pog-mcp 0.9.9 → 0.9.11

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pog-mcp",
-  "version": "0.9.9",
+  "version": "0.9.11",
   "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

@@ -1907,9 +1907,12 @@ runs from 1 to 5 depending on the core.
 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.
+Simulating that needs a recovery rate and a neutral point. **At this run's publication H's
+maximum stayed open.** Result 7e later crosses the selected F-2 curve with the maximum-role
+carrier and establishes the recovery prize; its first two-vector H treatment was later withdrawn
+under CAP=0. Result 7g closes the product value at `N=17` across every N17 registration in one
+immutable 18-asset owned universe;
+the figures below remain market-quality evidence, not that final bound.
 
 **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
@@ -2282,8 +2285,9 @@ 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.
+**0.2 was provisional when this run was published.** The role×condition interaction was then
+unmeasured and could move it either way. The 2026-09-06 follow-up below closes that gate for the
+pre-registered #728 role ladder and makes 0.2 final within that contract.
 
 ### The document already had this in its own table
 
@@ -2306,7 +2310,7 @@ five grid points give 17.5–18.0 and 26.6–27.4 respectively. (An earlier revi
 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
+### What this run did not cover at publication
 
 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
@@ -2320,14 +2324,228 @@ judgments AWAY from thresholds and reduce sensitivity. The axis can move 0.3 in
 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.
+### Role×condition and cup-depletion follow-up — the missing gates close (2026-09-06)
+
+`role-cond.mts` reused the six legal §3-15 squads over 36 ordered matchups and both engine modes.
+Scanning `computeBuff` directly under the product invariant `0≤tenure≤career`, through potential
+1 and through the `career=700` boundary after which every buff is zero, gives the actually
+reachable dense integer support **0..4**. The first maximum-4 witness is tenure/career **67/67**,
+and lower potential cannot escape that envelope. The growth
+axis uses four uniform endpoint combinations, five rotated heterogeneous vectors, and one
+reachable extreme with **the focal role slot at 4 and all other 21 players at 0**. Every slot of
+both XIs visits every value 0..4 across the rotated states; the final state directly covers the
+previously omitted input where the concentrated carrier has grown more than every teammate and
+opponent. This is a pre-fixed rotated/role-extreme sample, not an exhaustive enumeration of the
+per-player `5^22` combinations, and the guarantee below is limited to those ten growth states.
+From inputs alone, offense chose each focal
+squad's largest FW (`total → pass → shoot → slot`) as
+#728's pre-registered carrier. Defense swept **all ten outfield slots** rather than selecting one
+DF; the keeper stays out consistently with `appr-role.mts`'s contract that keeper A0 is outside
+roles. The engine was patched with #728's same per-player, per-scene `appr`/`apprDef` sheets;
+the **whole focal XI moved from condition 5.0 to 4.8** while keeping the role, and the opponent
+kept the shipped role table.
+
+Home/away and regulation/knockout were first joined within each seed cluster, with the latter
+mixed at canonical `KO_EXPOSURE=8.14%`. Six offensive and 60 defensive shards (six squads ×
+ten slots), **66 shards** in total, all paid against the unsharded simultaneous family
+**67,320 = 35,640 effects + 31,680 interactions versus ×1**. There are
+**3,960 = (6 offense + 60 defense) × 6 opponents × 10 growth states** structural cells. The
+final run used the independent seed column `role-cond-r37-final` and pre-fixed looks
+`50,000→100,000→200,000`. Each structural cell stopped at the first look where all nine effect
+intervals closed below 5pp. Both endpoints and all three looks were Bonferroni-priced at endpoint
+`δ=1.238e-7 = .05/(2×67,320×3)`, so the interaction intervals remain valid after optional
+stopping. **3,959 cells stopped at 50,000, one at 100,000 and none reached 200,000**, for
+**14,259,600,000 measured matches**. Empirical-Bernstein support widths are 400pp for effects and
+800pp for interactions. Each shard's 80 stock↔patch whole results and 80 ×1 role results were
+byte-identical. At the maximum pre-registered rung ×1024, the focal player's minimum nominal
+share among its non-zero draws is **98.71–98.91%** on offense and **94.03–99.15%** on defense;
+×1024 is therefore not called a uniform defensive “practical asymptote.”
+
+These are the cells with the largest absolute simultaneous upper bound among the 360 offensive
+and 3,600 defensive cells at every multiplier. All 360/360 offensive and 3,600/3,600 defensive
+cells at every row are established below 5pp.
+
+| role multiplier | offensive worst effect | offensive absolute upper | defensive worst effect | defensive absolute upper |
+| ---: | ---: | ---: | ---: | ---: |
+| 0 | +3.62 ±0.98pp | 4.60pp | +3.80 ±0.99pp | 4.79pp |
+| 0.25 | +3.62 ±0.98pp | 4.60pp | +3.85 ±0.99pp | 4.84pp |
+| 0.5 | +3.69 ±1.00pp | 4.69pp | +3.85 ±0.99pp | 4.84pp |
+| 1 | +3.65 ±1.00pp | 4.65pp | +3.86 ±0.99pp | 4.85pp |
+| 2 | +3.68 ±0.99pp | 4.67pp | +3.80 ±0.99pp | 4.79pp |
+| 4 | +3.75 ±0.98pp | 4.73pp | +3.81 ±0.99pp | 4.80pp |
+| 8 | +3.70 ±0.99pp | 4.69pp | +3.81 ±0.99pp | 4.80pp |
+| 32 | +3.84 ±0.99pp | 4.83pp | +3.89 ±1.01pp | 4.90pp |
+| **1024** | **+3.97 ±1.01pp** | **4.98pp** | **+3.95 ±1.01pp** | **4.96pp** |
+
+The global ×1024 witnesses are `352/stars/flat` slot 9 versus `433/even/flat`, growth
+`rotated-1`, `n=50,000` on offense, and `442/stars/atk` slot 1 versus `433/even/flat`, growth
+`rotated-4`, `n=50,000` on defense. Against ×1, offense has **5 established increases / 8 established
+decreases / 2,867 undecided interactions** (2,880 total), while defense has
+**42 / 20 / 28,738** (28,800 total). The largest positive upper endpoint is
+**+1.82 ±1.49pp = 3.31pp** on offense and **+2.23 ±1.41pp = 3.64pp** on defense. The most
+negative lower endpoint is **−1.62 ±1.47pp → −3.09pp** and **−1.55 ±1.47pp → −3.02pp**,
+respectively. Every effect and interaction interval endpoint over the whole ladder is inside
+5pp. **The 0.2 band ceiling is final for this role contract.**
+
+The scope is the focal side's **pre-registered offensive carrier, every outfield defensive
+slot, the whole-XI 0.2 band and the ten pre-fixed growth states**. The exhaustive `5^22`
+growth-vector product, opponent role sheets, other offensive slots, a keeper defensive role
+and continuous multipliers outside the ladder are not covered. Since roles do
+not exist in the product yet, #735 must adopt this measured ladder or a subset; expanding that
+space requires a new gate.
+
+The 2026-08-29 qualifier premise is also **withdrawn**. The generic
+`tournament-runner.ts` `runTournament()` does run qualifiers when given more than 48 teams, but
+that file explicitly says production cups do not use it. The production daemon combines ranked
+entrants with fill-bots into exactly 48 teams; registration replay and the legacy path also slice
+to 48 before building the group schedule. A champion therefore plays at most three group plus
+five knockout matches. There is no shipped eleven-match path and no three-match qualifier-luck
+gap to calibrate.
+
+The corrected product criteria cover the shipped eight-match path. In the verified production
+entry population plus explicit persistent-new, daily-fill-bot and observed-extreme stresses,
+every pre-registered offensive carrier and every non-keeper defensive role must finish with
+`cond8>4.8`; independently of entry state, that cup itself must consume at most 0.2 condition.
+Each cell passes only when the simultaneous upper bound on each violation probability is below
+1%. The first is an empirical population gate, not a universal floor for arbitrary future
+history: F-2 has neither a load cap nor a cup-entry reset, and its curve asymptote is below 4.8.
+
+#### Withdrawn scale-only result — 0.375 is historical, not the product value
+
+This result remains as an audit trail. It applied each scale to prior cup history and the new
+eight-match path, but kept the neutral point, `L₀`, and safe amplitude from result 7c, whose
+lifetime integration treated cup events at ×1. Since the fold changes the whole EWMA history,
+that was not a self-consistent product combination. Its conditional 0.375 pass below is superseded
+by the re-integrated two-dimensional boundary that follows.
+
+`engagement-load.mts` integrated the canonical production export with SHA-256
+`1c7335ba9fa11388363de57628207ecc4820dc9bb02285066bda849adf571b6f` at τ=24h on the
+claim clock under predicate D. Taking each human team's first pre-match vector in each cup after
+the 2τ warm-up yields **108 profiles across nine cups**. The hand-off file has SHA-256
+`dfd3032c9bab22c42c0616e708705e2f9a187fde719766af37cc68cc6c60515e`, and a fresh export
+from the current code was byte-identical. It preserves ladder and raw-cup components separately:
+candidate `s` enters at **`L=ladderL+s·cupRawL`**, so already-accumulated cup history is scaled too
+rather than leaving old cups at ×1 while scaling only the new cup. At `s=1`, entry-slot L has
+median **4.959667**, p95 **10.902379**, and maximum **33.888645**.
+
+Form was not deployed in that production window. The historical `cupRawL` column is therefore the
+growth-zero workload on the committed fixture ledger while condition was fixed at 5, not a
+closed-loop equilibrium in which an earlier candidate scale had already changed condition and then
+changed later engagement draws. Scaling that fixed historical column is exact for its event counts;
+only the new eight-match path below directly replays condition-to-engagement feedback. This is
+another reason the entry result is a fixed-basket empirical gate rather than a universal future-state
+claim.
+
+`cup-role-load.mts` crosses persistent-new load zero, the daily fill-bot neutral load, the observed
+population, a role-slot maximum and a whole-XI condition-deficit maximum in six entry scenarios.
+It measures all six offensive carriers and all 60 non-keeper defensive slots at ×1024. Six
+opponent squads × home/away × six entry scenarios produce **4,752 cells per scale**, each with
+3,000 independent paths through the shipped three group plus five knockout matches. Every match
+adds both XIs' growth-zero F-1 predicate-D raw engagements, with no inter-round recovery. The
+repeated synthetic opponent keeps the shipped role sheet while carrying its own accumulated load.
+
+The **16-point** grid is
+`{1, .5, .45, .4, .375, .35, .325, .3, .275, .25, .225, .2, .125, .1, .075, .05}`. Measuring
+the six largest candidates established 0.375 as passing and rejected every larger value, so the
+ten smaller points are irrelevant to selecting the largest pass. Every shard still paid against
+the full family **152,064 = 16 scales × 4,752 cells × 2 criteria**, using one-sided
+Bernoulli-KL/Chernoff `α=3.288089e-7`. That is **114,048,000 measured matches per scale** and
+**684,288,000 measured matches** total. Each shard also passed 80 stock↔patch involvement-result
+and 880 ×1-role byte-identity controls. The independent seed column is `cup-role-r36-final`.
+
+The two explicit product criteria require each cell's simultaneous upper bound on (1)
+`cond8≤4.8` and (2) `initialCond−cond8>0.2` to be **below 1%**. The first asks for an absolute
+floor in the observed entry basket; the second isolates this cup's band consumption from entry
+state. A sampled minimum or maximum may cross its threshold while the scale passes—the gate is on
+violation probability, not deterministic zero violations.
+
+| consumption scale | worst floor violations (simultaneous upper) | worst cup-drop violations (simultaneous upper) | min initial/cond8 | max drop | raw8 mean/max | verdict |
+| ---: | ---: | ---: | ---: | ---: | ---: | --- |
+| 1.000 | 3,000/3,000 (100.00%) | 2,996/3,000 (100.00%) | 4.897/4.692 | 0.330 | 70.3/148 | FAIL |
+| 0.500 | 405/3,000 (17.15%) | 139/3,000 (7.04%) | 4.911/4.770 | 0.238 | 70.2/148 | FAIL |
+| 0.450 | 67/3,000 (4.04%) | 21/3,000 (1.88%) | 4.912/4.780 | 0.224 | 70.2/148 | FAIL |
+| **0.400** | **6/3,000 (1.02%)** | 2/3,000 (0.72%) | 4.914/4.791 | 0.209 | 70.2/148 | **FAIL** |
+| **0.375** | **1/3,000 (0.63%)** | **1/3,000 (0.63%)** | **4.915/4.797** | **0.201** | **70.2/148** | **PASS** |
+| 0.350 | 0/3,000 (0.50%) | 0/3,000 (0.50%) | 4.915/4.803 | 0.192 | 70.2/148 | PASS |
+
+At 0.375, the floor witness is defensive slot 6, `433/stars/atk` versus `433/even/flat`,
+home, `observed-slot-max/team-max`; the cup-drop witness is defensive slot 4,
+`433/even/flat` versus `442/even/def`, away, `persistent-new/new`.
+
+That run therefore selected **0.375 per raw F-1 predicate-D engagement conditional on the old
+curve**. It is not the current product decision; the self-consistent rerun below replaces it.
+
+Inter-match time is zero in this declared zero-gap product stress distribution. It includes actual
+entry history and observed extremes, but remains an empirical guarantee for this fixed nine-cup
+basket. Without a load cap or cup-entry reset, it cannot guarantee an absolute floor for every
+possible future history—the curve's asymptote itself is below 4.8. Nor does it mathematically
+dominate changing real-bracket opponents or opponent role sheets; adding those spaces to the
+product contract requires re-gating.
+
+#### Final re-integrated boundary — `s=0.1375`, `b=0.4914` (2026-09-06)
+
+To remove that feedback error, the canonical lifetime timeline was re-integrated from scratch
+with structural cup events at `s=0.1375` and ladder events at 1. At τ=24h the active-appearance
+median is **1.146304**, p95 is **14.756304**, and the 4× curvature scale is therefore
+**`L₀=59.025216`**. The all-τ safe-amplitude search and an independent selected-curve
+confirmation pass **`b=0.4914`**. The neutral offset is derived, not copied from rounded prose:
+**`a=b·L_neutral/(L_neutral+L₀)=0.009361468442213193`**. Zero load maps to
+5.00936146844 and the asymptote to 4.51796146844. These are the canonical result-7d and F-2-v2
+constants.
+
+The same 108 profiles across nine cups keep ladder and raw-cup components separate, and each
+candidate enters as `L=ladderL+s·cupRawL`. At the selected scale, the 1,188 entry slots have
+minimum 0, median **0.698231**, p95 **3.050740**, maximum **26.617183**, and mean
+**1.264655**. The remaining cup contract is unchanged: six offensive carriers plus 60 outfield
+defensive slots, six opponents, home/away and six entry scenarios make 4,752 cells per scale;
+each cell runs 3,000 independent eight-match paths and accumulates both XIs' growth-zero
+predicate-D raw engagements.
+
+The pre-fixed scale grid is
+`{1, .5, .4, .3, .25, .2, .175, .15, .1375, .125, .1125, .1, .0875, .075, .0625, .05}`.
+Every point above 0.1375, the selected point, and the next point 0.125 were actually run—no
+monotonicity inference substitutes for a shard. The amplitude family was also pre-fixed at
+`{0.4914, 0.5026}`, and both were run at the scale boundary. The global family is
+**304,128 = 2 amplitudes × 16 scales × 4,752 cells × 2 criteria**, with one-sided
+Bernoulli-KL/Chernoff `α=1.644045e-7`. Each shard measures 114,048,000 matches; ten scale shards
+plus one amplitude-neighbour shard total **1,254,528,000 matches**. Every shard's 80
+stock↔patch involvement and 880 ×1-role results were byte-identical. The independent seed column
+is `cup-role-r44-final`.
+
+| amplitude `b` | cup `s` | worst floor violations (simultaneous upper) | worst cup-drop violations (simultaneous upper) | min initial/cond8 | max drop | raw8 mean/max | verdict |
+| ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- |
+| 0.4914 | 1.000 | 3,000/3,000 (100.00%) | 3,000/3,000 (100.00%) | 4.830/4.642 | 0.346 | 70.3/149 | FAIL |
+| 0.4914 | 0.500 | 3,000/3,000 (100.00%) | 1,969/3,000 (70.36%) | 4.845/4.702 | 0.267 | 70.3/149 | FAIL |
+| 0.4914 | 0.400 | 3,000/3,000 (100.00%) | 299/3,000 (13.30%) | 4.848/4.721 | 0.239 | 70.3/149 | FAIL |
+| 0.4914 | 0.300 | 2,978/3,000 (99.83%) | 1/3,000 (0.65%) | 4.851/4.744 | 0.205 | 70.3/149 | FAIL |
+| 0.4914 | 0.250 | 2,665/3,000 (91.78%) | 0/3,000 (0.52%) | 4.853/4.758 | 0.184 | 70.3/149 | FAIL |
+| 0.4914 | 0.200 | 921/3,000 (35.53%) | 0/3,000 (0.52%) | 4.855/4.773 | 0.159 | 70.3/149 | FAIL |
+| 0.4914 | 0.175 | 176/3,000 (8.57%) | 0/3,000 (0.52%) | 4.855/4.782 | 0.145 | 70.2/149 | FAIL |
+| 0.4914 | **0.150** | **9/3,000 (1.24%)** | 0/3,000 (0.52%) | 4.856/4.791 | 0.130 | 70.2/149 | **FAIL** |
+| **0.4914** | **0.1375** | **1/3,000 (0.65%)** | **0/3,000 (0.52%)** | **4.857/4.796** | **0.122** | **70.2/149** | **PASS** |
+| 0.4914 | 0.125 | 0/3,000 (0.52%) | 0/3,000 (0.52%) | 4.857/4.800 | 0.113 | 70.2/149 | PASS |
+| **0.5026** | **0.1375** | **9/3,000 (1.24%)** | 0/3,000 (0.52%) | 4.853/4.791 | 0.125 | 70.2/149 | **FAIL** |
+
+At the selected point, the floor witness is defensive slot 6, `433/stars/atk` versus
+`442/stars/atk`, home, `observed-slot-max/team-max`; every cell has zero cup-drop violations.
+Thus **`s=0.1375`, `b=0.4914` is a passing boundary point in the pre-fixed two-dimensional
+grid while its upper and right neighbours fail**. It is not claimed as a continuous or global
+optimum. Points below 0.125 were unnecessary to select the largest passing grid point and were
+not run; no pass is inferred for them by monotonicity.
+
+F-1 continues to retain integer raw D counters. F-2 v2 alone reads source-match structural
+identity (`tournament_id IS NOT NULL`) and applies 0.1375 to cups and 1 to ladder events, through
+one helper shared by current accumulation and historical rebuild. The fixed population, zero-gap,
+opponent-role, and changing-bracket limitations are the same as in the historical run above.
+
+## What a same-sum vector transfer is worth (`attr-vs-cond.mts`) [historical input]
+
+This probe was run while F-6 still proposed letting a registration redistribute each player's
+four attributes **within that player's own sum**. Two equal-sum core players could then submit
+each other's vectors and move the strong build to a rested ID. The intermediate §3-12 design
+slowed that transfer with a per-registration L1 cap. Result 7e later measures the missing distance
+family and rejects every effective positive integer cap, choosing **`CAP=0`**. Production F-6
+therefore permits no vector redistribution; this section remains evidence for why.
 
 ### Method
 
@@ -2414,12 +2632,13 @@ 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
+### What this could not 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.
+One squad family, one tired/fresh gap (4.5 vs 5.5), and one cross-position pair could not choose
+the cap. It also did not measure how often rejecting edits blocks a legitimate respec. Result 7e
+adds all 66 maximum roles and exact even distances 2–32, finds a distance-2 witness already worth
+11.70pp [10.50, 12.89], and selects `CAP=0`. The distribution and UX value of honest requested
+edits remain unmeasured; the product deliberately chooses the closed fairness boundary anyway.
 
 ## What is one point of condition worth? (`cond-effect.mts`)
 
@@ -2553,15 +2772,20 @@ product choice. Predicate C versus D was also open at publication and F-1 subseq
 **D** on 2026-09-04. The C-conditioned tables and correction trail below remain historical evidence;
 they are not relabelled as D.
 
-**Product decision, 2026-09-06 — F-2 reference curve.** Re-integrating the same owner-only export
-and stored seed column under D, then independently validating only the selected 4× curve, chooses
-**`τ=24h`**, active-appearance neutral load **`L_neutral=8.076776`**,
-**`L₀=90.717176`** (4× D p95 `22.679294`), **`b=0.5399`**, and therefore
-**`a=b·L_neutral/(L_neutral+L₀)=0.044139`**. The versioned equation is
-`cond=5+a−bL/(L+L₀)`. This does not activate fatigue: F-2 first lands behind a default-OFF flag,
-and G-3 can flip only after the maximum legal role×condition interaction, 8–11-match cup depletion,
-the launch-amplitude prize/cost ordering, H ceiling, and I-c CAP clear against this same curve.
-The D product rerun is recorded below; the older C run remains the audit trail.
+**Product decision, 2026-09-06 — F-2 v2 reference curve.** Re-integrating the same owner-only
+export and stored seed column under D, with every structural cup event folded at 0.1375 and every
+ladder event at 1 throughout the lifetime timeline, chooses **`τ=24h`**, active-appearance neutral
+load **`L_neutral=1.146304`**, **`L₀=59.025216`** (4× D p95 `14.756304`),
+**`b=0.4914`**, and therefore
+**`a=b·L_neutral/(L_neutral+L₀)=0.009361468442213193`**. The versioned equation is
+`cond=5+a−bL/(L+L₀)`. This does not activate fatigue: F-2 remains behind a default-OFF flag.
+The §3-15 follow-up closes #728-ladder role×condition and the shipped eight-match cup at the
+pre-fixed two-dimensional boundary `s=0.1375, b=0.4914`. Result 7e then closes the
+launch-amplitude focused-role recovery prize and I-c at `CAP=0`; result 7g closes H at `N=17`
+by comparing every admissible N17 registration against N18 in one immutable owned universe.
+F-2 stays default-OFF until
+F-3/F-4/F-6/J and G-3 are implemented. Result 7d records
+the current rerun; the older C run and cup×1 result 7c remain audit trails.
 
 ### F-1 implementation status — predicate D is now durable
 
@@ -3424,7 +3648,11 @@ rule-of-three bound was bolted on for it. Under empirical Bernstein the range te
 `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.
 
-### Post-F-1 product rerun — certify only the selected D · 4× curve
+### Withdrawn result 7c — the D · 4× · cup×1 curve
+
+> **This is not the current product curve.** It correctly used D after F-1, but did not feed the
+> subsequently selected cup fold back through lifetime integration. The figures and then-current
+> rationale remain as an audit trail; self-consistent result 7d supersedes them.
 
 Results 6 through the safe-amplitude section above preserve the historical predicate-C run and
 its joint 1×/2×/4× candidate family. The product values were not obtained by scaling that table.
@@ -3462,15 +3690,269 @@ The product choices are:
   condition 5. Zero load maps to 5.044, D p95 to 4.936, this window's observed maximum to 4.871,
   and the asymptote to 4.504. At `b=0.5399`, the independent worst upper bound is 3.69pp across
   all 18 actual cells.
-- **Prize versus cost remains unresolved.** At the safe amplitude's maximum recoverable deficit
+- **Prize versus cost was unresolved at this revision.** At the safe amplitude's maximum recoverable deficit
   0.173, the prize is 0.19pp [−0.07, 0.45], the selected cheapest one-point replacement costs
-  0.16pp, and paired cost−prize is −0.03pp [−0.50, 0.43]. Together with the unmeasured legal
-  role×condition and cup extremes, that keeps F-2 default OFF until the follow-up gates and G-3.
+  0.16pp, and paired cost−prize is −0.03pp [−0.50, 0.43]. The role×condition and cup extremes
+  now clear in the §3-15 follow-up, but this ordering was not measured directly on the practical-
+  asymptote carrier. At result 7c, that focused prize/cost gate, H, CAP, the cup fold and G-3 kept
+  F-2 OFF; results 7d/7e/7f later close every item in that list except G-3.
 
 The canonical constants are therefore `τ=24h` and
 `cond = 5 + 0.044139 − 0.5399·L/(L + 90.717176)`. Implementation derives `a` from the other
 three constants so documentation rounding cannot move the exact neutral point.
 
+### Result 7d — structural-cup-folded D · 4× · v2 product curve (current decision)
+
+Applying 0.375 only after result 7c left prior cup history at ×1 inside the curve while giving
+new cups a different unit. Result 7d removes that inconsistency by rebuilding the canonical
+lifetime timeline from scratch with every structural cup event at 0.1375 and every ladder event
+at 1. It uses the same export SHA-256
+`1c7335ba9fa11388363de57628207ecc4820dc9bb02285066bda849adf571b6f`, predicate D,
+only the 4× curve, and all-τ seed column `c2-cupfold-r45-final`. The joint all-τ family is
+**612**, the sum of each τ's actual-cell upper bound times six opponents.
+
+| τ(h) | D `L` median | D `L` p95 | 0.2-derived `b` (4×) | one match (pp)\* | cells over | searched safe-`b` interval | worst upper at low end | established-over `b` |
+| ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| 6 | 0.38 | 4.01 | 8.695 | 1.6454 | **16/17** | 0.4076–0.4161 | 3.68pp | 4.3474 |
+| 12 | 0.57 | 7.54 | 10.753 | 1.1750 | **10/15** | 0.4725–0.4830 | 3.79pp | 5.3765 |
+| **24** | **1.15** | **14.76** | **10.498** | **0.6147** | **3/18** | 0.4716–0.4819 | 3.64pp | 5.2492 |
+| 48 | 2.34 | 26.06 | 9.102 | 0.3087 | **4/18** | 0.5156–0.5244 | 3.67pp | 4.5511 |
+
+\* “One match” is the secant at each row's 0.2-derived amplitude with raw p95 increment 4, not
+the cost at the selected product amplitude. The interval's left endpoint is the largest grid
+point established below; its right endpoint is merely the smallest point that this run did not
+establish below. It is neither a failure nor an upper threshold. Only the final column, where a
+lower bound exceeds 5pp, establishes that the threshold lies below a candidate.
+
+τ is still a product choice, not an estimate fitted to observed fatigue. The 24h derived curve
+has the fewest over-gate cells in this price table, while 48h halves the one-match secant again,
+0.6147→0.3087pp. The product therefore retains **τ=24h** as its daily-rotation/fairness
+compromise. Its unrounded active-appearance median is **1.146304**, p95 is **14.756304**, and
+observed maximum is **32.07**, giving **`L₀=59.025216`** at 4×.
+
+The joint all-τ search prices the four-row table. Its right endpoint is not an upper bound and
+does not reject `b=0.4914`. After the cup-boundary diagnostic, that exact candidate was fixed
+without looking at the final result and checked on independent seed column
+`c2-cupfold-r42-product` under the selected τ24 curve's **family 156**: **18/18 cells were
+established below**, with worst upper **3.48pp** and maximum deficit **0.173**. The independent
+§3-15 `cup-role-r44-final` two-dimensional family then rejects the right neighbour `b=0.5026`
+at `s=0.1375`. Thus 0.4914 is a passing pre-fixed product-grid point, not a continuous threshold
+estimate.
+
+The current constants are:
+
+- **`τ=24h`, `L_neutral=1.146304`, `L₀=59.025216`, `b=0.4914`.**
+- **`a=b·L_neutral/(L_neutral+L₀)=0.009361468442213193`.** Neutral maps exactly to 5;
+  zero load maps to 5.00936146844, D p95 to 4.911081, the observed maximum to 4.836364, and
+  the asymptote to 4.51796146844.
+- Under the same curve, the 108-profile cup-entry p95 maps to 4.985211 and its maximum load to
+  4.856637.
+- The equation is **`cond = 5 + a − 0.4914·L/(L + 59.025216)`**; implementation derives `a`
+  from the other three constants rather than copying rounded prose.
+
+In the general-carrier diagnostic, a 0.173 deficit gives prize 0.26pp [0.04, 0.48], cost 0.10pp,
+and paired cost−prize −0.16pp [−0.53, 0.21], still unresolved. It does not replace the
+focused maximum-role carrier gate. Result 7e closes the recovery prize and I-c, but its first H
+cost treatment writes a second incumbent vector and is invalid under CAP=0; result 7g supplies the
+corrected same-owned-universe measurement.
+
+### Result 7e — launch role economics and I-c `CAP=0` (H cost verdict withdrawn, 2026-09-07)
+
+`role-economics.mts` imports the result-7d product function directly and fixes the exact treatment
+at observed maximum `L=32.07`: **fresh 5.009361468442 versus tired 4.836364480324**, a deficit of
+**0.172996988118**. It uses the same six legal squads as §3-15, six offensive carriers plus 60
+non-GK defensive slots, the maximum pre-registered role multiplier ×1024, and a uniform 60-cell
+basket of six opponents by ten growth states. One observation first groups home/away ×
+regulation/knockout for one independently sampled basket cell, then applies knockout exposure
+**8.14%**.
+The unit is §3-15's win-rate-advantage pp, `win-equivalent-share difference × 2`, so one
+observation lies in [−200, 200]pp and the support width is 400pp.
+
+Candidate selection uses **5,000** clusters from `role-economics-r1-select`; independent
+publication uses `role-economics-r1-final` with **200,000** economics clusters and **50,000** CAP
+clusters. For each role the selector exhausts **56–160** legal one-point-weaker replacements and
+**83–633** legal same-sum vectors, then confirms the top four economics candidates and top four
+high/low pairs at every exact even L1 distance on the publication column. The simultaneous family
+is **4,818 = 66×73**, each endpoint has `δ=5.188875e-6`, and the bounded empirical-Bernstein
+support width is **400pp**. Minimum–maximum context counts are **53–119 / 3,087–3,553 /
+741–937** for selection/economics/CAP, with all 60/60 cells present for every role. The 80
+stock↔patch and 80 ×1-role identities per shard are all byte-identical, **5,280** of each in total.
+The run replays **1,477,620,000 matches**. SHA-256 of the 66 sorted `RESULT_JSON` records is
+`ea666c7e275a6559b2aa02db77f8069e618821b439d9a9bf518166f12e5c381f`.
+
+**The valid part — the recovery prize is real.** The lower endpoint of the fresh-versus-tired
+prize is above zero in 66/66 roles. R1 then chose, per role, the finalist with the largest
+publication-column net `fresh one-point-weaker − tired anchor` and produced the following table.
+Its simultaneous-family accounting is sound, but the physical treatment does not implement H;
+the table is retained as audit history and is not a product verdict.
+
+| role | prize point-estimate range | fresh replacement better | tired starter better | unresolved |
+| --- | ---: | ---: | ---: | ---: |
+| offense 6 | 0.44–0.84pp | **5** | 0 | 1 |
+| defense 60 | 0.28–1.15pp | **5** | 38 | 17 |
+
+All finalists belong to the global simultaneous family, so choosing within the publication column
+does not invalidate their intervals. The largest reported value was
+`433/stars/atk|offense|slot8`, candidate `defense→slot1.defense`: prize **+0.44±0.16pp**,
+effect of the one-point-weaker role player plus the compensating point elsewhere
+**−1.08±0.37pp**, hence fresh replacement net **+1.52±0.37pp [1.15, 1.88]**. However,
+`costCandidates` lowers the role player's total by one and simultaneously raises a different
+incumbent's persisted vector and total by one. Adding asset 18 cannot perform that second write when
+`CAP=0`. This is a counterfactual two-vector XI effect, not the N17→N18 marginal slot, so its use as
+H evidence is **withdrawn** [Codex P1, 2026-09-07]. Result 7g owns the corrected H verdict.
+
+**I-c — the assumption that nearby vectors are not worth changing is refuted.** The `ĝ(d)` rows
+below choose the largest publication mean among the four independently selected pairs. Selection
+error can miss the true argmax, so these are holdout-confirmed existence witnesses in the lower-
+bound direction, not mathematical upper bounds over the whole space. That direction is sufficient
+to reject a CAP.
+
+| L1 distance d | holdout `ĝ(d)` | simultaneous interval | roles established over 5pp |
+| ---: | ---: | ---: | ---: |
+| 2 | **11.70pp** | [10.50, 12.89] | **64/66** |
+| 4 | 22.81pp | [21.35, 24.27] | 66/66 |
+| 6 | 33.75pp | [32.16, 35.35] | 66/66 |
+| 8 | 44.21pp | [42.54, 45.88] | 66/66 |
+| 10 | 53.74pp | [52.01, 55.46] | 66/66 |
+| 12 | 63.19pp | [61.44, 64.94] | 66/66 |
+| 14 | 70.59pp | [68.87, 72.32] | 65/65 |
+| 16 | 77.78pp | [76.06, 79.50] | 62/62 |
+| 18 | 76.64pp | [75.09, 78.20] | 50/50 |
+| 20 | 76.61pp | [75.04, 78.18] | 42/42 |
+| 22 | 76.77pp | [75.19, 78.35] | 39/39 |
+| 24 | 76.85pp | [75.27, 78.43] | 36/36 |
+| 26 | 77.11pp | [75.53, 78.69] | 28/28 |
+| 28 | 77.18pp | [75.60, 78.77] | 24/24 |
+| 30 | 77.22pp | [75.63, 78.80] | 17/17 |
+| 32 | 55.61pp | [54.02, 57.20] | 2/2 |
+
+The distance-2 maximum witness is `433/stars/atk|defense|slot2`,
+`(1,6,1,4)→(2,6,1,3)`; both are legal same-sum vectors. Two is the minimum non-zero L1 distance
+for a sum-preserving integer edit, so every effective positive integer `CAP≥2` admits this
+11.70pp move in one registration. `CAP=1` admits no change and is behaviourally equal to zero.
+Waiting more weeks or reasoning from maximum distance cannot rescue a positive CAP.
+
+The explicit product value is therefore **`CAP=0`**. The four attributes fixed at creation or
+acquisition never change in later weekly registrations. This deliberately gives up honest respecs;
+it does not freeze pool membership, mutable positions/names for unminted players, or FK/PK choices.
+F-6 must enforce one shared vector-identity gate immediately before every squad write. This result
+closes C-2's recovery prize and I-c; result 7g below closes H. F-2 and the player pool remain
+default-OFF until F-3/F-4/F-6/J and G-3 are implemented.
+This is an existence result over the pre-fixed uniform 60-cell **mean** with shipped-default
+opponent role sheets. It does not bound crossed opponent-role sheets or the global engine space,
+but the positive allowed cases needed to reject every effective positive integer CAP already exist
+inside this scope.
+
+### Result 7f — first immutable actual-pool N17⊂N18 correction (H evidence withdrawn, 2026-09-07)
+
+The Codex P1 reproduces exactly: result 7e's H candidate changes two incumbent vectors, so it is
+not an action that adding one asset can perform under the selected `CAP=0`. The corrected
+`MODE=pool-ceiling` directly imports and asserts
+`WEEKLY_PLAYER_POOL_PRODUCT_SIZE=17` and `WEEKLY_PLAYER_VECTOR_L1_CAP=0` before it runs.
+
+The next Codex P1 also reproduces exactly. This run compares only one pre-selected N17 with N18.
+An owner holding the same 18 assets can register core eleven, the new fresh focal and any five low
+clones at N17, so that N17 already has both outcomes from this run. The exhaustive counts, values
+and hashes below remain an exact audit of those two specific pools, but they are **not the marginal
+effect of registration slot 18 or evidence for H. That attribution is withdrawn; result 7g owns
+the current verdict.**
+
+The pre-fixed witness is `433/stars/atk|offense|slot8`. N17 consists of the legal source team's
+core eleven plus six distinct minted assets carrying the source's legal total-12 DF vector. N18
+preserves those 17 complete rows as a prefix and adds only a new minted asset with the tired slot-8
+player's same FW position, total 29 and `(pass,dribble,shoot,defense)=(6,9,9,5)` vector. No acquired
+name, position, attribute or designation is rewritten; the role sheet and condition are
+match-time tactic/derived values. N17 and N18's preserved-prefix SHA-256 are both
+`a43d28b29ee2474e2233f4f54315c3198d2a3d18f45ddee1834fc4761f3e1f9e`; full N18 is
+`147a312a0b3932e41f6530b19170301ef1037d4d215365ccc41458e03b7985cc`.
+
+The probe exhausts `C(17,11)` and `C(18,11)` and sends every 212-point subset through the engine
+validator. N17 has **210 legal subsets**, all containing the tired focal, and **one** competitive
+projection. N18 has **420/two**; **210** use the new asset instead of the tired focal and form the
+second outcome. The projection contains position, attrs, total, condition, FK/PK and both role
+sheets, excluding only asset identity and display names, which do not enter the winner calculation.
+The unchanged N17 outcome remains available in N18, so the treatment is one fresh playerId rather
+than an incumbent rewrite.
+
+The independent `role-economics-r2-pool-final` publication has **200,000 clusters** and covers all
+60 contexts **3,188–3,481 times per cell**. Each observation is the same home/away × REG/KO group;
+support width is 400pp and the pre-fixed family of one uses two endpoint `δ=0.025` bounded
+empirical-Bernstein intervals. Across **1,600,000 matches**, the N18-only fresh outcome gains
+**+0.431679±0.078950pp [0.352728, 0.510629]**. Two separate processes reproduce byte-identical
+numbers and `POOL_RESULT_JSON`; its SHA-256 is
+`48615a458ee8e5696f6259bf8b9ff5d48056be96e50509db9f3d2a3abfcde606`.
+N17 has one competitive outcome and N18 has only that outcome plus the fresh one, so the positive
+simultaneous lower endpoint also establishes the fresh arm as N18's best on this pre-fixed mean.
+
+The positive difference is valid between the pre-selected N17 and N18 above. A better N17
+registration contains both outcomes, however, so this value alone cannot select H's product maximum.
+
+```bash
+MODE=pool-ceiling POOL_SEEDS=200000 CONTROL_SEEDS=20 \
+  POOL_COLUMN=role-economics-r2-pool-final \
+  pnpm --filter pog-mcp exec tsx skill/reference/probes/role-economics.mts
+```
+
+### Result 7g — every N17 registration versus N18 in one owned universe, H `N=17` (current product decision, 2026-09-07)
+
+The second correction fixes the owned universe at 18 assets first and changes only the registration
+ceiling. The eleven core-XI IDs are mandatory on both sides. The optional assets are seven distinct
+minted clones byte-identical to the seven core vectors in slots 4–10. Their totals are
+**10, 13, 16, 19, 27, 28 and 29**; all are FWs, creating seven pre-fixed maximum offensive roles at
+×1024. The other four players force GK/DF/DMF/OMF. The XI sums to 212 and exactly saturates
+`TEN_MAX=3` and `EIGHT_MAX=5`. SHA-256 of the 18-asset owned universe is
+`a62626d300ea675e82f61c11b06ef9b15577ad52bc173cf3bdf3b73d15adb73b`.
+
+Because core eleven is mandatory, the maximal N17 registrations are **all seven** choices that omit
+one clone. Any smaller registration is weakly dominated by one of these because adding an owned
+asset cannot remove a legal XI. The probe exhausts each N17's `C(17,11)` and N18's `C(18,11)`, then
+sends every 212-point candidate through the actual engine validator. Each N17 has **293–385**
+212-point candidates and exactly **64 legal XIs**; N18 has **876/128**. Every legal XI contains the
+four structurally mandatory cores and exactly one member of each of the seven core/clone pairs, and
+has one base competitive outcome after excluding asset identity and display name. In each focused
+role, N18 has tired/fresh outcomes; only the N17 missing that role's clone has one tired outcome,
+while the other six N17 registrations have both. The probe exhaustively asserts this **7×7
+frontier**, removing any opportunity to manufacture the gain by selecting a weak N17.
+
+After a defensive pilot, the more direct seven offensive roles were fixed on a separate pilot
+column. The independent `role-economics-r3-pool-final` publication hashes one role and one of the 60
+opponent×growth cells uniformly per cluster, then replays that role's fresh/tired arms on identical
+home/away × REG/KO seeds. It has **1,000,000 clusters and 8,000,000 matches**, covering all 420/420
+role×context cells **2,255–2,511 times**. For each N17 registration, an observation is fresh−tired
+when the sampled role is its omitted clone and zero otherwise because the pools' best outcomes are
+identical. The pre-fixed family is seven, endpoint `δ=0.0035714285714285718`, with 400pp support
+width.
+
+| omitted N17 clone | N18−that N17 best (pp) | simultaneous 95% interval |
+| --- | ---: | ---: |
+| `clone-04` | +0.096454±0.021319 | [0.075136, 0.117773] |
+| `clone-05` | +0.080273±0.020059 | [0.060214, 0.100332] |
+| `clone-06` | +0.066324±0.018963 | [0.047362, 0.085287] |
+| `clone-07` | +0.035576±0.016225 | [0.019351, 0.051801] |
+| `clone-08` | +0.024480±0.014288 | [0.010192, 0.038768] |
+| `clone-09` | +0.020601±0.014259 | [0.006342, 0.034861] |
+| `clone-10` | **+0.013387±0.013130** | **[0.000257, 0.026517]** |
+
+All **7/7 lower endpoints are positive**. Selecting the lowest publication mean, `clone-10`, as the
+“best N17” is valid because all seven intervals are simultaneous. Two separate processes reproduce
+byte-identical complete output; SHA-256 of `POOL_RESULT_JSON` is
+`3013554c387daa709125b2077c17b22710761d70bf4a4c1d2c053fb21570b51b`.
+
+H therefore sets the global product maximum to **`N=17`**. Core eleven plus the largest free grant
+of six already needs 17, and there is an allowed case where even the best N17 registration in the
+same owned universe is inferior to N18. This is an existence witness for one pre-fixed CAP0 owned
+universe under shipped-default opponent roles and the uniform role/60-cell mean, not a global lower
+bound on N18−N17 over every pool. H's P13 question needs one allowed case to show slot 18 is not
+structurally inert. N=17 still does not erase purchase advantage: the one-free-player branch can
+leave as many as five optional selected slots. Runtime accepts only
+`FORM_PLAYER_POOL_MAX_SIZE=17`; N18 remains a complete-enumeration regression case.
+
+```bash
+MODE=pool-ceiling POOL_SEEDS=1000000 CONTROL_SEEDS=20 \
+  POOL_COLUMN=role-economics-r3-pool-final \
+  pnpm --filter pog-mcp exec tsx skill/reference/probes/role-economics.mts
+```
+
 ### 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`
@@ -3479,9 +3961,12 @@ in-division; the cooldown is initiator-only. Observed at-appearance `L` reaches
 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. The historical `g=0.4`
 single-player prize is 0.59pp [0.25, 0.93]; at the selected D product amplitude the maximum
-recoverable deficit gives 0.19pp [−0.07, 0.45], and the independently certified team-level worst
-upper bound is 3.69pp across the 18 observed cells. The unmeasured legal role×condition extreme
-is why that bound is not yet an activation verdict.
+recoverable deficit gives 0.26pp [0.04, 0.48] in the general-carrier diagnostic, and the
+independently certified team-level worst upper bound is 3.48pp across the 18 observed cells. The
+§3-15 follow-up establishes the #728 role ladder through ×1024 below 5pp, and result 7e measures
+the 0.173 recovery prize directly on all 66 focused roles. Its original two-vector H cost is
+withdrawn; result 7g provides the same-owned-universe H comparison. The feeder decision itself remains a
+product tradeoff; the now-closed H/I-c numeric gates do not decide it.
 
 ### What this does not settle
 
@@ -3515,11 +4000,12 @@ states while `allCells` did not. A deficit of exactly zero is not evidence of cl
 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 D product amplitude remains unresolved: at its maximum
-recoverable deficit 0.173, prize is 0.19pp [−0.07, 0.45], cheapest selected cost is 0.16pp, and
-paired cost−prize is −0.03pp [−0.50, 0.43]. The neutral-point fork is now closed at the D active-appearance median;
-what remains is 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 —
+The general-carrier prize-versus-cost ordering remains unresolved on its own: at maximum
+recoverable deficit 0.173 its prize is 0.26pp [0.04, 0.48], cost is 0.10pp, and paired
+cost−prize is −0.16pp [−0.53, 0.21]. Result 7e does not relabel that diagnostic; it supplies the
+missing direct 66-role recovery-prize and I-c measurement. Its two-vector H use is withdrawn;
+result 7g closes H across all admissible N17 registrations in one immutable owned universe. 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
@@ -3542,15 +4028,98 @@ itself. The cost is that nothing with a period longer than 11 days is visible.
 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
-# Post-F-1 product curve: predicate D, testing only the selected L₀=4×p95
+# F-2 v2 all-τ price table: predicate D, structural cup fold 0.1375,
+# testing only the selected L₀=4×p95.
 MATCHES=3000 SNAPSHOTS=/tmp/prod-tl.json REPLAYS=5 TAU_SWEEP=6,12,24,48 \
-  LOAD_PRED=D L0_SWEEP=4 LAMBDA_MED=6 LAMBDA_MAX=21 \
+  LOAD_PRED=D L0_SWEEP=4 CUP_FOLD_SCALE=0.1375 EXPLORE_WORST=0 \
+  CALIBRATION_SEED_COLUMN=c2-cupfold-r45-final LAMBDA_MED=6 LAMBDA_MAX=21 \
+  pnpm --filter pog-mcp exec tsx skill/reference/probes/engagement-load.mts
+
+# Independent confirmation of the pre-fixed exact product point.
+MATCHES=3000 SNAPSHOTS=/tmp/prod-tl.json REPLAYS=5 TAU_SWEEP=24 \
+  LOAD_PRED=D L0_SWEEP=4 CUP_FOLD_SCALE=0.1375 EXPLORE_WORST=0 \
+  FIXED_AMPLITUDE=0.4914 CALIBRATION_SEED_COLUMN=c2-cupfold-r42-product \
+  LAMBDA_MED=6 LAMBDA_MAX=21 \
+  pnpm --filter pog-mcp exec tsx skill/reference/probes/engagement-load.mts
+
+# Canonical hand-off for the cup gate: the first post-warm-up human entry vector
+# for each cup/team, with ladder and raw-cup components kept separate.
+SNAPSHOTS=/tmp/prod-tl.json TAU_SWEEP=24 TIME_BASE=claim LOAD_PRED=D \
+  CUP_FOLD_SCALE=0.1375 \
+  EXPORT_CUP_ENTRIES=/tmp/cup-entry-r36.json EXPORT_CUP_ENTRIES_ONLY=1 \
   pnpm --filter pog-mcp exec tsx skill/reference/probes/engagement-load.mts
 
 # Historical C-conditioned 1×/2×/4× table only
 MATCHES=3000 SNAPSHOTS=/tmp/prod-tl.json REPLAYS=5 TAU_SWEEP=6,12,24,48 \
   LOAD_PRED=C L0_SWEEP=1,2,4 LAMBDA_MED=6 LAMBDA_MAX=21 \
   pnpm --filter pog-mcp exec tsx skill/reference/probes/engagement-load.mts
+
+# 2026-09-06 final role×condition run. All 66 shards retain the full
+# 67,320-test × 3-look family and share one independent seed column.
+for focal_index in 0 1 2 3 4 5; do
+  ROLE_SIDE=offense FOCAL_INDEX="$focal_index" \
+    SEED_CHECKPOINTS=50000,100000,200000 SEED_COLUMN=role-cond-r37-final \
+    CONTROL_SEEDS=20 \
+    pnpm --filter pog-mcp exec tsx skill/reference/probes/role-cond.mts \
+    > "/tmp/role-r37-final-offense-${focal_index}.out" &
+done
+wait
+for focal_index in 0 1 2 3 4 5; do
+  for role_slot in 1 2 3 4 5 6 7 8 9 10; do
+    ROLE_SIDE=defense FOCAL_INDEX="$focal_index" ROLE_SLOT="$role_slot" \
+      SEED_CHECKPOINTS=50000,100000,200000 SEED_COLUMN=role-cond-r37-final \
+      CONTROL_SEEDS=20 \
+      pnpm --filter pog-mcp exec tsx skill/reference/probes/role-cond.mts \
+      > "/tmp/role-r37-final-defense-${focal_index}-${role_slot}.out" &
+  done
+  wait
+done
+
+# 2026-09-06 v2 cup boundary. 0.1375 is index 8; run it, every larger point,
+# and its lower neighbour without assuming monotonicity. Both pre-fixed
+# amplitudes are priced in the full 304,128-test family.
+for scale_index in 0 1 2 3 4 5 6 7 8 9; do
+  ENTRY_PROFILES=/tmp/cup-entry-r36.json CUP_SCALE_INDEX="$scale_index" \
+    TRIALS=3000 CONTROL_TRIALS=20 FORM_AMPLITUDE=0.4914 \
+    AMPLITUDE_FAMILY_SIZE=2 SEED_COLUMN=cup-role-r44-final \
+    pnpm --filter pog-mcp exec tsx skill/reference/probes/cup-role-load.mts \
+    > "/tmp/cup-r44-final-b04914-s${scale_index}.out" &
+done
+wait
+ENTRY_PROFILES=/tmp/cup-entry-r36.json CUP_SCALE_INDEX=8 TRIALS=3000 \
+  CONTROL_TRIALS=20 FORM_AMPLITUDE=0.5026 AMPLITUDE_FAMILY_SIZE=2 \
+  SEED_COLUMN=cup-role-r44-final \
+  pnpm --filter pog-mcp exec tsx skill/reference/probes/cup-role-load.mts \
+  > /tmp/cup-r44-final-b05026-s8.out
+
+# 2026-09-07 r1 role recovery / I-c run. Its two-vector H diagnostic is
+# audit-only. Each one-role shard still pays for the full 4,818-interval
+# family; grouping only limits concurrency.
+for focal_index in 0 1 2 3 4 5; do
+  FOCAL_INDEX="$focal_index" ROLE_SIDE=offense \
+    SELECT_SEEDS=5000 ECON_SEEDS=200000 CAP_SEEDS=50000 CONTROL_SEEDS=20 \
+    SELECT_COLUMN=role-economics-r1-select PUBLISH_COLUMN=role-economics-r1-final \
+    pnpm --filter pog-mcp exec tsx skill/reference/probes/role-economics.mts \
+    > "/tmp/role-economics-r1-${focal_index}-offense.out" &
+  for role_slot in 1 2 3 4 5 6 7 8 9 10; do
+    FOCAL_INDEX="$focal_index" ROLE_SIDE=defense ROLE_SLOT="$role_slot" \
+      SELECT_SEEDS=5000 ECON_SEEDS=200000 CAP_SEEDS=50000 CONTROL_SEEDS=20 \
+      SELECT_COLUMN=role-economics-r1-select PUBLISH_COLUMN=role-economics-r1-final \
+      pnpm --filter pog-mcp exec tsx skill/reference/probes/role-economics.mts \
+      > "/tmp/role-economics-r1-${focal_index}-defense-${role_slot}.out" &
+  done
+  wait
+done
+rg --no-filename '^RESULT_JSON ' /tmp/role-economics-r1-*.out \
+  | sed 's/^RESULT_JSON //' | jq -s 'sort_by(.roleIndex)' \
+  > /tmp/role-economics-r1-results.json
+shasum -a 256 /tmp/role-economics-r1-results.json
+
+# Corrected H: every N17 registration versus N18 in one owned universe.
+MODE=pool-ceiling POOL_SEEDS=1000000 CONTROL_SEEDS=20 \
+  POOL_COLUMN=role-economics-r3-pool-final \
+  pnpm --filter pog-mcp exec tsx skill/reference/probes/role-economics.mts
+# POOL_RESULT_SHA256 3013554c387daa709125b2077c17b22710761d70bf4a4c1d2c053fb21570b51b
 ```
 
 **The window's end is now stated explicitly, so these figures can be rebuilt.** The default is