pog-mcp 0.9.1 → 0.9.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pog-mcp",
-  "version": "0.9.1",
+  "version": "0.9.2",
   "type": "module",
   "description": "MCP server that lets an AI agent play Proof of Goal — wallet, sign-in, squad building, and matches as typed tools.",
   "license": "MIT",

package/skill/reference/measurements.md

@@ -644,6 +644,588 @@ secondary edge on top. An agent that only differentiates its roles and never
 touches the keeper beats a naive squad 44.0% of the time and loses 9.7% — better
 than four wins to one loss, with the rest drawn.
 
+## Which engine constant is a detectable lever? (`constant-sweep.mts`)
+
+Every other section here sweeps a SQUAD and holds the rules fixed. This one does
+the opposite. It pins a basket of published archetypes and moves the hardcoded
+numbers in `packages/engine/src/engine.ts`, one at a time, because that is the
+shape a "weekly rule" would ship: one global constant moved for everybody.
+
+The statistic is not "did the game change" — that is easy — but **"did the
+ordering of squads change"**. A rule that makes every match higher-scoring
+without moving anyone's best answer is a texture knob, not a lever.
+
+**How the sweep reaches a `const`.** `engine.ts` is read, one exact string is
+replaced (the match count is asserted, so an engine edit fails the probe loudly
+instead of silently sweeping nothing), its two relative imports are rewritten to
+absolute paths, and the result is written to a `mkdtemp` directory and imported.
+Nothing is written inside the repository. `fair` needs no patch: production never
+sets it — there is no `fair:` outside the engine's own tests, and every engine
+input is built from an explicit field list — so it is a global constant in
+practice, and setting it on both squads IS the rule change.
+
+**Not swept: the zone-press coefficient.** Its selector is broken — every
+candidate weight is `0.0`, so `weightedPick` falls back to "the first non-GK
+slot" — and a coefficient in front of an undefined selection measures nothing.
+Forbidden until the selector is decided.
+
+### The statistic, and why not the obvious one
+
+Rows below are in **win-equivalent share**: a win is 1, a draw is 0.5, a loss is
+0, and every match counts. The obvious alternative — win rate among DECIDED
+matches — conditions on something the swept constants themselves move. Free-kick
+occurrence takes the regulation draw rate from 57.9% to 38.1% across its range,
+so a decided-only percentage mixes "who wins more" with "which matches got
+decided at all". It also inflates the number the activation gate is read
+against: an 11/80/9 record reads 55% decided, clearing a +5pp gate, while the
+squad's actual advantage over all matches is one point. The gate's own
+justification is stated in weekly LEAGUE POINTS, so the statistic it is compared
+to has to include the draws that earn them.
+
+Deltas are **paired**. Every condition replays one seed column, so the default
+cell and the swept cell are matched match for match, and the test differences
+the columns rather than treating them as independent samples. An unpaired
+denominator would only be safe if the residual covariance stayed positive, and a
+rule that REVERSES a fixture is exactly where it need not — which would make the
+unpaired test anti-conservative on the cells this probe most wants to be sure
+about. (Both of these were review findings on the first version of this section,
+which used decided-only percentages and an unpaired z.)
+
+The same seed column is shared across the two MODES as well, and `gameFlg` is
+read only after minute 90 — so a match decided in regulation keeps that result
+in knockout mode, and the two modes' estimates are positively correlated. The
+exposure-weighted test therefore does not add their errors in quadrature; it
+takes the Cauchy-Schwarz upper bound, which is looser than the truth and cannot
+be anti-conservative. That is the only direction an acceptance gate may err in.
+
+### Controls
+
+| Control | Result |
+| --- | --- |
+| Mirror, `gameFlg 0` (one squad against itself; true z = 0) | z **+0.31** |
+| Mirror, `gameFlg 1` | z **+0.28** |
+| Patch identity — the temp-dir pipeline at the SHIPPED values vs the stock engine | 1,000 matches identical |
+| `fair` identity — `fair: 5` written out vs `fair` omitted | 1,000 matches identical |
+| Paired alignment — a re-patched shipped-value engine vs the default COLUMN | 200,000 matches, largest per-match difference **0** |
+| Counter trigger, swept 0% to 100% of eligible turnovers, in both baskets | must be REJECTED, and is — the run ABORTS if it is not |
+
+**Every control in that table aborts the run; none of them warns.** A mirror
+that fails is the condition that makes every table unreadable, and a negative
+control that PASSES is an acceptance bar that has failed its own calibration.
+Printing "bar too loose" and exiting 0 was itself a defect — an invalid run
+could be consumed as a successful one.
+
+One more number is counted every run, and it is a MEASUREMENT rather than a
+control — nothing aborts on it, because a knockout draw is a legitimate outcome
+to count rather than a failure:
+
+| Measured every run | Result |
+| --- | --- |
+| Knockout draws, across both baskets | **0 in 5,440,000** knockout matches |
+
+That draw count is a claim this section originally got wrong in the other
+direction. It said knockout mode "has no draws by construction". It nearly does:
+`runExtraTimeLoop` breaks a tie on INEQUALITY, so extra time either produces a
+golden goal or falls through to a sudden-death shootout. But an earlier version
+of that check read "a score changed", and one attack CAN score for both sides —
+a post-play layoff gives a second shot, and a `gk_catch` can launch a counter
+that scores — which ended extra time level and skipped the shootout. The
+constants swept here move goal production, which is exactly what that path keyed
+on, so the column is counted every run instead of declared.
+
+The patch-identity control is the one this section could not do without: without
+it, "the constant did it" and "the import rewrite did it" are the same
+observation. The alignment control is the one the PAIRED test could not do
+without — a paired test cannot detect misaligned columns from its own output, so
+a freshly patched shipped-value engine is required to reproduce the default
+column match for match, in order.
+
+3,400,000 matches in the narrow basket (5 archetypes, 10 fixtures, 10,000
+matches each, 17 conditions, both modes) plus 7,480,000 in the wide one.
+**Bonferroni over m = 8,320 pre-registered comparisons, two-sided: |z| >= 4.53.**
+That m prices the SEARCHES, not just the tests. A gate cell does not spend one
+hypothesis, and BOTH of its sides are data-chosen. The baseline side picks a
+leader out of K and builds a band around it — noisy standings could have named
+any of the K, so it could have surfaced any pair, K(K−1)/2, not just the K−1 the
+leader it happened to pick needed. The swept side then reads a head-to-head
+against that band, another K(K−1)/2. So K(K−1): **20 hypotheses per narrow gate
+cell and 110 per wide one**, not 5 and 11. Every gate is also derived at all three
+exposure weights, and all three are charged. On top of that sit the two fixture
+families: the narrow basket's 10 fixtures and the wide basket's 55, at every
+swept value — without the second, an axis could be called INERT while
+significantly moving fixtures among the six archetypes the wide basket exists to
+cover.
+
+**The bar is a z and does not move with the sample size; the effect it
+corresponds to does.** There are two test families here and they have different
+answers — the fixture deltas are a PAIRED test between two conditions, the
+activation gate is a ONE-SAMPLE test of a head-to-head share, and their standard
+errors are unrelated. For the fixture deltas, per basket (they run at different
+N, so averaging would let the better-powered one speak for the other):
+
+Two numbers, because they are two different things. `bar x SE` is the effect
+whose EXPECTED z equals the cutoff — a significance threshold at about 50%
+power, not a minimum detectable effect. The 80%-power figure adds z(0.80):
+
+| median / worst | significance threshold | 80%-power effect |
+| --- | --- | --- |
+| narrow, regulation | 1.0 / 2.2pp | 1.2 / 2.6pp |
+| narrow, knockouts | 1.8 / 3.0pp | 2.1 / 3.6pp |
+| wide, regulation | 1.8 / 3.8pp | 2.1 / 4.5pp |
+| wide, knockouts | 2.9 / 4.8pp | 3.4 / **5.7pp** |
+
+Read the worst column, not the median, since "every test is powered enough" is a
+claim about the tail — and read it honestly: **the worst wide-knockout fixture
+cell is not powered to catch a 5pp effect at 80%.** Every other cell is. That
+qualification applies to the MARGINS classification, which is what the fixture
+deltas decide; the gate is a different test, and its own answer is below.
+
+For the gate itself the answer is different and sharper: **no pair was ever
+formed.** At every swept value, in both baskets, the new optimum was already
+inside the default band, so there was nothing to test.
+
+So the gate has no power line to print in THIS run. It has one for a rerun that
+does form a pair, and it reports the same two quantities as the table above plus
+a warning when the 80%-power effect exceeds the 5pp gate itself — because a pair
+that just misses under an SE that wide is underpowered, not refuted. Forcing a
+band of one makes both branches fire, which is how they were checked: the wide
+basket lands at a 7.1pp threshold and an 8.5pp 80%-power effect, i.e. above the
+gate, and says so.
+
+That only settles anything if the band is SHARP — noisy standings widen it, and
+a wide enough band absorbs every new leader, which would be a power limitation
+wearing the costume of a result. So the probe reports the band's own
+discrimination: it holds **2 of the 5** narrow squads and **3 of the 11** wide
+ones, with its edge **1.45pp** and **1.60pp** from the leader. A band that tight
+is not swallowing the basket. That is the number a smaller rerun moves, and the
+one to check before reusing this conclusion.
+
+### The table
+
+Basket: `role-433`, `def-433`, `def-532`, `shootonly-433`, `flat-433` — five
+outfield templates, the shared `squad-lib.mts` builders, so a row here and a row
+in the round-robin above are the same squads.
+
+- **`sig`** — fixture deltas clearing the corrected bar, out of 40 in the narrow
+  basket (10 fixtures x 4 swept values) and out of 220 in the wide one (55 x 4).
+  Both are printed, because an axis can be classified on the wide basket's
+  evidence alone and a table that hid it would publish a verdict whose decisive
+  measurement is invisible.
+- **`rev`** — of those, how many crossed 50%: the fixture's ANSWER flipped, not
+  just its margin.
+- **`gate`** — the new optimum against the whole stale band, under the new rule:
+  the SMALLER of the two baskets' gains, and only if both are significant. In
+  WIN-RATE advantage points, the unit the plan writes its 5pp threshold in. Track 3.1 activates on this at +5pp, but **on an
+  exposure-weighted mix of the two modes, not per mode** — see below.
+
+| Axis (shipped) | Mode | sig (narrow) | sig (wide) | rev | Largest narrow delta | gate | What that row leaves FREE |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| **FK occurrence** `rand(10)<1` | reg | 21/40 | 102/220 | 0 | **+11.8pp** `role-433` vs `flat-433` @x4 | none | the FK taker: every basket squad kicks from slot 9, so this is "more free kicks", never "more free kicks AND a better taker" |
+| | ko | 18/40 | 100/220 | 0 | +8.8pp `def-532` vs `flat-433` @x4 | none — +7.5pp against `role-433` alone, 0.0pp against the band it belongs to | shootout exposure — it is a property of the PAIRING, and the basket fixes both halves |
+| **Post play** `rand(10)<1.5` | reg | 18/40 | 50/220 | 0 | −4.8pp `role-433` vs `flat-433` @x4 | none | the PP2 shooter is drawn by `getPlayer`, so the row confounds "more post plays" with "who the draw lands on" |
+| | ko | 5/40 | 40/220 | 0 | −3.4pp `role-433` vs `shootonly-433` @x4 | none | same, plus FW count: 433 and 532 field three and two forwards, and `fwIdx.length` gates the branch |
+| **Counter** `THRESHOLD=5` | reg | 3/40 | 5/220 | 0 | −2.0pp `shootonly-433` vs `flat-433` @0% | none | the four sampled thresholds span 0% to 100%, but changing this constant reroutes RNG consumption, so outcomes need not interpolate between them |
+| | ko | **0/40** | 7/220 | 0 | −1.2pp (below bar) | none | as above. This is the control: it never produced a gate in either basket, which is the only behaviour that matters here |
+| **`fair`** default 5 | reg | 13/40 | 37/220 | **1** | **+10.9pp** `shootonly-433` vs `flat-433` @22 | none | `fair` is uniform across all 22 players here; per-player or per-position `fair` is a different (and unmeasured) axis |
+| | ko | 7/40 | 23/220 | 0 | +10.9pp `shootonly-433` vs `flat-433` @22 | none | keeper `defense`, which every extra penalty routes through — held fixed inside each template |
+
+**Nothing here becomes a candidate.** All eight rows move margins somewhere — the
+control included, once the wide basket is counted — and
+the one pairing that re-prices anything does so against a squad that was already
+tied for best. The pass set at the plan's own activation gate is empty over both
+baskets, and the answer took a day rather than a release.
+
+**Empty for these four axes, which is not the same as empty for the pool.** The
+plan's weekly-rule pool also names set-piece `dBonus`, penalty attack noise, and
+base `dBonus`; none of them is swept here, and this run says nothing about them.
+The axes measured are the ones the issue behind this section listed — the two
+hardcoded rates, `fair`, and the counter control. Anyone reading this as
+"schedule nothing" should read it as "schedule none of these four, and go
+measure the other three before concluding anything about the pool."
+
+The verdict is deliberately absent from the rows above: it belongs to the AXIS,
+not to a mode. The next section is where it is issued.
+
+### Activation is per AXIS, not per mode
+
+The gate is specified on an exposure-weighted mode mix, and the two modes are
+nowhere near equally exposed: the ladder has no cap beyond a five-minute
+cooldown, while the cup runs once a day. **Cup entry is the top 48 of the whole
+ladder, ACROSS divisions** — division decides weekly matchmaking, not cup
+access, so what actually zeroes a manager's knockout exposure is never
+QUALIFYING, at any division. A rule that clears 5pp in knockouts alone has not
+cleared the gate, and reading the verdict per mode would say it had. The weight
+below is a regular qualifier's day: **1.33 / (1.33 + 12 + 3) = 0.081** — 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
+too. Leaving them out gives 0.100 and overweights the knockout evidence.
+`KO_EXPOSURE` is an argument, and the sensitivity below is why the choice does
+not matter here.
+
+**One value per row: the one that got FURTHEST through the ordered gates.** That is the same thing as "the best schedulable value" only when nothing blocks it, and here everything is blocked. A weekly rule ships one
+global number, so pairing regulation's best value with knockout's best would
+price a rule nobody can schedule — and could clear 5pp when no real value does.
+For the same reason the row is not simply the biggest number: a value blocked on
+one gate must not hide a smaller value that passes them all.
+
+**And one SQUAD per row.** A manager fields one build across a schedule that is
+mostly ladder and partly cup; the same eleven cover both. So the mix is applied
+to each squad's season first, and the optimum and the stale band are read off
+that one weighted table — averaging two separate per-mode contrasts would price
+an improvement no single squad achieves.
+
+**The wide basket vetoes; it does not compete.** Its eleven archetypes CONTAIN
+the narrow five (the probe refuses to run if they ever stop containing them), so
+a stale response that survives there is a build the manager could actually have
+fielded, and it disproves a re-pricing the smaller set claimed. Taking whichever
+basket happened to pass would let a subset's claim stand against its own
+superset's counter-example, so the gate takes the smaller of the two gains — and
+requires BOTH to be significant, rather than inheriting significance from
+whichever had the smaller point estimate.
+
+Track 3.1 has four activation gates, not one — a gain threshold, evidence, "no
+dominant build", and draw-rate and goals-per-match bands. All four are applied
+here, and the third is applied in TWO pieces because only one of them is
+decidable.
+
+The plan states that gate twice: absolutely ("no dominant build") and as a
+change ("dominance must not increase"). The absolute form is testable and is
+tested — `domAbs`, one squad significantly beating every other in the wide
+basket, refused at the corrected bar. The change form is not. `dom` is a
+difference of two standardised pairwise z-scores, which carries SD ~1.4 at ANY
+sample size, so a literal "> 0" would refuse half of all rules that changed
+nothing, and no amount of simulation shrinks that. Nor can the difference be
+paired away: the leader and its weakest opponent are argmax and argmin picks
+that can change IDENTITY between the two conditions, so there is no fixed pair
+to difference. What this probe does instead is a 2-sigma DIRECTION screen at
++3.0. It refuses far more than a corrected test would — that would sit near 6.4
+— which is the safe direction for a gate whose job is to refuse. It is a
+heuristic, it says so in the probe's own legend, and it is not the plan's
+sentence. The plan does not put numbers on the feel bands, so the ones
+used are the probe's declared defaults (|draw| <= 5pp, |goals| <= 20%,
+overridable) and every row prints its measured movement so a different band can
+be applied to the same table by eye.
+
+**The gate is measured against the whole stale optimum, not one member of it.**
+Where the default standings do not resolve a single best build — and in knockout
+mode they do not, `role-433` and `gkheavy-433` are not separated — the manager
+who never re-solves could already have been on any squad in that band. Crediting
+a rule for a switch that was free at the shipped rule would be measuring the
+tie, not the rule. So the gain is the WORST margin against the whole band, and a
+new leader drawn FROM the band scores zero by construction.
+
+| Axis | Best value | Gate (ko weight 0.081) | dom | domAbs | draw | goals | Blocked by |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| FK occurrence | x0.25 | **0.00pp** | +0.4 | 0.4 | **+5.1pp** | −16% | gain — and it would also fail the draw band |
+| Post play | x0.25 | **0.00pp** | −0.1 | 0.0 | −1.0pp | +3% | gain |
+| Counter (control) | 0% | **0.00pp** | +0.7 | 0.9 | +1.4pp | −6% | gain |
+| `fair` | 1 | **0.00pp** | +0.5 | 0.6 | +3.6pp | −12% | gain |
+
+`dom` is the CHANGE in the leader's weakest margin; `domAbs` is that margin
+itself, in the eleven-archetype basket. Both block, and the two are not
+interchangeable: a rule that leaves an already-dominant squad exactly where it
+stood scores a `dom` near zero while failing the requirement outright, which is
+why the absolute column exists at all. `domAbs` is the one carrying the plan's
+sentence — at or past the corrected bar, the same |z| the rest of this section
+uses, 4.53 at these sample sizes, one squad significantly beats every other,
+which is what a dominant build is. `dom` is the direction screen described
+above, at a threshold this probe declares rather than quotes.
+
+Both are computed on the exposure-WEIGHTED season, from weighted head-to-heads
+at the activation weight. Taking the larger of the two pure modes does not cover
+the mix: a squad's margins against different opponents are each linear in the
+weight, so the minimum of them can peak at an interior crossing even when the
+squad dominates at neither endpoint.
+
+**Every gate is zero, and the reason is sharper than "the effects are small".**
+The one cell that re-prices anything — FK x0.25 in knockouts — promotes
+`gkheavy-433`, which was already tied for best at the shipped rule. A manager
+sitting on it gains nothing from the rule; a manager sitting on `role-433` could
+have switched for free before the rule existed. Nothing here beats what was
+already available.
+
+And the exposure weight — the one judgement call in this section — does not
+carry it. The probe re-derives the GAIN at both extremes rather than asserting
+the choice does not matter. (Only the gain; dominance, feel and fixture movement
+are worst-of-both-modes by design and are not re-evaluated per weight, so this
+table is not four verdicts.)
+
+| Axis | Ladder-only (w=0) | Mixed (w=0.081) | Knockout-only (w=1) |
+| --- | --- | --- | --- |
+| FK occurrence | 0.00pp | 0.00pp | 0.00pp |
+| Post play | 0.00pp | 0.00pp | 0.00pp |
+| Counter (control) | 0.00pp | 0.00pp | 0.00pp |
+| `fair` | 0.00pp | 0.00pp | 0.00pp |
+
+Those three points do not settle the interior on their own: each squad's
+weighted score is linear in the weight, the leader is their upper envelope, and
+an envelope can hand the lead to a THIRD squad only somewhere in the middle. Nor
+does a grid settle it — an interval can be narrower than the step.
+
+So the interior is **solved**. Everything the gate reads is linear in the
+weight, which makes all three of its boundaries enumerable:
+
+- where two squads' win-equivalent shares cross — the leader changes;
+- where a leader's lower edge meets a candidate's upper edge — the stale band
+  changes;
+- where two head-to-head margins in one row cross — the gate takes a MINIMUM
+  over the band, so the controlling member switches there, and a minimum of
+  lines attains its maximum at exactly such a kink.
+
+The boundaries of BOTH baskets go into one set, because the activation rule
+reads both at the same weight and takes the smaller gain — solving each basket
+on its own grid answers a different question. Evaluating every boundary, and
+inside every resulting interval its midpoint plus both ONE-SIDED limits (the
+gain jumps where the leader changes, so a shared endpoint reports the wrong
+squad), plus the weights where the two baskets' gain lines cross — that is where
+`min(narrow, wide)` peaks when their slopes oppose — comes to 9,816 probe
+weights across every value:
+
+> **The largest gain reachable at ANY exposure weight is 0.00pp**, against a 5pp
+> threshold. At all 9,816 probe weights the swept optimum was already inside the
+> default band in at least one basket.
+
+That statement deliberately carries no significance claim, and does not need
+one: a gain that never reaches the threshold cannot clear the gate at any bar.
+Which also keeps the search honest — those thousands of data-derived weights are
+not in the Bonferroni count, so a "significant" result selected from among them
+would be a hypothesis nobody paid for. Reporting the maximum instead needs no
+correction at all. (If a rerun ever pushes that maximum ABOVE the threshold, the
+probe says so and stops claiming it: the satisfying set of the full predicate
+can be a narrow off-centre segment, and finding it would need its own boundaries
+solved and charged.)
+
+The boundaries are solved rather than sampled because sampling got this wrong
+twice on the way here. A 0.02 grid reported 0.00pp; solving only some of the
+boundary families reported 2.70pp; solving all of them but maximising per basket
+reported 2.87pp — and applying the veto at a common weight, which is what the
+gate actually does, brings it back to 0.00pp. Each of those was a different
+question, and only the last one is the gate's.
+
+`dom`, `draw` and `goals` are the WORST of the two baskets, deliberately — not
+the basket that supplied the gain. The gate asks "did anything get re-priced";
+those columns ask "at what cost to the field", and the second question is not
+scoped to the first one's basket.
+
+**A pass here would still not mean "schedule it."** The probe's best outcome is
+`CANDIDATE`, not `REPRICES`: track 3.1's gauntlet additionally requires a result
+to hold at growth buff 0 AND max, and to split the ACTUAL top-N live squads into
+two or more distinct best responses. Every squad here is tenure 0 and
+hand-built, so this probe structurally cannot run either check — it names them
+on every run instead of implying it did.
+
+**And the near miss is not a near miss.** FK x0.25 fails the gain gate outright —
+0.00pp, because the build it promotes was already inside the default band — and
+it would fail the draw band too, at +5.1pp against a +/-5pp limit. Its dominance
+is fine on the weighted season (+0.4 change, 0.4 absolute), which is worth
+stating precisely: an earlier version of this section read dominance off the two
+pure modes and reported +4.4, and that was an artefact of not weighting. The
+candidate that looked closest to shippable still fails, but on the gain and the
+feel band, not on concentration.
+
+### What the control calibrates
+
+The counter trigger is in this sweep to check that the bar rejects things, and
+under a paired test it is **not** inert: 3 of 40 narrow fixture deltas clear the
+bar, and 12 of 440 across the wide basket's two modes. They are small — and no
+value of it ever produced a gate, in either basket, at any exposure weight. That
+is the useful result:
+
+> **MARGINS means "detectable", not "worth shipping".** A knob that gates 3.9%
+> of goals reaches MARGINS at this sample size. The class that separates a lever
+> from a texture knob is the gate, not significance.
+
+It also supplies a floor. In the NARROW basket — same N, same ten fixtures on
+both sides, the only comparison here that is apples to apples — the largest
+movement this known non-lever produces is **2.0pp**. Anything a candidate does
+that is not comfortably above that is not distinguishable from what a knob
+nobody would ship already does:
+
+| Axis | Mode | Largest fixture move | vs the control's 2.0pp |
+| --- | --- | --- | --- |
+| FK occurrence | reg | 11.8pp | **6.0x** |
+| FK occurrence | ko | 8.8pp | **4.4x** |
+| `fair` | reg | 10.9pp | **5.5x** |
+| `fair` | ko | 10.9pp | **5.5x** |
+| Post play | reg | 4.8pp | 2.4x |
+| Post play | ko | 3.4pp | **1.7x** — the closest any candidate comes to the floor |
+
+The probe flags a row as indistinguishable from the control below **1.5x**, so
+post play in knockouts clears that line — but only just, and by less than any
+other candidate-mode pair in the table.
+
+The control is swept in BOTH baskets — that is what the gate needs — but this
+ratio is deliberately narrow-only on both sides. It compares a MAXIMUM over a
+family of cells, and the two baskets run at different N over different
+membership, so a cross-basket maximum grows with whichever family has more cells
+and less power rather than with whichever axis moves more. Mixing them swung
+these ratios by a factor of two between runs that changed no measurement.
+
+### Per axis, what actually moved
+
+**FK occurrence is the strongest axis, and what it does is CONCENTRATE.** From
+2.5% to 40% the regulation draw rate runs 57.9% -> 52.8% (shipped) -> 38.1% and
+goals per match 0.653 -> 0.779 -> 1.288. Every large delta at x4 has the same
+shape: the differentiated squads pull away from `flat-433` (`role-433` vs
+`flat-433` 68.4% -> 80.2% of the win-equivalent share). Turning it DOWN
+compresses the field instead — `flat-433`'s points per match rises 0.769 ->
+0.842 while every other squad's falls. So the axis is a skill-expression dial:
+more free kicks means more of the match decided by who invested, less means more
+of it decided by nothing. Turning it down therefore costs on both feel gates at
+once — more draws AND fewer goals — and in knockouts alone it also raises the
+leader's weakest margin from z 6.3 to 8.9, which points at concentration rather
+than rotation. Read that as the knockout-only reading it is: on the
+exposure-weighted season the same change is +0.4, well inside the direction
+screen, and the activation table above reports the weighted number.
+
+**`fair` is the only axis that ROTATES demand rather than amplifying it.**
+Raising it from 4.5% to 100% moves `shootonly-433` from 1.127 to 1.302 points
+per match in regulation and 1.370 to 1.518 in knockouts, while `role-433` and
+`def-433` stand still — more fouls means more free kicks and penalties, and
+those are converted by `shoot`. It carries the run's single reversal: `role-433`
+vs `def-433` goes from **47.5%** at the shipped default to **50.8%** at
+`fair 22`. Read that precisely — the DELTA is significant, the new level is a
+coin flip. The rule turned "the defensive lean is better" into "there is no
+difference", which is a re-pricing; it did not turn it into "the attacking lean
+is better".
+
+`fair` is also the only axis whose regulation dominance column does not move
+(+0.0). At the midpoint (`fair 11`, 50%) it still moves four fixtures
+significantly while taking the draw rate DOWN, 52.8% -> 48.6%, and goals up,
+0.779 -> 0.899 — the one cell in the whole sweep whose feel side-effects point
+in a direction anyone would ask for.
+
+**Post play is the weakest candidate.** Its largest move is 2.4x the control's
+in regulation and 1.7x in knockouts — the closest any candidate comes to a knob
+that gates 3.9% of goals. It needs x4 (60% of box entries) to reach even that,
+the feel barely registers (draw rate +4.3pp, goals −0.095), and across all eight
+of its wide-basket cells — four swept values in each mode — the top of the
+eleven-archetype table never changes.
+
+**The counter trigger never gates.** Swept from "counters never happen" to
+"every eligible turnover becomes one", goals per match moved 0.036, the basket's
+top never changed in either mode or either basket, and no cell produced a gate.
+**What rejects it is the compound activation gate, not the significance bar** —
+the bar finds its 3 narrow and 12 wide fixture movements perfectly real. That is
+the calibration: fixture-level significance is not evidence of a useful lever,
+and a gate worth believing when it accepts something has to reject this.
+
+### Does the answer move once the manager may re-optimise?
+
+The narrow basket's biggest weakness is that five archetypes may simply not
+contain the alternative a rule rewards. So EVERY swept value of every axis — the
+control included, so it is checked on the same surface as the candidates — is
+re-asked over the ELEVEN archetypes the round-robin publishes, 55 fixtures,
+4,000 matches each. That is 34 cells; the probe prints them all, and the shape
+is simple enough to state:
+
+Standings here are the **win-equivalent share** — the same 1 / 0.5 / 0 statistic
+the gate measures, so the leader they name is the squad the gate is then
+evaluated on. Ranking by league points instead is not a monotonic
+transformation of it when draw rates differ, and these constants move draw rates
+by fifteen points.
+
+| Mode | Shipped rule's top four | Last | Cells whose top changed |
+| --- | --- | --- | --- |
+| reg | `gkmin-433` .603, `def-433` .586, `def-532` .585, `role-433` .564 | `flat-433` .365 | **none of 16** |
+| ko | `role-433` .615, `gkheavy-433` .599, `def-433` .591, `def-532` .573 | `flat-433` .259 | **2 of 16**, both FK down |
+
+In regulation, `gkmin-433` leads every single cell — nothing moves the answer at
+any value of any axis. The two knockout exceptions are FK x0.25 and FK x0.5,
+which both promote `gkheavy-433` (+7.5pp z 4.7 and +6.0pp z 3.8 against
+`role-433`; 0.0pp against the tied band it belongs to). `flat-433` is last in 32
+of the 34 cells; the two it is not are FK x0.25 and x0.5 in regulation, where
+`stars-433` drops below it — that build wins and loses far more than it draws,
+so a metric that counts draws at half costs it more than a points table does.
+
+**That one axis-and-direction is the most interesting number in this section.**
+Read as a pair in both rules:
+
+> `gkheavy-433` vs `role-433`, knockout: **49.8% (z −0.3) at the shipped FK rate
+> -> 53.7% (z +4.7) at FK x0.25.** A coin flip becomes a significant win. Those
+> are shares; the +7.5pp above is the same reading as a win-rate advantage, which
+> is exactly twice the share advantage over 50%.
+
+Mechanically that is coherent: fewer free kicks means fewer set-piece goals,
+more matches level at full time, more shootouts, and the 29-point keeper the
+round-robin ranks sixth becomes the best build in the basket. It is a real
+re-pricing of the keeper budget by a rule.
+
+Read the "changed the answer" claim with the qualification the probe prints
+beside it: **the shipped knockout leader is not separated from `gkheavy-433`**
+either. So the honest description is not "the rule dethroned the leader" but
+"the rule turned a pair that was tied into a pair that is not" — which is still
+a re-pricing, and still the only one in the sweep.
+
+**It changes no advice in `SKILL.md`, and that is a decision rather than an
+oversight.** That file already refuses to settle the keeper total and says the
+knockout answer turns on shootout exposure. This sweep reaches the same place
+from the opposite direction — it moves a RULE instead of a squad, and the one
+thing that moved was the keeper budget in the mode where shootouts live. A
+second construction agreeing is a reason to trust the sentence that is there,
+not a reason to write a new one; nothing here is a fact an agent could act on,
+since none of these constants is scheduled.
+
+And it is worth **0.0pp at the activation gate** — in the mode with the LEAST
+exposure, since knockouts are cup-only and the cup runs once a day for the
+ladder's top 48. Two things reduce it from +7.5pp to nothing, and both are in
+the activation section above. `gkheavy-433` is inside the shipped rule's
+unresolved top band, so a manager could already have been on it for free. And
+the same cell moves the draw rate +5.1pp, outside the +/-5pp band a rule change
+is allowed to move it, so even a real gain there would have been refused.
+
+Incidentally, the shipped rows above are an 11-archetype round-robin at 4,000
+matches per fixture — roughly seven times the 300 seeds behind the published
+table — and they agree with it where that table says it is decidable:
+`gkmin-433` clear at the top of regulation, `role-433` at the top of knockouts,
+`flat-433` at the bottom of all but two. The middle band still reorders between
+the two runs, as that section says it must — and note that the bottom is where
+the metric matters: on a points table `stars-433` is above `flat-433`, and on
+the win-equivalent share the two swap in a couple of cells. The shipped `role-433` vs `def-433`
+fixture also independently reproduces the Skill's "solidity beats aggression"
+pair at 10,000 matches: the attacking lean takes 47.5% of the win-equivalent
+share in regulation and 53.1% in knockouts.
+
+**Do not read that 47.5% against the Skill's 56.5%** — they are different
+statistics on the same fact. The Skill quotes DECIDED matches, this section
+quotes the win-equivalent share, and half of all regulation matches are draws,
+so the share is compressed toward 50 exactly as the round-robin correction above
+warns. The knockout figures (53.1% here, 53.6% there) are directly comparable,
+because knockout draws are vanishingly rare — zero in the 5,440,000 knockout matches
+this run played, though not structurally impossible.
+
+### Degrees of freedom this sweep LEFT free
+
+Counting the degrees of freedom that were controlled is not enough; the keeper
+section above was wrong three times in one day and the cause was a different
+uncontrolled degree of freedom each time. So, explicitly:
+
+- **Manager response.** The basket is fixed; nobody re-solves a squad against
+  the swept rule. The wide-basket section shrinks this hole from five archetypes
+  to eleven but does not close it — eleven hand-built squads are a sample of the
+  legal space, not an argmax over it. Every verdict above is "…across these
+  baskets", and an axis that only rewards a build outside them reads as flat.
+- **Growth buffs.** Every squad is tenure 0. Production applies
+  `applyMatchGrowth`, which raises effective `total` and therefore `teamPow`,
+  and the zone press, the team-contribution terms and possession all read it.
+- **Set-piece slots.** `squad-lib` pins the FK taker to slot 9 and the penalty
+  taker to slot 10 for every candidate. The FK rows are the RATE axis only.
+- **Slot order.** Identical across the basket, which pins the broken zone-press
+  fallback to the same position in all of them. Held constant, not measured.
+- **Interactions.** One axis moves at a time. Nothing here says what FK x4 does
+  while `fair` is 22, and a pool built from two axes at once is untested.
+- **`cond`.** Fixed at 5 everywhere, as production does today. A form system
+  would invalidate every row.
+- **Opponent population.** Archetypes, not the live ladder — and production bots
+  are a single build, which is neither.
+
+Re-run after any engine change:
+
+```bash
+npx tsx packages/mcp/skill/reference/probes/constant-sweep.mts
+N=2000 WIDE_N=1000 npx tsx …   # faster; the bar is fixed, the detectable EFFECT moves
+```
+
 ---
 
 ## Note for the maintainers