biased upward**, so the worst cell is read as an
+interval, and a safety claim ("this band stays under the bar") uses the UPPER bound.
+
+```
+SEEDS=30000 npx tsx packages/mcp/skill/reference/probes/band-worstcase.mts   # ~20 min
+```
+
+### Result — in WIN-RATE points, on the exposure-weighted mix
+
+Two conversions have to happen before any number meets the 5pp bar, and getting either
+wrong moves the answer by more than the measurement does.
+
+**Unit.** `outcome` scores a draw as 0.5, so the probe produces win-equivalent SHARE, while
+the gate is written in win-rate advantage points. The canonical conversion lives in
+`constant-sweep.mts` as `WIN_RATE_PER_SHARE` and is exactly 2, always:
+`share − 0.5 = (w − l) / 2n`, so `(w − l) / n` is twice it regardless of how the draws
+fall. Comparing share against a 5pp bar silently applies a 10pp threshold.
+
+**Weighting.** The gate is defined over an exposure-weighted mode mix, not per mode, and
+the mix is the repository's own: `constant-sweep.mts`'s `KO_EXPOSURE` =
+`1.33 / (1.33 + 12 + 3)` = **8.1%** — 1.33 knockout matches against **12 ladder matches
+and the cup's own three group-stage matches**, which are played in regulation mode and so
+belong in the denominator. An earlier version of this entry used 24 ladder matches and
+omitted the group stage, giving 5.2%; the source file's comment warns about exactly that,
+noting it overweights the knockout evidence by about two points. Weighting is applied PER
+CELL, since the worst regulation cell and the worst knockout cell need not be the same
+matchup.
+
+| Δ (band) | regulation worst | knockout worst | **weighted upper** | cells over bar |
+| --- | --- | --- | --- | --- |
+| 0.2 **←ceiling** | +3.61 | +5.48 | **4.18** | 0/144 |
+| 0.3 | +5.36 | +8.06 | **6.07** | 24/144 |
+| 0.4 | +7.19 | +10.81 | **8.01** | 83/144 |
+| 0.5 | +8.88 | +13.47 | **9.81** | 113/144 |
+| 1.0 | +17.50 | +26.55 | **18.84** | 144/144 |
+
+**Only 0.2 clears the gate** (weighted upper 4.18, 0 of 144). 0.3 is already over at 6.07,
+and 1.0 puts all 144 cells over. The weighted slope is ≈17.7 pp per cond point, so the bar
+is crossed near **0.28**; 0.2 is the largest 0.1 grid point beneath it.
+
+**And the verdict barely depends on that weight.** Following what `constant-sweep` does,
+the gate is solved across a range rather than at one point (values are weighted upper
+bounds):
+
+| KO weight | Δ0.2 | Δ0.3 | Δ0.4 | Δ0.5 | Δ1.0 |
+| --- | --- | --- | --- | --- | --- |
+| 5.0% | ✅ 4.18 | ❌ 6.06 | ❌ 8.00 | ❌ 9.78 | ❌ 18.77 |
+| 8.1% **(canonical)** | ✅ 4.18 | ❌ 6.07 | ❌ 8.01 | ❌ 9.81 | ❌ 18.84 |
+| 10.0% | ✅ 4.18 | ❌ 6.08 | ❌ 8.02 | ❌ 9.83 | ❌ 18.89 |
+| 15.0% | ✅ 4.21 | ❌ 6.11 | ❌ 8.10 | ❌ 9.80 | ❌ 19.01 |
+
+0.2 passes and 0.3 fails at every weight from 5% to 15% — so even the wrong exposure figure
+gives the same answer, which is a fact only the sweep can establish.
+
+The plan's "start at 0.5, never exceed 1.0" is therefore about **2.5× the safe value**.
+
+**0.2 is provisional.** The role×condition interaction is unmeasured and can move it either
+way — see below — so do not treat it as a fixed value.
+
+### The document already had this in its own table
+
+§2 measured cond per point by template: `role-433` +5.0/+9.2, `bal-442` +4.4/+9.8,
+**`flat-433` +8.2/+13.5**, `def-532` +1.9/+10.5. §3-2 sized the band from `role-433` —
+0.5 × 9.2 = 4.6 pp, apparently under the bar — while the same table's `flat-433` gives
+0.5 × 13.5 = 6.75 pp.
+
+This sweep reproduces that row independently: 8.18 mirror / 8.75 mixed in regulation,
+13.27 in knockout. A flat squad is worst because its judgments cluster near the engine's
+thresholds, so a small condition difference flips the most outcomes.
+
+Two things are new. Mixed matchups and growth do NOT meaningfully raise the worst case —
+almost every worst cell is the flat mirror at growth 0/0, with one exception where a mixed
+pair beat the mirror by 7%. And the response is close to linear, in WIN-RATE
+points: **≈17.8 per cond point in regulation, ≈27.0 in knockout, a ratio of ≈1.5** — the
+five grid points give 17.5–18.0 and 26.6–27.4 respectively. (An earlier revision quoted
+8.5 and 13.4; those were share values left behind when this entry switched units.)
+
+Growth turned out not to be an amplifier at all. That was one of the three axes §3-2 asked
+for, and the answer is negative.
+
+### What this does not cover
+
+Role concentration — the third axis §3-2 named — is not measured here. It requires patching
+the engine to carry per-player, per-scene weight sheets (`appr-role.mts` does this), which
+is expensive, and rejecting 0.3 and above does not need it.
+
+**Do not read 0.3 as an upper estimate.** #728's +31.3 pp is the STANDALONE effect of role
+concentration, not the role×condition interaction, and a main effect does not fix the sign
+of an interaction. The explanation above cuts the other way if anything: if condition
+matters most where judgments cluster near thresholds, concentrating roles could move those
+judgments AWAY from thresholds and reduce sensitivity. The axis can move 0.3 in either
+direction, and it has to be measured before the value is adopted — 0.3 is what two axes
+give, not the answer for three.
+
+## What a same-sum vector transfer is worth (`attr-vs-cond.mts`)
+
+F-6 lets a registration submission redistribute each player's four attributes **within
+that player's own sum**. Two core players with equal sums can therefore submit each
+other's vectors: every check passes, and the strong build lands on whichever ID the
+manager picks. §3-12 answers with a per-registration cap on how far a vector may
+move from its own previous one, which slows a transfer to `ceil(distance / CAP)`
+registrations rather than blocking it; load is never moved. This probe measures what the transfer would have been worth.
+
+### Method
+
+Mirror match — both sides the same squad, the treatment on one side only, so the
+baseline is exactly 50% and the untreated cell IS the true-zero control. Seeds depend on
+(scenario, target, index) only, never on the arm (CRN). Win-equivalent share. **600,000
+seeds** — that is 600,000 independent clusters, 1,200,000 simulations, since each seed
+is reused for its two mirrored sides before they are averaged. Bonferroni family 72,
+equivalence margin ±0.25 pp.
+
+**The two sides of a seed are averaged into one observation before any variance is
+computed** — they share an RNG seed. Counting them separately is wrong, and the control
+proves it: with identical teams the two scores sum to exactly 1 every seed, so
+seed-level variance is zero, yet that estimator reported ±0.17. After the fix the
+control reads ±0.00.
+
+```
+SEEDS=600000 npx tsx packages/mcp/skill/reference/probes/attr-vs-cond.mts   # the tables below
+npx tsx packages/mcp/skill/reference/probes/attr-vs-cond.mts               # default 3000 — smoke only
+```
+
+### Within a position, a swap moves almost nothing
+
+| same-position swap | Δ(pp) | equivalence CI (±0.25) | equivalent? |
+| --- | --- | --- | --- |
+| DF regulation | -0.01 ±0.11 | [-0.12, +0.10] | yes |
+| DF knockout | +0.02 ±0.15 | [-0.12, +0.17] | yes |
+| DMF regulation | -0.00 ±0.11 | [-0.12, +0.11] | yes |
+| DMF knockout | +0.02 ±0.15 | [-0.13, +0.17] | yes |
+
+A swap leaves the team's attribute multiset unchanged; only the labelling moves.
+
+### Across positions it does — and that is a legal edit
+
+F-6 freezes the four attributes but NOT a core player's position, so equal-sum players
+in different positions may exchange vectors and be deployed in the exchanged roles.
+
+| cross-position swap | Δ(pp) | equivalence CI (±0.25) | equivalent? |
+| --- | --- | --- | --- |
+| DMF knockout | -0.57 ±0.12 | [-0.69, -0.45] | **NO** |
+| DMF regulation | -0.42 ±0.09 | [-0.50, -0.33] | **NO** |
+| FW knockout | -0.03 ±0.12 | [-0.15, +0.08] | yes |
+| FW regulation | +0.00 ±0.09 | [-0.08, +0.09] | yes |
+
+The DMF cells (the probe pairs DMF↔FW here; the FW target pairs FW↔OMF) sit well
+outside the margin, at the same magnitude as one `cond+1`
+in the same cell. "The engine is nearly indifferent to the assignment" holds only
+*within* a position. The sign here is negative because the probe swaps in one
+direction; a launderer picks the favourable one. The magnitude is the point.
+
+This is why §3-12 rejects transfers rather than pricing them: once the swap itself
+moves value, no load-based charge can cancel it — you would have to price the build
+movement too, and that varies by pair and by position combination.
+
+### What laundering buys where pricing WAS plausible
+
+Measured against **the compliant manager's best option** — fielding the strong build
+while tired, not resting it. Getting that baseline wrong makes any remedy look
+ineffective.
+
+| threat/433 | laundering gain | cond+1 | compliant best |
+| --- | --- | --- | --- |
+| DF regulation | +0.80 ±0.13 | +0.93 | +5.26 |
+| DF knockout | +1.07 ±0.18 | +1.32 | +6.23 |
+| DMF regulation | +0.62 ±0.14 | +0.58 | +3.70 |
+| DMF knockout | +0.80 ±0.19 | +0.83 | +4.18 |
+
+The gain (+0.62 to +1.07 pp) tracks the same-cell `cond+1` (+0.58 to +1.32 pp): within a
+position, laundering buys exactly one band of freshness.
+
+### A trap worth keeping
+
+An early **20-seed smoke run** returned **exactly 0.00** for the swap in every cell, with
+intervals of ±18 to ±27 pp. Read as a true zero it ends the investigation. It was an
+inert operation: the squad builders give equal sums equal vectors, so the swap produced
+an identical team. Eight cells reading 0.00 to the decimal while the interval is ±18 is a
+symptom, not a result.
+
+(The ±18 belongs to that 20-seed smoke run and is not in tension with the ±0.17 above,
+which is the 300,000-seed run under the uncorrected estimator. Both figures are real;
+only the run differs.) The probe now skips
+vector-identical pairs and prints what it skipped (35 cells).
+
+The exit code covers the control only. A cell failing equivalence is a **result** here,
+not a harness fault — §3-12's rule applies whether or not the swap moves value.
+
+### What this cannot decide
+
+One squad family, one tired/fresh gap (4.5 vs 5.5), and one cross-position pair. It does
+not measure how often the rejection blocks an edit a manager legitimately wants; §3-12
+bounds the geometry (four registrations to reach L1 distance 2, never closer), but the
+distribution of wanted edits is unmeasured. C-2's stress cases own that.
+
+## What is one point of condition worth? (`cond-effect.mts`)
+
+`cond` is pinned to 5 in every real match today, so this measures a lever that is
+not switched on yet. Task E of #732 needs it: the form-system plan proposes a
+ladder band of 0.5 rising to a 1.0 ceiling, calibrated under mirror matchups with
+growth and roles ignored.
+
+**This section reports the grid and the tests. It deliberately does not convert
+them into calibration guidance** — the confounds below are large enough that any
+such reading would outrun the evidence, and six review rounds of this probe were
+spent withdrawing exactly those readings.
+
+### Method
+
+Rows are **win-equivalent share** for the condition-deficient side (win 1, draw
+0.5, loss 0). `Δcond` is the deficit: the short side plays at `5 − Δ`.
+
+The seed depends on (scenario, direction, index) and NOT on mode, growth or Δ, so
+every cell within a scenario replays the identical fixture and any two form a
+paired contrast. Within-cell figures are per-seed paired differences against that
+cell's own Δ=0, with 95% CIs.
+
+Cross-cell contrasts — growth (each configuration against `0/0`) and mode
+(knockout against league) — are per-seed **difference-in-differences** carrying
+**Bonferroni-adjusted** intervals over the WHOLE reported family (m=90 — 54
+growth plus 36 mode contrasts across all three scenarios — α=0.05, z=3.4524),
+because the interesting cells are selected after inspecting all of them. Counting
+the family per-scenario would leave the published family-wide error near 0.09. Growth additionally
+gets an explicit **±0.5pp equivalence margin**; mode is tested only for direction,
+since the useful question differs.
+
+Controls: a true zero requires byte-identical teams (mirror, symmetric growth,
+Δ=0), and its z uses DECIDED outcomes only — a draw sits on the null mean and adds
+no variance. A failed control aborts the run nonzero rather than printing a
+warning beside normal-looking tables.
+
+Growth uses the imported production `computeBuff`. `MAX_BUFF` is unreachable:
+core players are created at `DEFAULT_POTENTIAL_BP = 5000` (P=0.50) and generated
+free agents draw from [3000, 8500), so for these core-shaped squads the feasible
+ceiling is **+2 total per player**, not the +5 the constant names.
+
+8,000 seeds × 2 directions (16,000 fixtures per cell).
+
+
+### League (gameFlg 0)
+
+| scenario | growth | Δ=0 share | Δ=0.5 | Δ=1.0 | Δ=2.0 |
+| --- | --- | --- | --- | --- | --- |
+| mirror | 0/0 | 50.00 | −2.50 ±0.28 | −5.18 ±0.39 | −10.10 ±0.50 |
+| mirror | max/max | 49.89 | −2.40 ±0.27 | −4.78 ±0.37 | −9.76 ±0.48 |
+| mirror | 0/max | 37.86 | −2.53 ±0.27 | −4.81 ±0.37 | −8.95 ±0.46 |
+| mirror | max/0 | 62.31 | −2.34 ±0.27 | −4.92 ±0.37 | −10.00 ±0.49 |
+| gap:base-short | 0/0 | 70.88 | −3.52 ±0.34 | −7.16 ±0.46 | −14.68 ±0.62 |
+| gap:base-short | max/max | 70.97 | −3.42 ±0.34 | −6.81 ±0.46 | −13.88 ±0.61 |
+| gap:base-short | 0/max | 54.66 | −3.73 ±0.37 | −7.47 ±0.50 | −13.99 ±0.63 |
+| gap:base-short | max/0 | 84.33 | −2.56 ±0.28 | −5.41 ±0.39 | −11.60 ±0.53 |
+| gap:other-short | 0/0 | 29.06 | −3.41 ±0.34 | −6.53 ±0.44 | −11.79 ±0.54 |
+| gap:other-short | max/max | 28.70 | −3.18 ±0.32 | −6.18 ±0.43 | −11.28 ±0.53 |
+| gap:other-short | 0/max | 16.14 | −2.14 ±0.26 | −4.16 ±0.34 | −7.13 ±0.42 |
+| gap:other-short | max/0 | 45.04 | −3.90 ±0.36 | −7.84 ±0.49 | −14.43 ±0.61 |
+
+### Knockout (gameFlg 1)
+
+| scenario | growth | Δ=0 share | Δ=0.5 | Δ=1.0 | Δ=2.0 |
+| --- | --- | --- | --- | --- | --- |
+| mirror | 0/0 | 50.41 | −4.83 ±0.51 | −9.53 ±0.67 | −18.29 ±0.82 |
+| mirror | max/max | 50.39 | −4.77 ±0.52 | −9.56 ±0.67 | −18.89 ±0.81 |
+| mirror | 0/max | 30.34 | −4.32 ±0.47 | −8.36 ±0.61 | −14.98 ±0.71 |
+| mirror | max/0 | 69.75 | −3.96 ±0.48 | −8.32 ±0.64 | −18.06 ±0.80 |
+| gap:base-short | 0/0 | 74.63 | −4.21 ±0.44 | −8.66 ±0.60 | −18.09 ±0.77 |
+| gap:base-short | max/max | 74.67 | −3.72 ±0.44 | −7.96 ±0.60 | −17.29 ±0.77 |
+| gap:base-short | 0/max | 54.77 | −4.60 ±0.49 | −9.34 ±0.66 | −17.69 ±0.80 |
+| gap:base-short | max/0 | 89.04 | −2.45 ±0.34 | −5.59 ±0.47 | −12.52 ±0.63 |
+| gap:other-short | 0/0 | 25.38 | −3.84 ±0.43 | −7.29 ±0.55 | −12.91 ±0.65 |
+| gap:other-short | max/max | 24.61 | −3.59 ±0.42 | −6.94 ±0.54 | −12.33 ±0.65 |
+| gap:other-short | 0/max | 11.14 | −2.17 ±0.31 | −3.82 ±0.39 | −6.33 ±0.46 |
+| gap:other-short | max/0 | 45.27 | −5.30 ±0.49 | −10.42 ±0.65 | −18.50 ±0.79 |
+
+### What the tests support
+
+**Condition matters, at every Δ, in every cell.** All 72 within-cell
+effects are negative, and 72 of 72 have 95% intervals
+that exclude zero. At Δ=1 in the league the
+balanced mirror costs 5.18pp; §3-10's published
+"cond 1점 = +5.0pp" is consistent with that cell on this unit.
+
+**Knockout differs from league.** Of 36 adjusted cross-mode contrasts,
+29 exclude zero and 7 do not; 1 of those 29 points the
+other way (knockout cheaper). The direction is therefore not uniform, and no
+single multiplier describes it.
+
+**Growth: nothing is established either way.** Of 54 adjusted contrasts,
+**0** meet the ±0.5pp equivalence margin and **18** exclude zero. So this
+grid establishes NEITHER direction: growth is detectable in a third of the
+contrasts, yet no contrast is tight enough to call it negligible at a margin that
+could move a 0.5–1.0 band. Both "growth can be ignored" and "growth is bounded by
+X" are unsupported here. An earlier revision reported a large asymmetric effect;
+that was an artefact of an unattainable +4 buff and is withdrawn.
+
+### What this grid CANNOT decide — read before using any of it for task E
+
+- **Role concentration is absent.** §3-10's +31.3pp came from patching the
+  engine's `*_APPR` tables (`appr-role.mts`), a probe-level mechanism not
+  replicated here. Task E is not closed until that axis is run.
+- **Growth is a uniform scalar, not a per-player vector.** Production growth
+  differs per player, so two squads can be equally grown in aggregate while
+  concentrating buffs on different positions. Same question as role
+  concentration, asked about a different quantity.
+- **Strength gaps come from squad construction, not from a varied axis.** The
+  two `gap` scenarios ARE lopsided under symmetric growth (~71/29 in league,
+  ~75/25 in knockout), so strength-gap evidence here does not depend on growth
+  asymmetry — but those baselines are produced by swapping in a different BUILD,
+  not by moving a strength dial while holding everything else fixed. Any pattern
+  across baselines is therefore an association among four hand-picked
+  configurations.
+- **The most extreme baselines are only reachable via asymmetric growth.** In
+  this grid nothing sits far outside ~71/29 without uneven growth, so at those
+  ends the two causes cannot be told apart. Reaching an extreme baseline with
+  symmetric growth needs a build pair with a larger intrinsic gap.
+- **One build pair, one formation pair, and a uniform deficit** — every player on
+  the short side drops by the same Δ. A real fatigued squad is uneven.
+
+
+## What load does a player actually accrue? (`engagement-load.mts` + `c2-team-load.ts`)
+
+**This section does NOT close C-2.** An earlier version said it did; that was wrong. `τ` was
+never measured — the probe's 24h default was simply used — and the neutral point was left
+unchosen between its two branches. An F-1 implementer following that decision would have made
+a probe default the production recovery rate. And `τ` cannot be measured from this data: F-1
+does not exist, so production has no fatigue and there is no observed recovery curve to fit.
+What measurement CAN do is price the choice, which is what the τ sweep does.
+
+**The gate is not a per-mode ceiling — it is defined on the exposure-weighted mix.**
+`band-worstcase.mts` holds the canonical statement (`01-engine-expansion.md:319-329` puts
+~5pp on that mix; `measurements.md` pins "not per mode"), and **E's 0.2 came from that gate**.
+This section had been holding ladder and knockout to 5pp SEPARATELY — calibrating against the
+weighted gate while testing against a stricter per-mode one. With knockout at **8.1%** of
+exposure, that difference reverses the conclusion.
+
+Three things are established:
+
+- **Carrying E's ceiling across exceeds NOTHING at τ=24h.** Each team state is turned into a
+  fresh and a tired variant and both are played against the SAME REAL OPPONENTS on the same
+  seed, with the modes mixed by exposure: at τ=24h, 66 cells give **0 over, 3 established
+  below, 63 undecided**. The short τ exceed badly — **32 of 52** at 6h, **12 of 63** at 12h.
+  Only three established below is the price of the guarantee, not a smaller effect: the cell is
+  now (team × curve × STATE RULE) with six opponents inside it, so the family is **1,836**. **"none over" and "all established below" are different claims**: the first says
+  exceeding cannot be shown, the second says clearing can. τ=24h is the first; getting the
+  second means lowering the amplitude. The earlier round played the tired team only against its
+  own fresh CLONE — a mirror matchup, not what the gate is defined on. **The modes still
+  diverge** — knockout point estimates run about **1.5×**
+  ladder's, which is why holding each mode to 5pp separately showed 38 of 78 over. That is
+  what the earlier "almost every team, in knockout" conclusion was: **the wrong bar**.
+- **So what this section delivers is a safe amplitude**: re-evaluating ALL 66 cells at each
+  candidate under a simultaneous correction and then validating the choice on INDEPENDENT
+  seeds gives **`b` ∈ 0.1988–0.2014** at τ=24h, worst cell 3.86pp. **All four τ values have
+  such an amplitude** (0.16–0.21), so fairness closes on the AMPLITUDE, not on the choice of τ.
+  That is **0.07×** the value derived by putting E's 0.2 on the median player. Testing one cell with a pointwise
+  interval leaves winner's-curse bias: on a small sample a search that passed at 4.11pp
+  validated at 10.48pp.
+- **Whether rotation pays — not at the amplitude that would ship.** At `g=0.4` the prize is
+  0.59pp [0.25, 0.93] and the paired difference is −0.43pp [−0.76, −0.10]: the prize does exceed
+  the cost there. But **that is not where the decision lives.** A rested player can only recover
+  the deficit the shipped amplitude creates, and the **safe amplitude produces a deficit of
+  0.135** (τ=24h, the largest across all validated cells — taking it from the cell with the
+  largest BOUND instead understates what can be recovered). At that deficit the prize is
+  **0.16pp [−0.08, 0.41]** against a cheapest cost of **0.16pp [−0.36, 0.67]**, paired difference
+  **−0.00pp [−0.47, 0.46] — undecided**, and the same at all four τ (the eight repricing calls
+  publish a prize interval and a paired-difference interval apiece, so all **16 intervals** split
+  one 5% budget). The cost of a one-point-weaker slot runs 0.16 to
+  2.21pp depending on WHICH point moves, with selection and publication on SEPARATE seed columns
+(the selection column's own spread across the 16 is 2.57pp).
+  **Which one is cheapest is not established**: measuring all 16 on the publication column too,
+  its minimum is `dribble→shoot` (−0.01pp), not the selection's `dribble→pass` (0.16pp). The
+  comparison keeps the selection's pick — taking the publication column's minimum would bias that
+  column — so the cheapest substitute costs somewhere in **0 to 0.2pp** and which one it is stays
+  open. Every interval here is distribution-free (empirical Bernstein), which is why this section
+  runs at 60,000 seeds.
+
+**The measurement was rebuilt five times and the conclusion moved each time.** A synthetic
+squad grid gave "this axis is too small" (prize 0.46pp). Replaying production matchups gave
+a longer tail and a prize the same order as one ability point. Integrating the production
+TIMELINE gives the numbers above. The first two are withdrawn.
+
+### λ_team × multiplier cannot produce L
+
+That product multiplies two upper quantiles, so it puts `L_top` at 63 — while integrating
+the real timeline puts the peak `L` at **p95 21.3** (observed max 50.5). The two quantiles do
+not arrive on the same player-day. And bucketing by day discards the SPACING, which is
+everything to an EWMA: cup fixtures cluster into a few hours after opening, so eight closely
+spaced calls do not equilibrate like eight spread over a day.
+
+So the layer is gone. `matches.input_snapshot` (#694) freezes the exact two teams handed to
+the engine, so the last 30 days' **1,201 real kickoff matchups** are replayed in time order
+and the decay is integrated over them:
+
+    L ← L·exp(−Δt/τ) + (that match's engagements)
+
+Neither `λ` nor a multiplier appears. Each snapshot carries its own `gameFlg`, so the mode
+mix is right by construction — removing an estimate beats getting it right.
+
+Three further corrections: only **credited** appearances accumulate (a human ladder away side
+and an AI cup entrant carry zero F-1 load, and an earlier version pooled them into the human
+multiplier); each matchup is replayed **five times and averaged**, because a single replay's
+integer counts read within-match noise as between-player heterogeneity; and the synthetic
+grid is kept as a control only.
+
+**τ is therefore not a free parameter.** An earlier version used the equilibrium form and
+concluded τ cancels. On the timeline τ sets the load itself: every table below is τ=24h and
+another value needs another integration.
+
+### Stage 1 — team call frequency, 30 completed UTC days
+
+| humans, 32 teams | median | p90 | p99 | max |
+| --- | --- | --- | --- | --- |
+| **calendar days** (zeros included, 865 team-days) | **0** | 8 | 10 | 21 |
+| active days (zeros excluded, 270 team-days) | 6 | 8 | 19 | 21 |
+| **per-team window total** (32 teams, median 30 observed days) | **0** | **183** | — | **199** |
+
+**A team-day median cannot say how many teams sat out.** An earlier version read the
+calendar median of 0 as "half the human teams played nothing"; a median of 0 over team-days
+only says half the team-DAYS were empty — every team could have played on a minority of its
+days, and late-created teams contribute fewer days besides. Counting each team's total over
+its own observed window instead: **18 of 32 human teams (56.3%) played nothing** in the
+window (bots: 919/1,536 = 59.8%).
+
+**The grid includes each team's creation day — that was removed and then put back.** Removing
+it avoids counting a partial day as a full team-day of zero, but it also **deletes every daily
+fill-bot match**: fill-bots are created the morning their cup opens, play only that day
+(AGENTS.md L712-717 — a deployment with zero humans still runs a 48-bot cup daily), and
+contribute nothing but zeros on the window's other days. Losing the numerator outright is far
+worse than one partial day in the denominator. Putting it back moves the bot cohort a long
+way — active team-days 664 → **1,224**, "never played in the window" 96.3% → **59.8%** — and
+leaves the human quantiles alone, since only 8 human team-days (42 calls) are creation days.
+
+**That count also exposes what the median hid — the distribution is bimodal.** Median 0, p90
+**183**, max 199. Human teams either barely play or play throughout; **there is no typical
+team in between.** So §3-1's neutral point at "today's typical load" has no typical to
+anchor to: wherever it goes, one end sits in the wrong place. On the calendar reading it is
+0, which sets `a` = 0 and puts everyone who plays into penalty — the mirror of the all-bonus
+failure §3-1 warns about; on the active-day reading (6) the 56% who sit out all land on the
+bonus side. Not chosen here.
+
+Buckets are anchored to UTC explicitly: `date_trunc('day', NOW())` follows the session
+timezone, so a non-UTC database silently measures partial days under a "completed UTC" label.
+
+**"Has a snapshot" counts only what the repository would accept.** An earlier version checked
+that `home.team` and `away.team` were JSON objects, which lets `{"team": {}}` or a squad
+without eleven players count toward 100% coverage — while `parseMatchInputSnapshot` throws on
+them (fail-closed, #694), so the exporter either drops the row silently or ships it and
+`playOne` dies. The coverage query now applies the same predicate as `isSideSnapshot`, plus
+the one condition only this consumer needs (exactly eleven players, since the probe indexes
+slots 0–10), and the coverage and export queries share the same SQL fragment — if they
+diverge, coverage says 100% while the export loses rows. A row that passes SQL but fails the
+JS parse is a contradiction, so the script exits rather than dropping it. Tightening changed
+nothing here: the same 1,201 matches, byte-identical export. "Fully covered" is now verified
+rather than assumed.
+
+### F-1's predicate: counting actors only drops half, and concentrates the rest
+
+Production replay, 1,201 matchups over 5 replays, 8,283 credited human player-matches:
+
+| predicate | team total | p95 | median | p95/median | zero-load share |
+| --- | --- | --- | --- | --- | --- |
+| **A.** hyoka credit sites, unsigned | 6.60 | 1.60 | 0.40 | **4.00** | 13% |
+| **B.** actor only | 8.97 | 2.00 | 0.80 | **2.50** | 14% |
+| **C.** actor + responder | **17.88** | 3.20 | 1.60 | **2.00** | **4%** |
+| **D.** C without the ZP fallback | 16.61 | 2.80 | 1.40 | 2.00 | 4% |
+
+The pool is credited HUMAN team-sides only — λ is a human-cohort figure, so the pool has to
+be as well. **This table alone uses expected RATES** (the 5-replay mean per player-match).
+Counting realized integers instead would put a player whose rate is 0.4 into "no load" on
+every draw that came up 0, and that player does accrue load across matches; the question here
+is whether a player is STRUCTURALLY load-free, not how often they were picked in one draw.
+The timeline integration asks the opposite question and uses integers.
+Counting actors only drops 49.8% of the engagement — (17.88−8.97)/17.88 — and what
+remains is MORE concentrated: p95/median 2.50 against 2.00 (4.00 at the hyoka sites), with a
+zero-load share three and a half times higher. Establishing §3-10's premise across the whole eleven needs responders counted.
+
+**The earlier "65% accrue nothing" is withdrawn.** That came from single-replay integer
+counts; on expected rates it is 14%. Within-match noise had been read as between-player
+structure. The argument survives in a different form.
+
+The keeper still accrues almost nothing — it enters a contest only at the A0 one-on-one and
+at penalties — so no rule here makes one tired, and F-1 has to fill that blank deliberately.
+
+Predicate D now subtracts fallbacks on BOTH sides of the contest: zone press picks its ACTOR
+through the fallback too, and an earlier version recorded only the responder's, so "C without
+the ZP fallback" was removing half of what it claimed.
+
+**A and C rose again in this pass** — A from 5.72 to **6.60**, C from 16.54 to **17.88** —
+because two sites were not being counted at all:
+
+- **The goal-scorer's credit does not go through `applyHyoka`.** `scoreGoal`
+  (`engine.ts:1581`) writes `atk.hyoka[i] += 25 + rand(50)` DIRECTLY. Defining A as "the
+  current credit sites" and then counting only `applyHyoka` drops every scorer.
+- **The defender picked after a failed dribble was uncounted.** `engine.ts:653` selects it
+  with `getPlayer(def, scene, false)`, which never passes through `getPoint`, so it sat
+  outside the responder tally — yet that player commits or avoids the foul and can be booked.
+  In the production distribution: **2,894 selections, 13.2% of all defensive selections,
+  7.1% of C**. If C means "defensive selections", this is one.
+
+**B did not move by a single digit** (8.97 → 8.97), which is the control that says both fixes
+touched only what they were meant to.
+
+### The parameters
+
+E capped the ATTAINABLE spread at 0.2, so the derivation fixes that and solves for `b`.
+`L` is measured, not estimated — timeline integration at τ=24h over the 143 credited human players:
+
+143 human players over an **11-day window with 100% snapshot coverage** (2026-08-21 onward,
+1,201 matches), τ=24h.
+
+**"the match that happened" is verified per match, not inferred from a date.** Calling a
+stored-seed replay that only holds if the engine replaying it is the engine that ran it, and
+when that breaks fixture consistency breaks with it — a different knockout winner leaves a
+team that cannot exist in that timeline still accruing load, exactly why the derived-seed
+timelines were dropped.
+
+That boundary was once taken from `git log`'s COMMIT date. A commit date is not a deployment:
+it can ship after the following midnight, be rolled back, or roll out gradually. So the
+inference is gone and the replay is compared against what production STORED instead — reproduce
+it and the match happens the same way under today's engine, otherwise drop it. The result is
+**1,201 of 1,201 (100%)**: the two `engine.ts` changes inside the window (`74af7010`
+golden-goal control flow, `b587f5d4` scene tags) did not change these matches' outcomes. The
+209 matches cut by date were fine.
+
+**The comparison runs on three axes, each closer than the last to what this probe counts:**
+
+| axis | what it covers | this window |
+| --- | --- | --- |
+| final score | coarsest — "same score, different path" passes | 1,201/1,201 |
+| event-sequence digest (`seq|minute|type|player|player2`) | the REPORTED path | 1,201/1,201 |
+| stored **per-player rating** | a projection of the hyoka accumulator | 2,402/2,402 sides · 26,422 player-cells · 0 mismatches |
+
+The second axis alone is not enough: selections that go through `getPoint` and emit NO event
+at all (a completed pass, a successful dribble) still land in the `off`/`def` arrays being
+integrated. The third covers that hole — `player_match_stats.rating` is
+`clamp(round(hyoka/100), 1, 10)`, so it is a projection of the hyoka accumulator, and hyoka is
+moved by exactly those selection sites (a successful dribble calls `applyHyoka(atk, kiten, 1)`).
+
+**That axis carries little information, and that is measured too.** 97.4% of the 26,422 cells
+are `4`, only three values occur, entropy is **0.176 bits per cell**, and the 2,402 eleven-vectors
+take only **57 distinct values**. So "26,422 cells agree" must not be read as 26,422 independent
+tests. What it actually constrains is whether hyoka moved enough to push some player across a
+rounding boundary — a weak constraint, but one independent of the event axis.
+
+**`scene` and `detail` are deliberately excluded from the digest.** Both are presentation
+metadata and both changed inside the window (#725 added scene tags on 08-22); including them
+would reject the 08-21–08-22 matches for a reason unrelated to the estimand. `player` already
+carries the side.
+
+**And the functions that produce the selection path did not change in the window — checkable at
+the source level.** Six commits touched `packages/engine/src` inside it, but only two touched
+simulation (the rest are test files and `validator.ts`): `74af7010` (golden-goal termination)
+and `b587f5d4` (scene tags on events). Neither touches `getPlayer`, `getDefPlayer`,
+`getPoint`, or `weightedPick`:
+
+```
+git log --since=2026-08-14 --oneline origin/main -- packages/engine/src/
+git show b587f5d4 -- packages/engine/src/engine.ts   # all pushEventAt / scoreGoal arguments
+```
+
+**What remains unverified.** Production has NEVER persisted the `off`/`def` tally, so for a
+window that has already passed, "persist the tally and compare it" is not available at all.
+What survives all three axes is a revision that leaves the events AND the ratings identical
+while selecting different players — which requires hyoka credits to cancel exactly. Cutting the
+window by deployment date is not the alternative: a commit date is not a deployment, and that
+inference is precisely what this verification replaced.
+
+**Engagements are integrated as integers.** An earlier version replayed each matchup 5 times,
+AVERAGED the counts, and fed those fractional expectations into the EWMA. That is wrong:
+every published quantity — per-player peak, p95, the gate deficits — is a NONLINEAR
+functional, so `E[f(X)] ≠ f(E[X])`, and averaging erases within-match selection streaks and
+narrows the load vector. A narrower vector passes the gate more easily, so the error runs in
+exactly the direction that says "safe". What runs now is **one realized timeline**, replaying
+each match's stored `matches.rng_seed` (**1,201 of 1,201**) — that pass is not a different draw
+of the same matchup but the match that happened.
+
+A middle version ran five timelines, four of them on derived seeds, and that was wrong too:
+the fixture list comes from production and is fixed, so a derived seed that changes a knockout
+winner leaves a team that **could not exist** in that timeline still accruing load in the next
+round. Those four were inconsistent histories, and they were feeding the published medians and
+the worst-state search. The cost of dropping them is that the across-draw spread is no longer
+measurable here — it needs bracket progression regenerated, which this probe does not do.
+
+Fixing the averaging moved the vectors up: max deficits rose 5–14% per cell and the per-match
+increment rose **31–33%**. The safe amplitude moved about 2% (measured then against the
+per-mode gate, so the absolute values differ from today's) — the conclusion was insensitive to
+the error, but that is only knowable after re-measuring.
+
+**Which instant you read `L` at decides the policy.** Three values differ:
+
+| instant | median | p95 | observed max |
+| --- | --- | --- | --- |
+| **at appearance** (what cond reads, 7,139 reads) | **8.46** | **26.21** | 55.27 |
+| at the window end (an arbitrary clock time) | 7.03 | 16.24 | — |
+| path maximum (after its own match's engagements) | 19.35 | 38.17 | — |
+
+**The load event is not placed at kickoff.** F-1 pins the decay reference to the time the
+match's condition is FIRST SNAPSHOTTED, which is when the scheduler picks the match up —
+closer to `completed_at`, since simulation takes milliseconds. An earlier version substituted
+`kickoff_at` and justified it with "the lag inside the window is under a minute". **That was
+false**: the window's last day has a median lag of **428 minutes** and the overall p99 is
+**508** (35% of τ=24h) — some cups are processed well after their scheduled time.
+
+So it is measured rather than argued. The published `claim` axis uses **1,144** cup
+`completed_at` values plus **57** playoff `finalized_at` values: playoff writes the result as
+`live` first, then that latter column records the completion transition. A per-row
+`claimSource` proves which hand supplied it; if either the completion time or its source is
+missing, the probe refuses the whole window instead of falling back to kickoff [Codex P2,
+2026-09-04]. Integrating on the explicit `kickoff` alternative instead gives median 8.66 and
+p95 26.35 — **2.3%** and **0.5%** away.
+
+**That comparison holds the warm-up boundary fixed.** Left to itself each variant recomputes
+`collectFrom` from its own first event — and switching the time base, or dropping the
+long-lag matches, is exactly what moves that first event. The printed difference would then be
+a clock effect PLUS a changed sample rather than the substitution alone. The published run's
+boundary is passed in so every variant collects from the same absolute instant. Here the
+correction moved the median by 0.01 (8.65 → 8.66): small, but measured rather than assumed.
+
+**Neither completion timestamp is the first claim.** The daemon PRESERVES a failed attempt's
+snapshot, resets the match to `scheduled`, and picks it up later, while cup `completed_at`
+records the attempt that finally succeeded. Playoff `finalized_at` is coarser still: it closes
+the live window roughly two to three minutes after the result was computed. The store has no
+first-snapshot column, so it cannot be recovered; excluding the **104** matches with more than
+60 minutes of lag gives median 8.66 and p95 27.26, **+2.4%** and **+4.0%**. The uncertainty on
+this axis is 2–4%, and F-1 persisting the first snapshot time
+removes it.
+
+**That this axis is a real claim clock is now proved by the ledger, not by lag** [Codex P2,
+2026-09-02]. Two hands write `completed_at`: a live completion stamps the moment it finished,
+and `repairMissingCompletionPositions` stamps the moment it was DISCOVERED. Nothing on the row
+says which. An earlier version separated them by size — "more than 24 hours of lag means
+backfill" — and that is the wrong axis: the repair also runs during rolling deployments (its
+own comment says so), so an old replica's just-finished match discovered minutes later carries
+a small lag and passes that test. Magnitude cannot tell the two hands apart.
+
+For `completed_at`, the `data_backfills` ledger can. `runOnce` returns at the `<name>:closed` marker BEFORE doing
+any work, so once that marker is written the repair never runs again — which makes "closed at
+T" a proof that every `completed_at` at or after T was written by a live completion. In this
+store the marker was written at **2026-08-17T09:56:13Z**. The repair itself ran at 06:37Z that
+day: of the 1,502 completed competitive matches kicking off between 08-03 and 08-16, every one
+that carries a `completed_at` carries it **inside 06:37:21.9–23.5Z**, a 1.6-second window —
+which is the fingerprint the repair's own `base + rn milliseconds` produces. The window opens on
+08-21 and its earliest claim is 08-21T04:00:08Z. The 57 playoff rows use a separate durable
+proof: `finalizePlayoffMatch` writes `status='complete'` and `finalized_at` in the SAME update,
+and the completion backfill never edits that column. The export carries both timestamp and
+source. The probe checks that all 1,144 `completed_at` values postdate the ledger close, that
+`finalized_at` was not assigned to a cup, and that every row has both fields. With no proof it
+does not drop rows, it **refuses to compute**: dropping them would punch a hole in the window
+and bring the same bias back wearing a different face.
+
+**The EWMA is warmed before observations are collected.** Truncating the window (coverage,
+engine boundary) leaves players carrying load they arrived with, while the integration starts
+them at `L=0` — treating the boundary as a league-wide fatigue reset. An earlier version merely
+listed this as unmeasured; measured, it is real: discarding the first **2τ** as warm-up raised
+the at-appearance median from 7.32 to 7.76 (+6%) — **that run's numbers**; the current run
+prints only the warmed value, and with the predicate corrected it is **8.46**. The cold start was biasing load DOWN,
+which makes the gate easier. The collection window is 8.8 days and the residual bias is
+`exp(−2) ≈ 0.135` — recorded, not claimed gone.
+
+Calibration uses **at appearance**. Condition is frozen at claim time (§7), so what a player
+actually carries is the `L` read as they enter a match. The window end is a state at an
+instant, not a state on entry — and **nobody ever experiences the path maximum**, because it
+includes the match's own engagements and no one reads condition after that. Earlier versions
+used each player's own last kickoff (which leaves a player who stopped un-cooled) and then the
+window end (right for "now", wrong for "on entry").
+
+| `L₀/L_top` | `b` | `a` (neutral) | team spread | one-player deficit | prize* | x per match | in pp |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| 1× | 0.820 | 0.200 | 0.200 | 0.410 | 0.602pp | 0.029 | 0.043pp |
+| 2× | 1.439 | 0.200 | 0.200 | 0.480 | 0.705pp | 0.046 | 0.068pp |
+| 4× | 2.679 | 0.200 | 0.200 | 0.536 | 0.788pp | 0.063 | 0.093pp |
+
+\* deficit × slope (1.47 pp/cond); those deficits sit above the largest directly measured one
+(0.40) and the response is concave, so they are biased high. The measured figure is 0.59pp,
+and at the safe amplitude the whole-vector team-level figure is 3.86pp.
+
+Repricing at each safe amplitude uses the largest deficit across all independently validated
+cells, not the deficit of the cell with the largest confidence bound:
+
+| τ | safe `b` | deficit | prize there | cheapest cost | paired cost − prize | verdict |
+| --- | --- | --- | --- | --- | --- | --- |
+| 6h | 0.1633 | 0.112 | 0.13pp [−0.10, 0.37] | 0.16pp | 0.02pp [−0.44, 0.49] | undecided |
+| 12h | 0.1639 | 0.110 | 0.13pp [−0.10, 0.37] | 0.16pp | 0.03pp [−0.44, 0.50] | undecided |
+| **24h** | **0.1988** | **0.135** | **0.16pp [−0.08, 0.41]** | **0.16pp** | **−0.00pp [−0.47, 0.46]** | **undecided** |
+| 48h | 0.2061 | 0.137 | 0.16pp [−0.08, 0.41] | 0.16pp | −0.00pp [−0.47, 0.46] | undecided |
+
+The eight repricing calls publish a prize interval and a paired-difference interval apiece.
+All 16 intervals split one 5% family budget; each two-sided interval splits its share again
+between its two tails [Codex P2, 2026-09-02; 2026-09-04].
+
+`x`, the number §3-1 asked to be stated: one more match costs a top-load player **0.029–0.063
+condition points (0.043–0.093pp)**. No "N matches per day" qualifier is needed — τ and the
+real spacing are already inside `L`. The increment is the p95 of the **realized integer**
+distribution (4 engagements; median 1) — one match is one draw, so the p95 of the expected
+RATE ("a busy player's average match") answers a different question. That distinction alone
+raised these two columns 31–33%.
+
+**The pp column is a CONVERSION, not a measurement.** The 1.47 slope is a SECANT taken at
+`g = 0.4` (that one slot's loss divided by 0.4), and win probability is nonlinear in condition
+and depends on role, attributes and opponent. So the same deficits were measured directly, on
+the same slot and the same published seed column:
+
+| τ | per-match increment (cond) | **measured directly** | secant conversion | ratio |
+| --- | --- | --- | --- | --- |
+| 6h | 0.189 | **0.21pp** [−0.05, 0.47] | 0.278pp | 1.3× |
+| 12h | 0.124 | **0.14pp** [−0.10, 0.38] | 0.182pp | 1.3× |
+| **24h** | **0.063** | **0.07pp** [−0.15, 0.28] | **0.093pp** | **1.3×** |
+| 48h | 0.031 | **0.06pp** [−0.14, 0.26] | 0.046pp | 0.8× |
+
+**None of the four resolves at this budget** — every interval covers zero once the alpha is
+split across the 16 intervals from eight repricing calls — so the pp column has to be read as a conversion
+throughout. Comparing point estimates, the secant runs about 1.3× the direct measurement, but
+that comparison is itself not established. The sweep table's per-match column is the same
+conversion.
+
+### τ cannot be measured — only priced
+
+`τ` is the recovery rate §6 C-2 asks for. F-1 does not exist, so production has no fatigue
+and there is no observed curve to fit: `τ` is a DESIGN choice, not a measurement. What
+measurement can do is re-run everything at each candidate and report what follows. Each row
+below re-integrates the timeline, re-runs that τ's full cell set, and re-searches for the safe
+amplitude (one run, one snapshot).
+
+| τ(h) | `L` med | `L` p95 | published `b` (4×) | per match (pp) | over gate | **safe `b`** | worst at its low end | `b` proven over |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| 6 | 2.68 | 11.80 | 3.716 | 0.278 | **32/52** | 0.1633–0.1669 | 3.64pp | 1.8581 |
+| 12 | 4.19 | 16.18 | 3.291 | 0.182 | **12/63** | 0.1639–0.1671 | 3.42pp | 1.6455 |
+| **24** | **8.46** | **26.21** | **2.679** | **0.093** | **0/66** | **0.1988–0.2014** | **3.86pp** | — |
+| 48 | 17.55 | 45.94 | 2.294 | 0.046 | **0/61** | 0.2061–0.2084 | 3.56pp | — |
+
+Three readings. **`L` is near-linear in τ** (2.68 → 17.55), so §3-6's `L* = λτ` holds on the
+timeline too and changing τ means rebuilding the whole table. **Longer τ makes a match cheaper
+and the field fairer** — marginal cost 0.278 → 0.046pp, cells over **32 → 0** — because a
+longer memory dilutes both the single match and the heterogeneity between players; a short τ
+buys "one match hurts" at the price of unfairness, and here that price is steep. **All four
+have a safe amplitude**, so fairness is closed by the AMPLITUDE, not by τ; τ is the "how much
+does one match hurt" dial.
+
+**And the bisection errs in one direction under noise.** One unlucky evaluation above 5 near
+the threshold means everything above it is never revisited, while an unlucky PASS is caught by
+the independent-seed validation — the asymmetry is structural. This run's τ=24h trace shows the
+scale (4.97pp at b=0.1988, 5.14pp at b=0.2014, 1.3% higher). The published amplitude is
+therefore a LOWER bound on the threshold, and the error runs toward the safe side. And the
+right-hand endpoint is NOT a confidence limit: "not established below" is not "over" — that
+would need a LOWER bound above 5. Across this sweep two candidates cleared that bar and survived an independent-seed confirmation
+(τ=6h at b=1.8581 and τ=12h at b=1.6455), and only those thresholds acquire an upper bound.
+
+**Whether the weight decides the verdict is measured too.** Re-counting the same cells at
+knockout weights of 5%, 8.1% (canonical), 10% and 15%: at τ=24h cells over run **0·0·0·0** —
+the verdict does not hang on that number at all. The canonical weight is used; where the
+dependence grows is recorded rather than assumed away.
+
+The safe `b` is a **bracket**, not a point: the largest candidate that passed and the smallest
+that failed. And that bracket is SEARCH GRANULARITY, not a confidence interval — each
+candidate's bound is itself estimated from 6,000 seeds PER OPPONENT. This run's lower endpoints
+increase (0.163 · 0.164 · 0.199 · 0.206), but the procedure does not test ordering across τ, so
+that shape must not be read as structural. The
+bisection errs in ONE DIRECTION under noise: a single unlucky evaluation above 5 near the
+threshold means everything above it is never revisited, while an unlucky pass is caught by the
+independent-seed validation. τ=24h's trace shows the scale — 4.97pp at b=0.1988 against 5.14pp
+at b=0.2014, 1.3% higher. So the published value is a LOWER BOUND on the threshold, and the
+direction it errs in is the safe one. **The right endpoint is not an upper bound either** —
+"not established below" is not "over". What is established is that every τ in this range has an
+amplitude in the reported safe range that puts all cells under the gate, and that cells over fall 32 → 0
+with τ.
+
+**The sweep caught two defects in the gate** that a single τ could never have shown. First,
+capping a degenerate sample's bound at infinity made cells where the amplitude is small enough
+that the two variants coincide fail the search outright — τ=6, 12 and 48 all reported "no safe
+amplitude" for that reason, and τ=24 only appeared to succeed because it happened to land
+ABOVE that point. The earlier τ=24 result was luck, not structure. A rule-of-three bound
+(`c + (−ln α / n)·(max − c)`) fixes it and all four then resolve. Second, five bisection steps
+were too coarse: τ=6 stalled at 5.35pp on b=0.122 and that was printed as "none exists", which
+is search depth, not a finding. Ten steps now, with an early exit on the first failing cell to
+pay for them.
+
+### Testing the vector against the gate instead of borrowing a ceiling
+
+E applies the same Δ to all eleven; fatigue makes a vector. An earlier version averaged
+engagement by ARRAY INDEX across every formation and applied that to a synthetic 4-3-3 — but
+index is slot order, so the average at slot s mixes unrelated roles and is nobody's load
+vector. Worse, it described the result as "played against its fresh copy" while the code put
+each variant against a rotating synthetic field and subtracted field-relative win rates.
+Match outcomes are nonlinear and opponent-dependent; that difference is not a head-to-head.
+
+Corrected: each team state is turned into a fresh and a tired variant, and BOTH are played
+against EACH OF THE SIX FIXED REAL-OPPONENT BASKET MEMBERS as its own cell — never averaged or
+screened away — on the same seed,
+then differenced. **Each seed runs BOTH home and away; those two
+differences are averaged before they become one observation** [Codex P2, 2026-09-04]. Alternating
+orientation by seed index would require home and away to be identically distributed, which the
+engine's home-only DMF branch disproves. Playing the tired side against its own fresh clone is a mirror matchup, and the
+gate is explicitly defined by VARYING the opponent, because the engine's thresholds make the
+condition response matchup-dependent. The opponent family is every distinct XI actually fielded
+in the window, folded on a hash of the fielded eleven rather than on `teamId` — folding on
+`teamId` keeps only each team's first snapshot and discards every later XI. The model's
+neutral offset `a` is applied to both sides, since the equation is `5 + a − bL/(L+L₀)` and
+omitting `a` overstates the gap by exactly that.
+
+**The opponent is a CELL AXIS — neither averaged nor screened.** The canonical gate
+(`band-worstcase.mts:161`) loops `for (A) for (B) for (growth) for (mode)`, so both the count of
+cells over and the choice of worst live on that grid. This got it wrong twice:
+
+1. **Averaged.** Drawing one pool member per seed and averaging them into a single cell dilutes
+   an effect that appears only against a particular opponent, and a mixture mean is never above
+   the worst, so the error ran toward "safe".
+2. **Then screened.** Testing all 62 and confirming only the winner leaves the discarded 61 with
+   NO guarantee and a family that never counts the opponent axis. Worse, that screen was NOISE.
+   This run measures one term's SD at a median of **24.0pp** (a win/draw/loss difference), so at
+   120 screen seeds an opponent mean has SE 2.19pp and the max−min of 62 draws is **≈9.4pp from
+   noise alone**. The 10.85pp range the previous round read as "the opponent axis matters" was
+   SMALLER than what noise produces.
+
+So screening is gone and the basket is fixed: the 62 fielded XIs sorted by team strength, six
+taken at even quantiles. Deterministic, never touching an outcome, and the family counts the
+whole axis (cell = state × curve, with six opponents inside; family = (78+78+78+72) × 6 =
+**1,836**). Publication and out-of-basket exploration do not divide seeds across the basket — each opponent gets the full 12,000 — because
+a degenerate cell's bound floors at `(−ln(α/2)/n)·SER_MAX`, so dividing `n` by the basket size
+widens that floor by the same factor; dividing it produced "no safe amplitude" at all four τ,
+which is arithmetic rather than a finding.
+
+**Bounding all 62 under 5pp is not affordable** — at that SD a 1.5pp half-width per opponent
+needs ≈4,092 seeds each, tens of times the cell budget across 62. So the basket is a SAMPLE of
+the 62, the same kind of limit the canonical gate accepts by using a designed grid. The question
+it leaves open — does an opponent outside the basket exceed? — is asked separately below.
+
+**What that fold actually changed was measured.** Across 2,402 sides there are 95 `teamId`s but
+only **62 distinct XIs**. The old pool held 95 objects, of which only **47** were distinct
+lineups — 48 were duplicates. And **35 of the 95 teams fielded more than one XI** inside the
+window, every later one discarded. The new pool CONTAINS the old one and adds 15 lineups.
+(Snapshot `cond` is 5 for everyone in this window — F-1 does not exist yet — so "fielded XI"
+is squad identity here, not condition.)
+
+**Taking only the top three was arbitrary — all 13 teams that played are measured.** Two
+things were wrong at once. Teams were ranked by *peak single load* while states were chosen by
+*summed deficit* — two currencies, and putting them on one **changed two of the top three**.
+With the currency fixed, the cut turned out not to exist:
+
+```
+70363723 5.93 · 30178fc1 4.62 · 6097e735 4.45 · 704b8e45 4.34 · 6aec8bd8 4.31
+118f1f13 4.20 · 6a37ed0a 4.20 · 2aa911db 4.12 · 2b865779 4.01 · 23baeba2 3.74
+930b4313 3.58 · 796efbfe 3.34 · e8dc63e1 0.15        (summed deficit, max over curves)
+```
+
+Second (4.62) through twelfth (3.34) sit close together, so "the worst team" would
+have been decided by which three I picked. Nothing is cut now, and the simultaneous correction
+widened accordingly. (τ=24h below; other τ in the sweep above.)
+
+| team | `L₀` | state (UTC) | max deficit | ladder | knockout | **weighted** | lo | hi | 5pp |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| `70363723` | 1× | 2026-08-25 12:58 | 0.524 | 5.00 | 7.86 | **5.24** | 2.59 | 7.88 | undecided |
+| `70363723` | 1× | 2026-08-27 09:57 | 0.556 | 5.10 | 7.33 | **5.28** | 2.60 | 7.95 | undecided |
+| `70363723` | 2× | 2026-08-25 11:55 | 0.684 | 6.30 | 9.09 | **6.52** | 3.72 | 9.33 | undecided |
+| `70363723` | 2× | 2026-08-27 09:57 | 0.739 | 6.32 | 8.86 | **6.53** | 3.76 | 9.29 | undecided |
+| `70363723` | 4× | 2026-08-25 11:55 | 0.834 | 6.78 | 10.16 | **7.05** | 4.16 | 9.94 | undecided |
+| `70363723` | 4× | 2026-08-27 09:57 | 0.925 | 6.71 | 10.45 | **7.02** | 4.14 | 9.89 | undecided |
+| `30178fc1` | 1× | 2026-08-31 12:56 | 0.404 | 4.55 | 7.14 | **4.76** | 2.16 | 7.35 | undecided |
+| `30178fc1` | 1× | 2026-08-27 07:26 | 0.434 | 5.47 | 7.74 | **5.65** | 3.00 | 8.31 | undecided |
+| `30178fc1` | 2× | 2026-08-31 12:56 | 0.471 | 5.24 | 8.66 | **5.52** | 2.84 | 8.19 | undecided |
+| `30178fc1` | 2× | 2026-08-27 07:26 | 0.518 | 5.34 | 7.91 | **5.55** | 2.84 | 8.26 | undecided |
+| `30178fc1` | 4× | 2026-08-31 12:56 | 0.524 | 6.25 | 8.27 | **6.41** | 3.67 | 9.16 | undecided |
+| `30178fc1` | 4× | 2026-08-27 07:26 | 0.588 | 5.93 | 8.35 | **6.13** | 3.36 | 8.89 | undecided |
+| `6097e735` | 1× | 2026-08-28 07:24 | 0.415 | 5.31 | 6.83 | **5.44** | 2.75 | 8.12 | undecided |
+| `6097e735` | 1× | 2026-08-31 12:56 | 0.441 | 4.22 | 5.88 | **4.36** | 1.75 | 6.96 | undecided |
+| `6097e735` | 2× | 2026-08-28 07:24 | 0.487 | 5.42 | 7.34 | **5.57** | 2.83 | 8.31 | undecided |
+| `6097e735` | 2× | 2026-08-31 12:56 | 0.530 | 4.61 | 6.84 | **4.79** | 2.11 | 7.47 | undecided |
+| `6097e735` | 4× | 2026-08-28 07:24 | 0.546 | 5.46 | 7.89 | **5.66** | 2.84 | 8.47 | undecided |
+| `6097e735` | 4× | 2026-08-31 12:56 | 0.605 | 5.04 | 7.67 | **5.26** | 2.54 | 7.97 | undecided |
+| `704b8e45` | 1× | 2026-08-28 07:26 | 0.470 | 4.47 | 6.35 | **4.62** | 2.05 | 7.19 | undecided |
+| `704b8e45` | 2× | 2026-08-28 07:26 | 0.578 | 5.14 | 6.71 | **5.27** | 2.61 | 7.93 | undecided |
+| `704b8e45` | 4× | 2026-08-28 07:26 | 0.673 | 5.19 | 7.49 | **5.38** | 2.65 | 8.10 | undecided |
+| `6aec8bd8` | 1× | 2026-08-23 07:26 | 0.438 | 5.52 | 6.53 | **5.60** | 2.90 | 8.29 | undecided |
+| `6aec8bd8` | 1× | 2026-08-26 07:26 | 0.466 | 5.02 | 6.25 | **5.12** | 2.44 | 7.80 | undecided |
+| `6aec8bd8` | 2× | 2026-08-23 07:26 | 0.524 | 5.95 | 7.61 | **6.08** | 3.30 | 8.87 | undecided |
+| `6aec8bd8` | 2× | 2026-08-26 07:26 | 0.571 | 5.70 | 7.40 | **5.84** | 3.10 | 8.58 | undecided |
+| `6aec8bd8` | 4× | 2026-08-23 07:26 | 0.596 | 6.19 | 7.85 | **6.32** | 3.50 | 9.14 | undecided |
+| `6aec8bd8` | 4× | 2026-08-26 07:26 | 0.663 | 5.94 | 7.38 | **6.06** | 3.28 | 8.83 | undecided |
+| `118f1f13` | 1× | 2026-08-30 07:24 | 0.468 | 3.78 | 6.70 | **4.02** | 1.54 | 6.50 | undecided |
+| `118f1f13` | 2× | 2026-08-30 07:24 | 0.576 | 4.42 | 7.42 | **4.66** | 2.09 | 7.24 | undecided |
+| `118f1f13` | 4× | 2026-08-30 07:24 | 0.669 | 5.18 | 7.98 | **5.41** | 2.78 | 8.04 | undecided |
+| `6a37ed0a` | 1× | 2026-08-25 07:26 | 0.441 | 4.17 | 6.13 | **4.33** | 1.76 | 6.89 | undecided |
+| `6a37ed0a` | 1× | 2026-08-28 07:26 | 0.459 | 4.04 | 5.30 | **4.14** | 1.59 | 6.70 | undecided |
+| `6a37ed0a` | 2× | 2026-08-25 07:26 | 0.530 | 4.94 | 7.10 | **5.12** | 2.46 | 7.77 | undecided |
+| `6a37ed0a` | 2× | 2026-08-28 07:26 | 0.560 | 4.42 | 6.39 | **4.58** | 1.94 | 7.22 | undecided |
+| `6a37ed0a` | 4× | 2026-08-25 07:26 | 0.605 | 5.50 | 7.58 | **5.67** | 2.99 | 8.36 | undecided |
+| `6a37ed0a` | 4× | 2026-08-28 07:26 | 0.647 | 4.75 | 7.67 | **4.99** | 2.35 | 7.62 | undecided |
+| `2aa911db` | 1× | 2026-08-29 07:26 | 0.404 | 4.04 | 5.57 | **4.17** | 1.61 | 6.72 | undecided |
+| `2aa911db` | 1× | 2026-08-30 07:26 | 0.452 | 3.92 | 5.66 | **4.07** | 1.52 | 6.61 | undecided |
+| `2aa911db` | 2× | 2026-08-29 07:26 | 0.471 | 4.24 | 6.51 | **4.43** | 1.86 | 6.99 | undecided |
+| `2aa911db` | 2× | 2026-08-30 07:26 | 0.548 | 4.63 | 6.02 | **4.74** | 2.15 | 7.32 | undecided |
+| `2aa911db` | 4× | 2026-08-29 07:26 | 0.524 | 4.93 | 6.97 | **5.10** | 2.45 | 7.74 | undecided |
+| `2aa911db` | 4× | 2026-08-30 07:26 | 0.630 | 5.06 | 6.86 | **5.20** | 2.55 | 7.86 | undecided |
+| `2b865779` | 1× | 2026-08-26 07:26 | 0.453 | 4.33 | 6.60 | **4.51** | 1.92 | 7.11 | undecided |
+| `2b865779` | 2× | 2026-08-26 07:26 | 0.549 | 5.02 | 7.92 | **5.26** | 2.58 | 7.94 | undecided |
+| `2b865779` | 4× | 2026-08-26 07:26 | 0.632 | 4.90 | 8.32 | **5.18** | 2.49 | 7.88 | undecided |
+| `23baeba2` | 1× | 2026-08-31 12:56 | 0.405 | 4.06 | 5.37 | **4.17** | 1.16 | 7.18 | undecided |
+| `23baeba2` | 1× | 2026-08-29 07:06 | 0.409 | 3.61 | 5.93 | **3.80** | 1.25 | 6.34 | undecided |
+| `23baeba2` | 2× | 2026-08-31 12:56 | 0.472 | 4.85 | 6.32 | **4.97** | 1.85 | 8.08 | undecided |
+| `23baeba2` | 2× | 2026-08-29 07:06 | 0.478 | 4.30 | 5.85 | **4.43** | 1.81 | 7.05 | undecided |
+| `23baeba2` | 4× | 2026-08-31 12:56 | 0.525 | 4.75 | 6.79 | **4.91** | 1.73 | 8.10 | undecided |
+| `23baeba2` | 4× | 2026-08-29 07:06 | 0.534 | 5.04 | 5.46 | **5.08** | 2.42 | 7.73 | undecided |
+| `930b4313` | 1× | 2026-08-23 07:01 | 0.383 | 3.77 | 5.64 | **3.92** | 1.38 | 6.46 | undecided |
+| `930b4313` | 1× | 2026-08-24 07:14 | 0.429 | 3.62 | 6.07 | **3.82** | 1.29 | 6.35 | undecided |
+| `930b4313` | 2× | 2026-08-23 07:01 | 0.439 | 4.14 | 6.92 | **4.37** | 1.78 | 6.95 | undecided |
+| `930b4313` | 2× | 2026-08-24 07:14 | 0.510 | 4.19 | 5.71 | **4.32** | 1.76 | 6.87 | undecided |
+| `930b4313` | 4× | 2026-08-23 07:01 | 0.482 | 4.85 | 6.92 | **5.02** | 2.41 | 7.62 | undecided |
+| `930b4313` | 4× | 2026-08-24 07:14 | 0.577 | 4.35 | 6.38 | **4.51** | 1.94 | 7.09 | undecided |
+| `796efbfe` | 1× | 2026-08-25 07:06 | 0.351 | 3.59 | 6.89 | **3.86** | 1.44 | 6.28 | undecided |
+| `796efbfe` | 1× | 2026-08-28 06:58 | 0.367 | 2.98 | 4.89 | **3.13** | 0.80 | 5.46 | undecided |
+| `796efbfe` | 2× | 2026-08-25 07:06 | 0.392 | 3.65 | 8.03 | **4.01** | 1.53 | 6.48 | undecided |
+| `796efbfe` | 2× | 2026-08-28 06:58 | 0.415 | 3.06 | 4.94 | **3.21** | 0.83 | 5.60 | undecided |
+| `796efbfe` | 4× | 2026-08-25 07:06 | 0.422 | 4.03 | 7.65 | **4.32** | 1.80 | 6.84 | undecided |
+| `796efbfe` | 4× | 2026-08-28 06:58 | 0.451 | 3.21 | 5.15 | **3.37** | 0.99 | 5.75 | undecided |
+| `e8dc63e1` | 1× | 2026-08-24 02:05 | 0.017 | 0.20 | 0.25 | **0.20** | -1.39 | 1.79 | **below** |
+| `e8dc63e1` | 2× | 2026-08-24 02:05 | 0.015 | 0.15 | 0.25 | **0.17** | -1.35 | 1.72 | **below** |
+| `e8dc63e1` | 4× | 2026-08-24 02:05 | 0.015 | 0.15 | 0.03 | **0.17** | -1.34 | 1.69 | **below** |
+
+**None of the 66 cells is over**; 3 are established below and 63 undecided. So many undecided is
+what the guarantee cost — the family is **1,836** = (78+78+78+72) × 6 opponents.
+
+**The state axis had the same defect.** The previous round chose each team's worst state by a
+low-seed screen — exactly what had just been thrown out on the opponent axis: discarded states
+carry no guarantee, the family never counts that axis, and the screen itself is noise. The state
+is now chosen by two DETERMINISTIC rules — largest summed deficit (how tired the team is overall)
+and largest single-slot deficit (a vector concentrated on one place) — and BOTH are measured. At
+τ=24h the two rules pointed at different states in **27 of 39** (team, L₀) pairs.
+
+**The previous round's "1 over" is withdrawn.** It came from screening 62 opponents, then
+correcting the winner against a family of 153 — the search over 62 was never counted. Corrected
+for it (α = 0.05/18972), nothing at τ=24h is established over. **"none over" and "all established
+below" remain different claims** — the first says exceedance cannot be established, the second
+says non-exceedance can — and the 63 undecided are neither.
+
+**At the short τ, however, an exceeding matchup does exist.** Since the basket is a sample of the
+62, "does an opponent outside it exceed?" is a separate question, asked separately: screen all 62
+per cell under common random numbers (120 seeds), measure the winner on DIFFERENT seeds at full
+count (12,000), and correct for having searched 62.
+
+| τ | (cell, opponent) established over | worst |
+| --- | --- | --- |
+| 6h | **23** | `30178fc1` 4× — weighted 13.32pp, lower bound 9.59pp |
+| 12h | **5** | `6aec8bd8` 4× — weighted 9.81pp, lower bound 6.36pp |
+| 24h · 48h | 0 | — |
+
+Inside the basket, the point estimates for one state spread by **1.98pp** median (4.14pp max) —
+the real measure of how much the opponent moves the verdict, and next to the ≈9.4pp noise range
+above it shows why a screen cannot rank them.
+
+This does NOT replace the verdict above: screening is drawn to noise, so it does not establish
+"we found the worst". What it establishes is an EXISTENCE claim — at τ=6h and 12h the published
+amplitude fails against a real matchup from this window. At τ=24h and 48h not even that stands. A cell is (team 13 × curve 3), and the simultaneous family
+is the SUM of each τ's cell bounds times the opponent basket, **(78+78+78+72) × 6 = 1,836**
+(τ=48h loses one team to the longer warm-up) — the document's claim spans every τ, and the counts differ per τ. A
+cell is (team × curve) and the two modes are mixed **inside** it by exposure, because that is
+how the gate is defined. The per-mode figures sit beside it because they show WHERE the effect
+lives: knockout point estimates run about 1.5× ladder's. Holding each mode to 5pp separately
+made 38 of 78 look over, and that was the earlier "almost every team, in knockout" conclusion.
+**It is withdrawn — it was measured against the wrong bar.**
+
+**Choosing the state by summed deficit was also wrong.** The sum ignores WHICH slot is tired,
+and influence differs by slot (§3-16's ZP fallback slot is the example) and by mode. Every one
+of a team's states is now measured on 200 seeds with COMMON RANDOM NUMBERS across states, and
+the worst weighted edge wins. That choice differed from the summed-deficit pick in **29 of 39**
+cells.
+
+**But "we found the worst" is not claimable — that was measured too.** The gap between first
+and second place is **0.16pp** median (min 0.00), while the gap between first and last is
+**4.41pp** median (min 0.00). The range is some 27 times the margin, so the screen carries
+real signal and **only the top is tied** — the difference between what was chosen and the true
+worst is about the margin, ≈0.2pp, which cannot move a 5pp gate. It also reframes the
+disagreement with the summed-deficit pick: not "the proxy was badly wrong" but "the top states
+are tied, so any rule picks one of several equals".
+
+The selection comes from an outcome, so winner's-curse bias remains. That is why the states are
+**re-screened at the chosen amplitude** before validation, then measured on independent seeds:
+freezing the selection at the published amplitude would leave validation unable to see a state
+that is worse at a much smaller `b`, because the response is nonlinear. The bisection itself
+runs on the frozen selection as a SEARCH HEURISTIC; only the published verdict is exact at its
+own amplitude, and if validation fails the amplitude halves and the whole step repeats.
+
+**Those repeats spend a divided alpha budget.** Up to five independently seeded confirmations
+that accept the FIRST to pass, each spending the full `0.05/1836`, would inflate the chance of
+falsely declaring all cells below the gate by as much as five times. Attempt `i` gets
+`1/2^(i+1)` of the budget, which sums to under one, so the union bound preserves the family-wise
+rate while the first attempt — usually the only one — pays a factor of two rather than five.
+The correction moved only the confirmation bound; the safe amplitude itself is unchanged,
+since the bisection spends the search budget.
+
+**And the family is not this τ's cells alone.** The document claims a safe amplitude exists for
+EVERY τ, so spending a separate 5% per τ would let the chance that at least one bound misses
+run well above 5%. The simultaneous correction is over the SUM of each τ's cell bounds times the basket (**1,836**), counted rather than
+multiplied — the burn-in is proportional to τ, so a team can have post-warmup states at a short
+τ and none at a long one, and here τ=48h does lose one. The cell count printed beside the tally
+stays this τ's, because labelling the family while the tally sums to something else is the next
+accident.
+
+**The vector has to be a SYNCHRONOUS state.** An earlier version took each player's maximum
+over all appearances independently and assembled the eleven — which puts one player's day-3
+peak beside another's day-9 peak in an XI that **never existed**. That is the same mistake as
+averaging across timelines, transposed onto the time axis, and it runs the other way: it
+inflates the vector and overstates the exceedance. In the three-team version, fixing it moved
+13 → **12** cells over and the worst cell down with it (the safe amplitude `b` was
+unchanged).
+
+**The "last lineup" machinery disappeared as a side effect.** The lineup now comes from the
+same state as the loads, so loads can never be applied to a player who has since left — the
+mismatch that machinery existed to prevent is now unrepresentable. Each state is the worst
+OBSERVED one per team and `L₀` (**649** post-warm-up candidates = that team's appearances),
+chosen on summed deficit, before a single match is played. `b` is a common factor of the
+deficits so it cannot reorder them — the bisection tests the same state throughout. `L₀` can
+and does reorder them: `70363723`'s worst kickoff differs at 1×, 2× and 4×.
+
+ Two things had been hiding that. The gate rebuilt the
+curve per team, which pins each team's MEDIAN deficit to 0.20 — the median column read
+exactly 0.200 everywhere and I took that for the design working, when it meant no published
+parameter set was being tested and the most-loaded team's deficit was suppressed. And it ran
+ladder only; knockout bites more than twice as hard.
+
+### So what amplitude is safe?
+
+Exceeding the gate does not mean the mechanism fails; it means the amplitude is too large.
+Bisect for the largest `b` whose worst cell stays under, measuring EVERY cell at each step
+rather than assuming linearity — a single proportional guess landed at 5.15, still grazing it.
+
+| `b` | cell hit / worst | upper | lower | max deficit | verdict |
+| --- | --- | --- | --- | --- | --- |
+| 1.3394 | `70363723` 1× | 13.38 | 3.51 | 0.856 | not established |
+| 0.6697 | `70363723` 1× | 8.66 | -0.12 | 0.428 | not established |
+| 0.3349 | `70363723` 1× | 5.70 | -1.67 | 0.214 | not established |
+| 0.1674 | `23baeba2` 1× | 4.84 | -2.28 | 0.083 | **below** |
+| 0.2511 | `70363723` 1× | 5.28 | -2.09 | 0.160 | not established |
+| 0.2093 | `30178fc1` 1× | 5.03 | -2.12 | 0.111 | not established |
+| 0.1884 | `6aec8bd8` 1× | 4.91 | -2.19 | 0.101 | **below** |
+| 0.1988 | `30178fc1` 1× | 4.97 | -2.12 | 0.098 | **below** |
+| 0.2041 | `30178fc1` 1× | 5.06 | -2.06 | 0.101 | not established |
+| 0.2014 | `30178fc1` 1× | 5.14 | -2.02 | 0.107 | not established |
+| **0.1988** (validated on independent seeds) | `6aec8bd8` 1× | **3.86** | — | 0.106 | **all cells below** |
+
+**`b` ∈ 0.1988–0.2014**, with all 66 cells established below the gate on independent seeds at
+the lower end — after re-screening the states at that amplitude — and the upper end being the
+smallest candidate that failed.
+
+**Do not read that bracket as precision.** The fail rows are not monotone even though the
+deficit falls with `b`, and the cell that trips changes between them: each bound is estimated
+from 6,000 seeds and near 5 the estimation error exceeds the spacing between candidates, so
+which candidate looks worse is decided by the seeds. The bracket is search granularity, not a
+confidence interval — which is why validation re-measures at twice the sample (12,000 per opponent) on a
+separate seed column. Note
+that the worst cell moves: at the published amplitude it is `L₀`=4×, but at low `b` it is
+**1×**, because a smaller `L₀` saturates sooner and so produces a larger deficit for the same
+`b`. Testing only the cell that was worst at the starting amplitude would have missed that. E's 0.2 was measured with a uniform Δ across all eleven, and
+putting it on the MEDIAN player is what let the top of a real vector (max deficit 0.02–0.88)
+run past the gate. That value is the worst over all 13 teams that played, but NOT over every appearance: the
+state comes from two deterministic rules (largest summed deficit, largest single-slot deficit)
+applied to the 649 post-warm-up candidates, and only those one or two states per (team, curve)
+carry a bound. The remaining appearances are unmeasured — the same kind of limit the opponent
+basket carries. What is gone is selection on an OUTCOME; what is not gone is that this is a
+sample of the observed states, and everything outside this window is unseen entirely.
+
+**The bound's support constant was wrong, too.** One `ser` term averages the home and away
+versions of `(win(fresh) − win(tired))`, then multiplies by `100 · 2`. Each difference lies in
+`[−1, 1]`, so their average still has support `[−1, 1]` and the term's support is **±200**: both
+orientations can attain the same extreme. The constant once said 100. Halving the support credits
+an unseen outcome with only half its possible effect, which can print an UNSAFE amplitude as below
+5pp. With it corrected to 200, a degenerate cell's
+bound has a floor independent of `b`: at SD = 0 the surviving empirical-Bernstein term is
+`3·R·ln(3/δ)/n` with `R = 2·SER_MAX = 400` and `δ = α/4` (α split once across the two modes and
+again across the two sides), `α = 0.05/family`. If that floor exceeds 5 the bisection cannot
+pass at ANY amplitude — not a finding, a seed shortage — so the probe PRINTS it (2.60pp at
+6,000 seeds per opponent and family 1,836 in this run). Cut the seeds and it climbs past 5 and
+the search reports "none found"; without that line one would read it as a result.
+
+**Student-t was the wrong interval here.** Each seed observation averages the two orientations
+and therefore takes one of nine values, from `−200` through `200` in 50pp steps. It is not a
+normal sample, and t coverage is exact only under normality. The tail this section uses is
+`α = 0.05/1836 ≈ 2.7e-5`; ordinary large-sample
+accuracy says nothing at that extreme, and an anti-conservative cell interval lets an unsafe
+amplitude pass. The gate now uses an **empirical Bernstein** bound (Maurer–Pontil), which
+assumes only boundedness and adapts to the variance — far tighter than Hoeffding here, where
+almost every observation is 0:
+
+    μ ≤ x̄ + SD·√(2 ln(3/δ)/n) + 3R ln(3/δ)/n,   R = support width = 400
+
+The price is width. At this family, a half-width of 2.9pp needs **12,000 seeds per opponent**;
+at 6,000 it is 4.9pp, which would leave the 5pp gate no room at all. So the seed budget per
+opponent went up 4×, the safe amplitude fell from 0.15 to 0.13, and only 3 of 66 cells are
+established below at the published amplitude. Every one of those is the cost of not borrowing
+normality.
+
+It also subsumes the degenerate case: with SD = 0 the second term survives, so the interval
+never collapses to zero width and the separate rule-of-three guard is gone — that guard had
+been a second argument resting on its own iid assumption.
+
+**Degenerate samples no longer need special handling.** Under the old t interval, a cell where
+every seed returned the same value had SD 0, so the interval collapsed to zero width and any
+mean at all printed as "established below the gate" — no variance means unmeasurable, not
+certain, and a small-sample smoke run produced exactly that false clear. A separate
+rule-of-three bound was bolted on for it. Under empirical Bernstein the range term
+`3R ln(3/δ)/n` survives at SD = 0, so the interval never collapses in the first place: no
+guard, no special case, and one fewer argument resting on its own iid assumption.
+
+### Stage 3 — a multi-wallet feeder is exempt, not merely advantaged
+
+F-1's side rule writes nothing for a ladder away side that is human; `playoff-finalize.ts`
+calls `applyResult` for both; `selectOpponentByStrength` draws from the five strength-nearest
+in-division; the cooldown is initiator-only. Observed at-appearance `L` reaches 55 for an
+honest player while a team taking the same match count as the away side accumulates zero. That
+is exemption, not a ratio — bounded only by what bounds the feature, since the prize is 0.59pp at g=0.4 (0.16pp at the τ=24h safe amplitude)
+and the team-level gap is 3.86pp at the safe amplitude, under the gate.
+
+### What this does not settle
+
+Fallback provenance is now bound to (team, index) and cleared the moment it is consumed.
+Predicate D ("C without the ZP fallback") depends on whether `weightedPick` took the fallback
+branch; holding that as a bare index and never clearing it let a FIXED actor — an FK kicker,
+PP1's forward, a penalty taker, none of which go through `weightedPick` — inherit the previous
+selection's mark and lose a legitimate involvement. There is a real producer: `engine.ts:653`
+picks a defender on a failed dribble and no `getPoint` ever consumes it. D moved 15.25 →
+15.27 after the fix; A, B and C were unchanged.
+
+**The same site then produced something larger.** The mark was left dangling only because that
+selection was not being COUNTED at all; it is now (see the predicate table). Counting it also
+consumes the fallback mark at that point, so nothing can be left dangling — the cause is gone
+rather than a second guard added on top of it.
+
+That fix created an adjacent defect, and the document is what caught it. Counting actor-side
+fallbacks gave one counter TWO consumers: predicate D subtracts from `off + def` and needs
+both, while the ZP-fallback table subtracts from `def` alone and needs only the responder's.
+Sharing one field made the table subtract actor fallbacks too, putting "slot 1 minus fallback"
+BELOW the other defenders (0.33 against 0.91) — which does not merely change a number, it
+breaks that table's conclusion. Splitting into `fbOff`/`fbDef` restored 0.90 against 0.91.
+How it surfaced matters: the probe disagreed with the document, and the document was right.
+The reflex was to update the "stale" document to match the new run.
+
+A team with zero load is not a cell. A team whose only appearance in the window is its first
+has eleven zero pre-match loads, so the tired side IS the fresh side, every seed returns the
+same value, and SD is 0 — the degenerate guard then sets that cell's bound to infinity and
+**one such team would block the amplitude search entirely**. The gate table skipped those
+states while `allCells` did not. A deficit of exactly zero is not evidence of clearing the
+gate either, so such teams are excluded outright and the output says how many. None occur in
+this window.
+
+The prize-versus-cost ordering AT THE AMPLITUDE THAT WOULD SHIP — it splits at `g=0.4`
+(−0.43pp [−0.76, −0.10]) but not at the safe amplitude's own deficit (prize 0.16pp, cheapest
+cost 0.16pp, paired difference −0.00pp [−0.47, 0.46]), and that second one is the decision. The neutral-point fork. Maximum LEGAL role concentration (the
+replay is what happened, not what the rules permit). Player identity is the snapshot's `actorIds` playerId, keyed GLOBALLY rather than per team —
+a minted player sold mid-window would otherwise get two chains starting from zero, understating
+the buyer's vector. A missing actor map or even one player that fails to join is rejected by the
+replay-completeness gate; there is no name-key fallback. **Two code paths went
+untested by this data**: no playerId appears under two teams in the window (0 of 143), and
+only 1 of 754 human team-sides was uncredited — credit now gates the increment alone, since
+condition is read on entry whether or not load is charged, and ownership moves on every human
+appearance. Both are right by construction, which is not the same as having been seen to
+run. All four τ values `{6, 12, 24, 48}h` are integrated separately, with the gate and safe-amplitude
+search rerun on that τ's own cells rather than converted from 24h. At τ=24h the published run has
+66 exposure-mixed cells: 0 established above the gate, 3 established below it, and 63 undecided.
+Snapshot coverage turned out to be a step function — 0% before 2026-08-21 and 100% after — so
+the window is 11 fully covered days rather than a 37% sample, and the script finds that step
+itself. The cost is that nothing with a period longer than 11 days is visible.
+
+### Reproduction
+
+```
+# ENGINE_SINCE is OPTIONAL — the probe verifies each match by replaying it against
+# the stored score, event digest and per-player ratings, so there is no date to infer.
+C2_ENV_FILE=<main checkout>/.env.deploy DAYS=30 WINDOW_END=2026-09-01 \
+  EXPORT_SNAPSHOTS=/tmp/prod-tl.json \
+  pnpm --filter @sws26/api exec tsx scripts/c2-team-load.ts
+MATCHES=3000 SNAPSHOTS=/tmp/prod-tl.json REPLAYS=5 TAU_SWEEP=6,12,24,48 \
+  LAMBDA_MED=6 LAMBDA_MAX=21 \
+  npx tsx packages/mcp/skill/reference/probes/engagement-load.mts
+```
+
+**The window's end is now stated explicitly, so these figures can be rebuilt.** The default is
+still "the last completed UTC day", so an unqualified re-export extends it by a day — but
+`WINDOW_END` makes that date the window's (exclusive) end. The numbers in this section come
+from **08-21 through 08-31, 1,201 matches**, and re-exporting with `WINDOW_END=2026-09-01`
+reproduces that `matches` array **byte for byte** (verified) — which came for free from
+collapsing a window boundary that had been spelled out in eighteen separate places into one
+definition. The window's length remains a limitation.
+
+**"Fully covered" cannot be counted from completed rows alone.** A day with rows still
+unfinished, or whose cup bracket has not completed, is not fully covered — knockout rows are
+inserted only after the preceding round finishes, so **rounds not yet inserted never appear in
+the denominator** and counting completed rows makes such a day look like 100%. Admitting it
+drops the later engagements and biases `L` and the safe amplitude DOWNWARD, the unsafe
+direction. The unfinished-row count and the open-cup count are now printed per day, and a day
+counts as fully covered only when both are zero. No day in this window is affected.
+
+**And that day's own cup has to exist — any completed cup will not do** [Codex P2, 2026-09-02
+×2; 2026-09-04]. A day whose cup was never created still has ladder and playoff rows, so it
+passes all four checks above. Requiring one completed cup is still not enough. The first repair
+still BUILT each group from match kickoff day and merely required the cup-ID date to equal that
+key. If `SEASON_OPEN_AT` pushes the entire cup into the next UTC day, every `wc-<yesterday>` row
+sits in today's group and even yesterday's completed cup contributes nothing to yesterday's
+`done_cups`.
+
+The cup calendar is now built INDEPENDENTLY from the bracket ID's `wc-YYYY-MM-DD` suffix, and
+competitive rows are attached to that cup's own date. Thus, even if every fixture kicks off the
+next day, the `wc-<yesterday>` rows do not invent activity today: they fill yesterday's
+denominator and snapshot count beside yesterday's completed bracket. The two calendars are
+`FULL JOIN`ed, so a bracket-only or match-only hole stays visible. A completed bracket with empty
+`match_ids`, or a day with zero match rows, is not evidence and cannot pass. Only legacy or
+hand-made IDs that carry no date, plus standalone playoffs, fall back to fixture UTC date like
+`cupDateOfFixture`. **The export start, claim check, and postcondition share that same competition-
+day expression**: fixing coverage alone would let the kickoff filter pull a previous cup's
+after-midnight rows back into the new fully covered window. Neither the verdict nor one byte of
+the 1,201-match production export changed; only the delayed-schedule boundary did.
+
+**The window has to close on both clocks** [Codex P2, 2026-09-02; 2026-09-04]. Load events are placed at the
+claim time while the window was cut on kickoff — so a match kicking off just before midnight and
+processed just after is dropped from the export while the coverage table still counts that day as
+100%. That team's engagement vanishes silently and `L` drops, the unsafe direction again. Such a
+match is no longer counted and reported: the window's end is pulled back a day at a time until
+both clocks close. Crucially, the trim now groups by COMPETITION DAY and checks that day's maximum
+kickoff and maximum claim together. Otherwise, pulling the end back to today's midnight for a
+claim can leave a `wc-<yesterday>` row retained by competition date but silently removed by its
+today kickoff. If either clock of a retained competition day crosses, the end retreats another
+day, so a cup stays whole or leaves whole. The postcondition does not pre-filter on the effective
+kickoff end: it counts `kickoff >= end OR claim >= end`, which makes that omission visible. Neither
+clock crosses in this production window, so the trim is zero days.
+
+**The export file is private too** [Codex P2, 2026-09-04]. An engine team inside the snapshot
+contains the effective `total` from which paid potential can be inferred. Writing
+`/tmp/prod-tl.json` with Node defaults commonly creates 0644, while merely passing
+`{ mode: 0o600 }` does not repair an EXISTING 0644 destination. The exporter now writes an
+exclusive 0600 temporary file in the same directory, applies `chmod(0600)`, and atomically renames
+it over the destination. Both a new file and an existing-file replacement are owner-only; a failed
+temporary is removed.
+
+**Replay verification replaces the engine boundary.** The probe replays every match from its
+stored seed and compares three things — the **stored score**, the **event-sequence digest**, and
+the **stored per-player ratings** — then drops the ones that do not reproduce and prints the
+rate. A rate below 100% means the window crossed an engine deployment — no date to infer, and no
+`ENGINE_SINCE` needed (given, it is only a pre-filter).
+A missing event digest, missing ratings, a missing `actorIds` map, or even one of the eleven player
+cells failing to join makes that evidence axis **unverifiable**, not equal; the other axes cannot admit that row.
+Rejecting any such row then trips the whole-window completeness guard below, so the probe neither
+counts absent evidence as agreement nor integrates a window with a missing increment.
+
+**And a window with holes is not integrated — the probe exits.** If some matches were rejected,
+their increments would vanish while every later appearance stayed, so those players would start
+from an artificially low `L` and `Lmed`, the gate vectors and the safe amplitude would all be
+understated. Engine changes are input-dependent, so interspersed failures are a normal
+possibility rather than a clean release boundary.
+
+`LOAD_PRED` selects the integration predicate (default `C`). C is published, and the probe also
+integrates at D at every τ and prints the difference — F-1 has not chosen, so the calibration
+has to carry which predicate it is conditional on.
+
+Without `TAU_SWEEP` only `TAU_H` runs. **Calling a single-τ run a "decision" makes a probe
+default the production recovery rate** — that is why the sweep exists.
+
+`REPLAYS` now applies only to the predicate table, where each fixture is replayed
+independently. The timeline is integrated once, from each match's stored `rng_seed`.
+
+**"`tsc` does not see these probes" is wrong, and it was checked.** #719 added
+`skill/reference/probes/**/*.mts` to `packages/mcp/tsconfig.typecheck.json`, and
+`packages/api/tsconfig.typecheck.json` already includes `scripts`. The local gate covers both
+files. The opposite claim sat here because a pre-#719 fact was copied forward unverified —
+not harmless, since believing there is no static gate changes how one writes.
+
+What is genuinely uncovered is the throwaway EDIT scripts. A backtick terminating a template
+literal early broke one seven times; the last was a markdown table's `` `70363723` `` breaking
+the edit script itself. No tsconfig includes those. The remedy is asserting the match count on
+every replacement, so a broken script dies before it touches the file — which is exactly what
+happened. A smoke run is a behavioural check, not the parse gate.