pog-mcp 0.8.1 → 0.8.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pog-mcp",
-  "version": "0.8.1",
+  "version": "0.8.2",
   "type": "module",
   "description": "MCP server that lets an AI agent play Proof of Goal \u2014 wallet, sign-in, squad building, and matches as typed tools.",
   "license": "MIT",

package/skill/SKILL.md

@@ -332,25 +332,101 @@ player is the worst thing you can build — it lost to every differentiated squa
 tested, most by 3:1. Defenders want defense, forwards want shoot. This is the
 single largest effect measured, and it is what a naive even split gets wrong.
 
-**Points are not equally valuable everywhere.** An outfield player's total feeds
-their positional group sum, which multiplies into *every* contested action their
-team takes. The goalkeeper contributes to neither — it defends only 1v1 chances
-and penalty shootouts, using its own stats with no team contribution. So a point
-given to the keeper does less work than the same point given outfield. Budget
-accordingly rather than treating all eleven slots as equal.
+**The goalkeeper slot reads exactly one attribute when defending.** Every save
+a keeper ever makes scores `defense + cond/2 + total/3`, at three sites: the 1v1
+chance, an in-match penalty (a regulation-time event too — roughly an eighth of
+all goals are penalties), and a shootout kick. Its pass, dribble and shoot are
+read by no *defensive* code path.
+They are read in exactly two offensive cases: if you nominate the keeper as a
+kick taker (measured worst of every option — do not), and in a long shootout,
+where the kicking order covers all eleven players and can eventually reach the
+keeper, scoring `(pass + shoot) / 2 + cond / 2 + total / 3` — the same
+`takePkShot` roll as any taker. Neither changes the advice: whatever total
+your keeper ends up with, put it in defense.
+
+**How much total to give it is NOT settled, and you should distrust anyone who
+tells you it is** — including earlier versions of this file. Four separate
+measurements over one day gave four different answers, each overturned by
+controlling something the previous one had left free. What survived:
+
+- A smaller keeper total is better in league matches (draws allowed). This held
+  under every construction tried.
+- In knockout matches the best total **depends on shootout exposure — how
+  often YOUR matches end level, which both sides' attacks set together.** The
+  keeper's `total/3` is in every save it makes — regulation 1v1s included — but
+  a shootout multiplies how many keeper rolls a match contains. Measured across
+  three same-profile pairings (shootout reach ~13% to ~75%): at the
+  starved-attack extreme the ranking REVERSES and a total-29 keeper beats a
+  total-13 one outright; in ordinary pairings the small total stays best. Your
+  own attack is the half of that exposure you control.
+
+So treat the keeper total as a squad-level question, not a lookup.
+
+**There is a second budget, and this file's own advice competes for it.** Beyond
+the 212 points, the whole TEAM may hold at most three attributes of value 10 and
+five of value 8–9 (`validator.ts`). "Forwards want shoot", "give the kicks to
+your best shooter", and a maximal keeper defense cannot all be satisfied — three
+forwards at 10 plus a keeper at 10 is four, and the validator rejects it. Decide
+where the scarce values go on purpose.
+
+Measured head to head, 20,000 matches each, squads identical except for where a
+single 10 sits:
+
+| A single 10 given to… | League | Knockout |
+| --- | --- | --- |
+| the set-piece taker's `shoot` | **1st** | 2nd |
+| an ordinary forward's `shoot` | 2nd | 3rd |
+| the keeper's `defense` | 3rd | **1st** |
+
+The keeper and the ordinary forward swap ends. The set-piece taker is first or
+second in both — the only allocation measured that was never wrong. (This too was
+measured on one outfield template; read it as a strong hint, not a law.)
+
+**One thing about kickers that is easy to get wrong.** Your `isFkKicker` and
+`isPkKicker` choices decide who takes free kicks and in-match penalties —
+together roughly a third to a half of all goals, with the exact share depending
+on squad construction. **They do not apply to a shootout.** That order is computed
+from `shoot + pass` across the whole squad, so a shootout is decided by your best
+shooters whether you nominated them or not.
+
+Budget in ranges, not in single points: a one-point change is below what even
+thousands of matches can distinguish.
 
 **Spreading beats star-building.** Concentrating points into a few 26–29 players
 and starving the rest conceded roughly three times as many goals as an even
 outfield spread, and lost overall despite scoring more.
 
-**Solidity beats aggression.** Defense-leaning shapes outperformed attack-leaning
-ones. Scoring is low — well under one goal per team per match, and about half of
+**Solidity beats aggression — in matches that can end level.** With identical
+formations and budgets, the defense-leaning 4-3-3 beat the attack-leaning one
+56.5% of the time in league play, and in knockout rounds the same pairing
+flips: the attack-leaning side wins 53.6%. Both directions are significant at
+4,000 matches. A 0-0 is worth a point in the league — and in a cup GROUP
+match too, which is regulation-only and can end level just like a league
+round. Only in a cup KNOCKOUT round does a 0-0 proceed to a tiebreaker, and
+that tiebreaker is anything but a coin flip — the shootout is decided by your
+best shooters' `shoot + pass` against the keeper's `defense`/total (see the
+keeper section above), so the draw-heavy defensive shape hands the decision to
+exactly the attributes it skimped on. Scoring is low — well under one goal per team per match, and about half of
 all matches end level — so conceding one fewer is worth more than scoring one
 more.
 
-**Give the free kicks and penalties to your best shooter.** Kicks taken by a
-forward beat kicks taken by a defender by about 5 percentage points of win rate,
-and by a goalkeeper by 8. One player may take both.
+**The two set-piece roles split — measured separately, one flag moved at a
+time (`pk-taker.mts`).** A free kick goes direct half the time (scored by the
+kicker's `shoot`) and crossed half the time (gated by the kicker's `pass`, then
+a DIFFERENT receiver shoots). Even with that pass gate, the shoot-heavy taker
+wins the FK slot: a shoot-10/pass-4 striker out-converts a 9/9 passer-shooter
+34.4% → 47.9% and out-scores him by ~330 points over 4,000 matches. (Caveat:
+the crossed half's payoff runs through your receivers' `shoot` — this fixture's
+receivers are flat, which favours the direct half.) The in-match penalty flips:
+`takePkShot` rolls `(pass + shoot) / 2 + cond / 2 + total / 3 + rand`, and with
+totals and condition held equal the 9/9 converts 83.9% vs 69.8% for the 10/4.
+Both branches roll through `getPoint`, which adds `total / 3` (and
+condition) to the selected skill — the probe held totals and condition equal
+to isolate the attribute terms. So between candidates with DIFFERENT totals
+compare full expressions, not attributes alone: `isFkKicker` by `shoot +
+total / 3`, `isPkKicker` by `(pass + shoot) / 2 + total / 3`. One player may
+hold both roles if he wins both criteria. (Neither nomination applies to a shootout — that order is computed
+from `shoot + pass` across the whole squad.)
 
 **There is no home advantage.** Identical squads win equally often from either
 side. Never explain a result by which side you were on.
@@ -362,11 +438,21 @@ worthless as evidence, and this is the mistake to avoid:
 
 | Friendlies played | Chance the better squad looks better |
 | --- | --- |
-| 1 | 35% |
-| 5 | 62% |
-| 10 | 73% |
-| 30 | 91% |
-| 50 | 96% |
+| 1 | 34% |
+| 5 | 58% |
+| 10 | 70% |
+| 30 | 88% |
+| 50 | 94% |
+
+(Measured on ONE specific gap — a keeper-10 squad against a keeper-24 squad,
+re-validated with independent seeding. A smaller true gap needs MORE matches
+than this table says, a larger one fewer; read it as orders of magnitude, not
+as thresholds. **Measured with `allowDraw: true` (`gameFlg 0`) — a default
+friendly resolves level scores through ET and a shootout, so these draw-heavy
+rows do not describe it.** Worse than that: in default-friendly mode the SAME
+two squads' true ordering FLIPS — the keeper-10 squad loses 48.5/51.5 over
+20,000 matches, because keeper weight itself is mode-dependent (see the keeper
+section). Measure in the mode you intend to play.)
 
 At one match the most likely outcome is not a wrong answer — it is **no answer**:
 roughly half of all friendlies end level.
@@ -378,8 +464,8 @@ are a coin flip:
 | Friendlies | Identical squads showing a 2+ win gap | …that a significance test would call real |
 | --- | --- | --- |
 | 10 | 51% | 1% |
-| 30 | 70% | 3% |
-| 100 | 82% | 4% |
+| 30 | 72% | 3% |
+| 100 | 84% | 4% |
 
 Read the two columns together. "One squad is clearly ahead" is the normal state
 between squads that are exactly the same, and it gets MORE common with more

package/skill/reference/measurements.md

@@ -9,6 +9,13 @@ from an installed copy of this package.
 Re-run them after any engine change. **A rebalance invalidates this file, and
 `SKILL.md` with it.**
 
+**The probes are repo-only, deliberately.** They import the workspace engine by
+a relative path, so from an installed npm package they are module-not-found —
+that exact breakage shipped once and is now pinned by `packaging.test.ts`. To
+reproduce any table here, run from a checkout of
+`github.com/1XP-AI/solana-world-soccer-2026` under
+`packages/mcp/skill/reference/probes/`.
+
 ## Method
 
 Candidate squads are built to exactly 212 points and passed through
@@ -16,12 +23,42 @@ Candidate squads are built to exactly 212 points and passed through
 actually field. Squads then play each other over many seeds; only the seed
 varies between repetitions of a fixture.
 
-**All matches run with `gameFlg: 0` — regulation only, draws allowed.** The
-engine defaults it to 1, which sends a level score to extra time and can produce
-a golden-goal winner. An earlier version of every table here was measured that
-way, so it described a mixture of regulation and knockout outcomes while the
-Skill tells agents to compare squads with draws enabled. The rankings survived
-the correction; the magnitudes did not, and draws roughly doubled.
+**Tables run with `gameFlg: 0` — regulation only, draws allowed — unless a
+section says otherwise.** The engine defaults it to 1, which sends a level score
+to extra time and can produce a golden-goal winner. An earlier version of every
+table here was measured that way, so it described a mixture of regulation and
+knockout outcomes while the Skill tells agents to compare squads with draws
+enabled. The rankings survived the correction; the magnitudes did not, and draws
+roughly doubled.
+
+**That correction then created a blind spot of its own, found 2026-08-19.**
+Measuring only with draws allowed describes the ladder and cup group stage — but
+cup knockout rounds and default friendlies are played with `gameFlg: 1`, and at
+least one conclusion REVERSES between the two. A table measured in one mode is
+evidence about that mode only. Where a result depends on the mode, both are now
+published side by side.
+
+**Probe seeding matters more than it looks.** Until 2026-08-20 every probe here
+derived seeds with a hand-rolled FNV-style nibble walk, and the engine reads
+only the first 32 hex chars of a seed. An adversarial re-measurement showed that
+similar tag strings can produce CORRELATED seed families — phantom |z| up to ~7
+on cells whose true value is zero, and a true ~12 measured as ~1.7, depending on
+tag-string luck. `budget-allocation.mts`, `gk-split.mts`, `pk-taker.mts` and
+`cycle-probe.mts` now seed with SHA-256 and carry a standing mirror cell (same
+team vs itself; it must stay near z 0 or the seeding, not the game, is what got
+measured) — the cycle probe additionally applies a Bonferroni-corrected
+threshold, and its re-run is recorded in "Strategy round-robin in BOTH modes"
+below. `sample-size.mts` was re-validated the same day (SHA-256; both tables below
+updated — shape unchanged, rows a touch weaker; its identical-squads section is
+its own true-zero control). Still queued: `strategy-probe.mts`,
+`skill-value.mts`, `gk-sweep.mts`, and any published verdict resting on
+|z| < 10 from the old seeder.
+
+**A probe that scores by comparing `homeScore` to `awayScore` is wrong under
+`gameFlg: 1`** and silently reports decided matches as draws. `MatchResult` has
+no `penaltyWinner` field — read `winner`, which carries the engine's verdict
+after every tiebreaker. This mistake was made and caught here; the note is left
+so it is not made again.
 
 One earlier version of this harness repaired the point total globally after
 shaping each player, which quietly flattened all eleven candidates into the same
@@ -71,8 +108,16 @@ the ten outfield players. Every squad plays the same fixed opponent (a keeper of
 | 23 | 1.144 | 0.47 |
 | 29 | 0.991 | 0.51 |
 
-Monotonic across all twenty values. Note the second column: a *better* keeper
-concedes *more*, because the points came out of the ten players in front of it.
+The trend falls across the whole range, but **adjacent totals do not separate**.
+Re-running the sweep, the order reverses at five of the twenty values — 13, 15,
+17, 19 and 28 each score above the total directly below them. Three thousand
+matches cannot split a one-point difference in keeper budget. What it does split
+is the range: the ~0.4 pt/match gap between the low teens and the high twenties
+is far larger than any reversal. Read this table as "the low end beats the high
+end", never as "19 beats 20".
+
+Note the second column: a *better* keeper concedes *more*, because the points
+came out of the ten players in front of it.
 
 **Mechanism** (`packages/engine/src/engine.ts`): the goalkeeper is selected as
 defender only in scene `A0`, the 1v1 chance — its weight is 0 in every other
@@ -83,8 +128,178 @@ through the rate tables. The keeper's stats also decide penalty shootouts
 (`takePkShot`).
 
 This is faithful to the original sgsoccer engine, not a defect introduced here.
-It is nonetheless a large enough effect to be worth knowing about — see the note
-at the bottom.
+
+### …and the "reversal" it seemed to show was an artefact
+
+An earlier version of this section reported that the keeper budget REVERSES in
+knockout matches — cheap best in regulation, expensive best in knockouts. That
+was wrong, and the way it was wrong is worth keeping.
+
+**The sweep above is not a single-variable sweep.** `gk-sweep.mts` builds the
+keeper from a fixed ratio, `a(1, 1, 1, 3)`, so the total and the keeper's
+`defense` move together and cannot be told apart. Worse, `defense` hits its cap
+of 10 at around total 20, so below that a "smaller budget" also means "a worse
+keeper" while above it the total grows with defense already pinned. The knockout
+column's apparent peak in the high teens is those two effects crossing, not a
+budget optimum.
+
+Separating them — total held while defense varies, then defense held while the
+total varies; 4,000 matches per cell; opponent fixed at total 20 / defense 10:
+
+**Defense pinned at 10, total varied**
+
+| GK total | Regulation | Knockout |
+| --- | --- | --- |
+| 13 | **1.492** | **1.613** |
+| 16 | 1.375 | 1.615 |
+| 20 | 1.244 | 1.505 |
+| 24 | 1.075 | 1.380 |
+| 29 | 0.945 | 1.253 |
+
+**Total pinned at 20, defense varied** (outfield byte-identical across rows)
+
+| GK defense | Regulation | Knockout |
+| --- | --- | --- |
+| 2 | 1.271 | 1.090 |
+| 4 | 1.240 | 1.158 |
+| 6 | 1.250 | 1.224 |
+| 8 | 1.301 | 1.364 |
+| 10 | 1.280 | **1.492** |
+
+**Neither axis reverses ON THIS OUTFIELD** — and that qualifier turned out to
+matter. A smaller total is better in both modes here; more defense is better in
+both modes, weakly in regulation and strongly in knockouts.
+
+> ⚠️ **The knockout half of that is outfield-conditional.** Reproducible from
+> this repo — `gk-split.mts` section ④ holds the keeper at defense 10, swaps
+> only the outfield across three legal fixtures (opponent = same-outfield
+> total 20 / defense 10), and prints the shootout-reach rate that drives it:
+>
+> | Outfield (shoot profile) | Shootout reach | Knockout: total 13 vs 29 |
+> | --- | --- | --- |
+> | weak (shoot starved to 1) | ~75% | **REVERSED — 29 wins by ~0.85 pt/m** (1.202 vs 2.050) |
+> | low (even spread, no 10s) | ~23% | monotone, 13 best (1.662 vs 1.298) |
+> | high (two forwards at shoot 10) | ~13% | monotone, 13 best (1.661 vs 1.352) |
+>
+> `total/3` is in every save (regulation 1v1 included); a shootout multiplies
+> those rolls, so the keeper's total is worth more the more often the match ends
+> level. **The best keeper total is a function of SHOOTOUT EXPOSURE — a property
+> of the pairing (both attacks together), not of the keeper and not of one squad
+> alone.** The reach column is the mediator itself. Construction matters and bit
+> us once: against a FIXED opponent the reversal disappears (exposure stops
+> varying between arms) — the fixture pairs same-profile outfields precisely so
+> exposure varies, and that also means the table demonstrates a pairing effect;
+> your own attack is the half of it you control. The flip was first seen in an
+> adversarial scratch run (z≈24); section ④ replaces that citation with
+> something a checkout can re-run. The regulation column stayed monotone under
+> every construction tried.
+
+**Mechanism**: every save a keeper makes scores the same way —
+`defense + cond/2 + total/3` against the attacker — at three sites: the 1v1
+(`engine.ts:1385` via `getPoint`), the in-match penalty, and the shootout kick
+(both through `takePkShot`, `engine.ts:1251`; `handlePk` calls it in regulation
+too, so keeper defense pays on ~an eighth of goals even in league play). Nothing
+else on the keeper is read by any defensive path, so its pass, dribble and shoot
+are dead points there. The penalty/shootout roll weighs harder because its
+random range is `rand(5)` against the taker's `rand(16)`.
+
+**A related trap in the set-piece tables**: `isFkKicker` and `isPkKicker` decide
+free kicks (`engine.ts:785`) and in-match penalties (`engine.ts:1214`), but
+`runPkShootout` ignores both — it sorts the whole squad by `shoot + pass`
+(`engine.ts:1062`). Any set-piece conclusion measured with `gameFlg: 0` therefore
+says nothing about shootouts, and the shootout cannot be steered by nomination at
+all.
+
+**That is still not the whole answer**, and the first version of this section
+claimed it was ("defense 10, everything else at 1, total 13 — best in both
+modes"). That claim came from a sweep in which the points freed from the keeper
+were spread evenly over the outfield — nobody lost anything, so no trade-off
+could appear. A sweep has no opportunity cost built in.
+
+Give those points a REAL competing use and the ranking changes. The squad has a
+second budget beyond 212: **at most three attributes of value 10 and five of 8–9
+across the whole team** (`validator.ts` `TEN_MAX` / `EIGHT_MAX` — team-wide, not
+per player). Where one of those three 10s goes, measured head to head at 20,000
+matches per pairing on otherwise byte-identical squads:
+
+| A single 10 given to… | League | Knockout |
+| --- | --- | --- |
+| the set-piece taker's `shoot` | **1st** | 2nd |
+| an ordinary forward's `shoot` | 2nd | 3rd |
+| the keeper's `defense` | **3rd** | **1st** |
+
+| Pairing | League | Knockout |
+| --- | --- | --- |
+| keeper-10 vs taker-10 | 45.1% (z −18.5) | 52.8% (z +7.9) |
+| taker-10 vs forward-10 | 51.9% (z +6.9) | 53.7% (z +10.4) |
+| forward-10 vs keeper-10 | 50.9% (z +3.5) | 43.3% (z −19.1) |
+
+*Re-measured 2026-08-20 with SHA-256 seeding: ranking and signs unchanged
+(z −15.5 / +11.0 / +5.5 regulation, +5.4 / +12.2 / −19.1 knockout; mirror
+control z −0.09).*
+
+The keeper and the ordinary forward swap ends completely. The set-piece taker is
+first or second in both — the only allocation measured that is never wrong.
+
+**Two lessons for this file, learned the hard way on the same day:**
+
+1. A sweep that varies one *named* quantity is not a single-variable sweep if the
+   construction ties other quantities to it (the keeper's ratio tied `defense` to
+   the total). State the construction, or measure the axes apart.
+2. **Separating the axes is not enough.** A sweep measures a quantity against
+   nothing, so it cannot rank allocations. Any claim of the form "X is optimal"
+   needs a control that spends the same budget somewhere else.
+
+**Which matches are which** (`packages/api/src`) — still worth knowing, because
+the SIZE of the defense effect differs by an order of magnitude between them:
+
+| Match type | `gameFlg` |
+| --- | --- |
+| Ladder (`routes/playoff.ts`) | 0 |
+| Cup group stage (`scheduler/daemon.ts`) | 0 |
+| Cup knockout (`scheduler/daemon.ts`) | 1 |
+| Friendly, default (`routes/matches.ts`) | 1 |
+
+## Strategy round-robin in BOTH modes (`cycle-probe.mts`)
+
+Recorded 2026-08-20 — the founding "no cycles" claim had circulated without a
+home in this file. Construction: the SAME squads as the round-robin above (the
+builders are shared code in `squad-lib.mts`, not a copy — an earlier hand-copied
+variant repaired the scarcity caps in a different order and silently built
+different squads from identical slot totals), SHA-256 seeding, mirror control
+z 0.89, eight candidates, 4,000 matches per pairing per mode.
+
+Significance is Bonferroni-corrected across all 56 pairings (two modes ×
+28): |z| >= 3.32 rather than a naive 1.96, which over 56 comparisons all but
+guarantees a few false edges — and one false edge can create or erase a
+dominance or cycle verdict.
+
+| | Regulation (draws) | Knockout (ET + shootout) |
+| --- | --- | --- |
+| Significant pairings | 25/28 | 24/28 |
+| 3-cycles (A>B>C>A) | **0** | **0** |
+| Dominant strategy | none | **`role-433`** beats all seven |
+
+Mode reversals that survived this re-measurement:
+
+- **Attack vs defense lean** (same formation, same budgets): `def-433` beats
+  `role-433` 56.5% in regulation; the SAME pairing flips to 53.6% for the
+  attack lean in knockouts. Both significant — the cleanest reversal measured.
+- **Formation**: `def-532` beats `bal-442` 73.9% in regulation and loses 56.6%
+  in knockouts.
+- **Cheap keeper**: `gkmin-433` collapses in knockouts (27~42% against the
+  other top squads). Its regulation edge over the top group did NOT survive the
+  corrected builders (50.8% vs `def-433`, 47.4% vs `def-532` — not significant);
+  what remains significant in regulation is its edge over `role-433` (57.0%)
+  and everything below the top group.
+
+Monotone in both modes, still: role differentiation (`flat-433` last
+everywhere) and spreading over star-concentration (`stars-433` next-to-last).
+
+Scope: all squads tenure=0, as everywhere in this file. Production applies
+`applyMatchGrowth()` before the engine, raising effective totals — these
+dominance/reversal claims are about zero-buff squads until the buff-0/max
+control runs (docs/tactics/01, track 0.1).
 
 ## Home advantage: none
 
@@ -103,34 +318,89 @@ Same squad, only the kicker slots move. Opponent fixed, 3,000 matches each.
 | DF / DF | 20.4 | 1.149 |
 | GK / GK | 17.4 | 1.066 |
 
-Conversion tracks the taker's shooting, so the kicks belong with the best
-shooter. Giving both roles to one player costs nothing.
+Conversion tracks the taker's shooting — for FREE kicks. This table moved both
+roles together and never crossed the two criteria, so it says nothing about the
+PK slot on its own. `pk-taker.mts` crosses them: one squad held byte-identical,
+only `pkKicker` moved between a 9-pass/9-shoot midfielder (the PK formula's
+pick — `takePkShot` scores the taker as `(pass + shoot) / 2 + …`) and a
+10-shoot/4-pass striker (the FK criterion's pick). Opponent fixed, regulation,
+4,000 matches per arm, mirror z −0.62:
+
+| pkKicker | (pass+shoot)/2 | PK attempts | Conversion |
+| --- | --- | --- | --- |
+| mid 9/9 | 9 | 299 | **83.9%** |
+| fw 10/4 | 7 | 301 | 69.8% |
+
++14.2pp, two-proportion z 4.12. In-match penalties are rare (~0.07 attempts
+per match here), so the points axis is underpowered — the verdict is the
+conversion rate.
+
+The FK half is measured the same way (section ② of the probe): only
+`fkKicker` moves, `pkKicker` pinned. A direct FK (half of `doFk`) scores by
+the kicker's `shoot`; a crossed FK gates on the kicker's `pass` and then a
+different receiver shoots. Attempt counts include foul-site direct FKs taken
+by the fouled ball-carrier — those ignore the flag and are identically
+distributed across both arms, so the BETWEEN-ARM gap is attributable to the
+designated-kicker FKs alone (and the per-attempt gap is an underestimate).
+
+| fkKicker | shoot | pass | FK attempts | Team FK goals | Per attempt | W-D-L |
+| --- | --- | --- | --- | --- | --- | --- |
+| mid 9/9 | 9 | 9 | 2,307 | 793 | 34.4% | 1289-1532-1179 |
+| fw 10/4 | 10 | 4 | 2,311 | 1,107 | **47.9%** | 1420-1470-1110 |
+
+−13.5pp, two-proportion z −9.34; points 5,399 vs 5,730. Even with the pass
+gate, the shoot-heavy taker wins the FK slot — in THIS fixture, whose
+receivers are flat: the crossed half's payoff runs through the receivers'
+`shoot`, so a squad with strong secondary attackers shifts value back toward
+the crossing taker. Both set pieces roll through `getPoint`/`takePkShot`,
+which add `total/3` (and condition) to the attribute term — the probe held
+totals and condition equal to isolate that term. So: free kicks to the best
+`shoot + total/3`, the penalty to the best `(pass+shoot)/2 + total/3`; with
+unequal condition include `cond/2` too; one player may hold both roles when
+he wins both criteria.
 
 ## How many friendlies a conclusion needs
 
 Two squads with a known true gap, resampled 20,000 times at each sample size.
 
+Re-measured 2026-08-20 with SHA-256 seeding (the original numbers came from the
+correlated FNV seeder; shape unchanged, every row a touch weaker). Scope: this
+is ONE matchup — a keeper-10 squad against a keeper-24 squad, ratio-locked
+construction — so the true gap is that specific ~1.8-win-in-ten one. A smaller
+true gap needs more matches than the row says; a larger one, fewer.
+
+Scope, part two (Codex P2, #688): **everything below is `gameFlg 0` —
+regulation, draws allowed** (`allowDraw: true` on a friendly), which is what
+`sample-size.mts` simulates. A DEFAULT friendly is `gameFlg 1`. Running the
+probe with `GAME_FLG=1` does not merely shrink the tie column: the same two
+squads' true ordering FLIPS — the keeper-10 squad, the better side in
+regulation, loses 48.5/51.5 over 20,000 default-mode matches, consistent with
+the keeper section's mode dependence (ET and a shootout multiply keeper
+rolls). A "which squad is better" verdict is mode-specific twice over: the
+sample sizes below apply only to draw-allowed friendlies, and the ANSWER can
+change when the mode does. Measure in the mode you intend to play.
+
 | Friendlies | Correct conclusion | Wrong conclusion | Tied |
 | --- | --- | --- | --- |
-| 1 | 35.3% | 16.7% | 48.0% |
-| 3 | 53.2% | 19.6% | 27.2% |
-| 5 | 61.6% | 18.4% | 20.0% |
-| 10 | 73.0% | 14.5% | 12.5% |
-| 20 | 85.0% | 8.7% | 6.4% |
-| 30 | 90.7% | 5.8% | 3.5% |
-| 50 | 96.4% | 2.3% | 1.3% |
-| 100 | 99.4% | 0.4% | 0.1% |
+| 1 | 33.6% | 17.2% | 49.2% |
+| 3 | 50.2% | 20.4% | 29.3% |
+| 5 | 58.3% | 19.7% | 22.0% |
+| 10 | 69.7% | 16.6% | 13.7% |
+| 20 | 81.5% | 11.1% | 7.4% |
+| 30 | 88.0% | 7.5% | 4.5% |
+| 50 | 94.1% | 3.9% | 1.9% |
+| 100 | 98.9% | 0.7% | 0.4% |
 
 And the false-positive side — **identical** squads, by three different
 definitions of "one side looks better":
 
 | Friendlies | Record not exactly level | Win-loss gap ≥ 2 | Significant (sign test, p<0.05) |
 | --- | --- | --- | --- |
-| 1 | 50.7% | 0.0% | 0.0% |
-| 5 | 75.4% | 34.9% | 0.0% |
-| 10 | 82.7% | 50.8% | 0.8% |
-| 30 | 89.8% | 70.1% | 3.0% |
-| 100 | 93.8% | 82.1% | 4.1% |
+| 1 | 51.4% | 0.0% | 0.0% |
+| 5 | 75.8% | 35.0% | 0.0% |
+| 10 | 82.8% | 50.9% | 1.0% |
+| 30 | 90.6% | 72.3% | 2.9% |
+| 100 | 94.6% | 83.9% | 3.5% |
 
 **The first column is a trap and was published on its own in an earlier version
 of this file.** "Not exactly level" approaches 100% for any pair of squads simply
@@ -174,19 +444,90 @@ rest drawn.
 
 ## Note for the maintainers
 
-Spending the minimum on the goalkeeper is monotonically better across the whole
-range (1.406 vs 0.991 points per match). Anyone who runs a few hundred
-simulations will find it, and agents playing through the MCP server will find it
-faster than people will.
-
-It is worth keeping in proportion, though: against a naive squad, role shaping
-alone wins 43.6% and adding the cheap keeper takes that to 52.1%. Head to head,
-the cheap-keeper version beats the conventional-keeper version 25.7% to 19.3%
-with 55% drawn. So it is an edge, not the game. The thing that actually
-decides matches is whether players have distinct roles at all.
-
-`SKILL.md` therefore states the *principle* — points spent on the keeper do less
-work, because of how the engine aggregates — rather than a "set your keeper to
-10" instruction. That is the honest version and leaves the choice with the
-player. If the keeper is ever given team contribution or a budget floor, re-run
-these probes and rewrite both files.
+> **2026-08-19: this note was rewritten twice in one day.** First it said the
+> cheap keeper was "monotonically better across the whole range". Then it said
+> the direction reverses in knockouts. Both were readings of a sweep that could
+> not separate the keeper's total from its defense. The section above has the
+> separated measurement; this note now describes that.
+
+The keeper slot reads exactly one attribute — `defense` — at exactly three
+sites: the 1v1, the in-match penalty (a regulation event too, ~an eighth of all
+goals), and the shootout kick. Everything else parked there is dead weight. So there
+are two independent facts, and conflating them produced both earlier errors:
+
+- **Total: smaller is better in matches that allow draws** — that held under
+  every construction tried. In knockouts it is OUTFIELD-CONDITIONAL (see the
+  warning above): a low-scoring squad reaches shootouts often enough that a
+  bigger keeper total pays there.
+- **Defense: higher is better on its own slot, in both modes** — weakly in
+  regulation, strongly in knockouts. But a 10 there is one of only three the
+  whole team may hold, and in the scarce-10 comparison above the keeper's 10
+  ranked LAST in league play.
+
+So even the two facts do not compose into one recommended shape. What survives
+is the negative advice: never leave points in the keeper's pass/dribble/shoot,
+and never scale a balanced keeper down.
+
+Keep it in proportion. Against a naive squad, role shaping alone wins 43.6% and
+adding a cheap keeper takes that to 52.1%. Role differentiation still decides
+more matches than the keeper does.
+
+`SKILL.md` states the two facts separately and warns against scaling a balanced
+keeper down, because that drags defense along with the total. If the keeper is
+ever given team contribution or a budget floor, re-run these probes in BOTH
+modes AND with the two axes separated — the ratio-locked sweep alone will not
+show it.
+
+### The decision, and what would reopen it
+
+Issue #604 settled on **leaving the engine faithful to sgsoccer** — and made
+that settlement *conditional* on the meta not collapsing onto one answer. The
+separated measurement makes that risk concrete again: defense 10 is the single
+best keeper defense everywhere measured, and under ordinary shootout exposure —
+regulation, and knockout pairings that reach shootouts at typical rates — the
+best shape is defense 10, total 13. The one measured exception is the pairing
+block above: against shoot-starved pairings reaching shootouts ~75% of the
+time, the optimum climbs the total axis (total 29 beat total 13 by ~0.85 ppm)
+while defense stays at 10. So what a converging population would collapse onto
+is per-context on the total axis but single-valued on the defense axis. The
+earlier hope that "the meta has nowhere to collapse to" rested on a reversal
+that turned out to be an artefact.
+
+What is worth watching therefore returns to the original question, with one
+correction: **watch the keeper's `defense`, not its total.** A population
+converging on total 13 while keeping defense at 10 has found the optimum; a
+population converging on a low total with a scaled-down defense has been misled
+and is playing worse. Those two look identical to a total-only metric.
+
+Read-only from production on 2026-08-18 (32 human squads, seeded AI bots
+excluded): median keeper total **19**, mean 17.6, **nobody** above 24, and only
+two squads (6%) at the minimum. So the direction the probes predict is visible —
+heavy keepers are unbought — but the meta has NOT converged. Two thirds sit at
+19–20, which is the default left untouched rather than a choice.
+
+That reading is limited by sample, not by engine design: in that season only two
+human teams had played more than three matches, so no comparison *by result* was
+possible. The re-review triggers therefore key on sample first:
+
+| # | Trigger | Threshold |
+| --- | --- | --- |
+| 1 | Human teams with enough matches to compare by result — AND a contrast: two keeper-total cohorts of >= 3 qualified teams each. UNDECIDABLE (reported, never firing) until matches record the squad they were played with: cohorts group season records by the CURRENT keeper, and FA-market trading makes mid-season keeper edits routine, so a cohort can be fabricated | >= 10 teams, >= 2 cohorts, per-match squad snapshots |
+| 2 | Share of human squads on the MONITORED allocation — total <= 13 (the legal minimum with defense 10 is 1+1+1+10) AND defense >= 9. Cheap-but-weak squads are the opposite event (stale-advice damage) and are reported separately as `misled`, never firing this trigger. Undecidable below 10 human squads — a share over fewer reads one manager's choice as a meta | > 40% of >= 10 squads |
+| 3 | Median keeper total of the top 10 ranked human teams, each with >= 10 matches (a one-match post-rollover ranking stays null). UNDECIDABLE (reported, never firing) until per-match squad snapshots exist — rank is season-long, the keeper total is current, and leaders swapping keepers after earning their records would fire it over zero matches at the new total | <= 14, plus squad snapshots |
+
+These live as code, not as prose here — prose does not get re-read:
+
+```bash
+DATABASE_URL=postgres://… pnpm --filter @sws26/api exec tsx scripts/gk-budget-watch.ts
+```
+
+It prints the current distribution, the record grouped by keeper total, and
+whether any trigger has fired. `packages/api/src/lib/gk-budget-watch.ts` holds
+the thresholds; the CLI reads squads through `TeamStore` rather than the
+`players` table, because a sold keeper leaves `roster_slots` without rewriting
+the legacy row.
+
+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.