Synthesis from Layers 1–3 below. Supported findings only — inconclusive signals deliberately omitted.
Layer 1 — Chip deployment
Herd TC timing was anti-information. GW26 drew 29 of 84 TCs but had 8.41 pts mean regret vs 4.04 pts off-modal (only 41.7% of TCs landed on the optimal GW).
Herd BB timing also anti-information. GW33 drew 34 BBs with 17.56 pts mean regret vs 5.59 pts off-modal — ~3× worse. Modal-GW BB on the obvious DGW underdelivered relative to quieter weeks.
Free Hit usage is broadly value-additive.80.0% of FH plays delivered positive delta vs the pre-FH squad (mean +15.0 pts). No evidence of systematic mistiming.
Recommendation: Do not auto-follow top-50 modal GW for TC or BB — the herd signal was anti-information this season. Continue eager Free Hit deployment; don't hold for a "perfect" DGW.
Layer 2 — Triage calibration
compute_verdict discriminates strongly. HOLD (n=483) averaged 4.22 pts/GW vs SELL (n=21,014) at 0.81 pts — a 3.41-pt gap. The model works.
Suggested threshold shift is overfit — keep current values. Grid-sweep winner (HOLD≥0.7) lifts the gap by ~0.4 pts but shrinks the HOLD bucket from 483 to ~84 rows (0.3% of population). Marginal lift is inside sampling noise. Do not change verdict_weights.json.
Recommendation:Keep verdict_weights.json at HOLD≥0.6 / MONITOR≥0.45. Re-evaluate after a full season of coverage; current grid-sweep lift is overfit to a tiny tail.
Layer 3 — Replacement quality
Premium IN-players dominate. ≥£10m tier wins 79.1% (mean +13.1 pts/3GW) vs <£6m at 54.6% (+3.5 pts) — 24.5pp gap. Bias trade targets toward premium when funds permit.
MID > FWD as trade targets. MID (n=1,379) wins 66.3% (+6.9 pts) vs FWD at 56.2% (+5.5 pts) — 10pp gap on substantive samples.
62.2% overall win rate on n=3,050 trades. Trades are net positive but ~38% are wash-or-loss; verdict-driven transfers are not free yield.
Recommendation: When triage flags multiple SELLs, prioritise replacements that are premium MIDs. Treat budget enabler trades as marginal — only when needed for chip funding or short-term punt, not as default value plays.
Caveats on these findings
Hindsight: regret figures assume perfect information. The "anti-modal" finding does not mean the herd was irrational — it means the obvious play didn't pay this season.
Audit subjects are top-50 managers, not our 3 teams. Layer 1/3 reflects league-leader behaviour; per-team rows for SlowBuild / Yamal1 / MbappeSalah are pending data capture (TODO.md).
n=1 season. Findings are calibration evidence, not laws. Re-run end-of-season and after 2026/27 H1 before any structural model change.
Layer 1 — Chip Deployment Audit
Top-50 managers · GW1–33 · Hindsight regret: best alternative GW within same chip half
Chip
Events
Modal GW
Median Regret
P90 Regret
Hit Optimal GW
Triple Captain
84
GW26
6.0 pts
14 pts
41.7%
Bench Boost
92
GW33
9.0 pts
21 pts
25.0%
Free Hit
60
GW13
delta vs pre-FH squad: +15.0 pts avg | 80.0% positive
Wildcard
89
GW32
descriptive only — yield analysis out of scope
Approximations: regret uses squad-actual-per-GW (manager's real captain/bench each week).
Hindsight regret assumes perfect information — see report footer for caveats.
Chip Timing — Yield & Regret by Chip Half
Each team receives a full chip set at GW1 and a second set at GW20, so TC/BB/FH decisions live in two independent windows. These Paretos split by chip half: H1 (GW1–19) covers TC1/BB1/FH1; H2 (GW20+) covers TC2/BB2/FH2. n≥5 per GW; tail grouped into Other.
First half — GW1–19 (TC1, BB1, FH1)
TC ceiling — H1
Managers whose TC1 ceiling landed here (n=50)
BB ceiling — H1
Managers whose BB1 ceiling landed here (n=50)
FH net Δ — H1
Sum of fh_delta by deployed GW (n=48); green = paid off, red = trap
Second half — GW20–33 (TC2, BB2, FH2)
TC ceiling — H2
Managers whose TC2 ceiling landed here (n=34)
BB ceiling — H2
Managers whose BB2 ceiling landed here (n=42)
FH net Δ — H2
Sum of fh_delta by deployed GW (n=12) small sample; green = paid off, red = trap
Regret concentration by half
TC + BB regret — H1
Stacked regret pts at each deployed GW · 100 events · 438 pts total
TC + BB regret — H2
Stacked regret pts at each deployed GW · 76 events · 949 pts total
TC regretBB regretOther (n<5 per GW)Cumulative %80% guideline
How to read these charts
Yield Paretos (rows 1 & 2)
Top bars = highest-yield GWs in that half. Each bar shows how many of the top-50 managers had their chip ceiling at that GW. The cumulative line crossing 80% marks the "vital few" deployment windows.
Free Hit bars are signed. Green = FH net-paid-off across the sample; red = systematic trap. GW33 (H2) is the clearest negative-EV FH week.
Concentration shape tells you timing importance. A sharp Pareto (H1 TC) says a single GW dominated; a flat one (H2 BB) says timing mattered less than squad quality in that window.
Regret Paretos (row 3)
Tall bars = where the herd deployed and paid for it. Top bars collect the most forgone yield within the half.
A small "Other" bar means regret is concentrated, not diffuse — reinforces the "target specific GWs" lesson for that half.
H1 vs H2 — cross-half reading
H1 TC is dominated by a single GW (GW6 accounts for the majority of H1 TC ceilings); H2 TC splits across GW31 and GW28. Interpretation: "when to play TC?" has a clearer answer in H1.
H2 FH is a sparse sample — only GW33 meets the n≥5 threshold. Read it as "the actionable finding is DGW33," not as a general H2 FH distribution.
H2 regret concentrates in 2 GWs (≈93% cumulative); H1 regret is diffuse across ~8 GWs. Different mistiming patterns apply to each half.
Combined interpretation — cross-reference yield and regret within each half:
Yield top bar?
Regret top bar?
Interpretation
Yes
Yes
Ceiling was real, herd missed it → strongest signal to target next year
No
Yes
Mediocre GW, slight ceiling miss → low-stakes mistake
Yes
No
Herd matched the ceiling (rare) → herd behaviour worked
No
No
Uneventful GW — no signal
Layer 2 — Triage Model Calibration
Reconstructed verdicts via compute_verdict · avg_fdr fixed at 3.0 (neutral) · chance_next/tsb_pct from current snapshot