# Gamma-wall history backfill — results · 2026-09-09 13:04

Pre-registration: `2026-09-09-gamma-backfill-prereg.md` (frozen before the pull). History days 730 (2023-10-10 → 2026-09-08); sessions evaluated 693 (quad-witching excluded). Regimes: neutral 281, short 268, long 144. NDX regime agreement with QQQ: 36%.

## Proxy validation (S−1 close book vs live 09:10 book)
n=22 overlap sessions: call wall within 0.5% **95%**, put wall **100%**, regime agreement **100%** → proxy **FIT**.

## Part C — wall policies
- **P10 cage held**: 68% of 693 sessions; by regime long 77%, neutral 72%, short 59%.
- **P8_longg_fade**: eligible 144, triggers 20, W/L/T 0/7/13, R/trigger **-0.285** (p=0.978), halves -0.667 (n=3) / -0.218 (n=17), $ at 1 NQ -8,360 → **REFUTED**.
- **P9_shortg_break**: eligible 268, triggers 136, W/L/T 13/50/73, R/trigger **-0.134** (p=0.964), halves -0.156 (n=61) / -0.116 (n=75), $ at 1 NQ -25,255 → **REFUTED**.

## Part D — first-touch wall fade (limit at the wall, stop max(12, 0.5×exp 5m range), 1R)
Walls on the correct side of the open: 1386; touched 449 (32%).
- Hit rate at 1R **42.1%** (W/L/close 189/260/0), mean R -0.158, $ at 1 NQ -21,902; halves 42.7% (n=211) / 41.6% (n=238).
- By side: call 42.5% (n=228, R -0.149), put 41.6% (n=221, R -0.167).
- By regime: long 45.3% (n=64, R -0.094), neutral 38.8% (n=178, R -0.225), short 44.0% (n=207, R -0.121).
- By touch time: after 10:00 44.0% (n=316, R -0.120), 09:30–10:00 37.6% (n=133, R -0.248).
- Per-year hit rate: 2023 40% (n=35), 2024 39% (n=142), 2025 47% (n=156), 2026 39% (n=116).
→ **REFUTED** (rule: ≥57% in both halves, long-γ ≥ short-γ).

## Part D2 — same fade, doctrine stop max(12, 2×ATR3 5m), 1R (single pre-declared variant)
Walls on the correct side of the open: 1386; touched 449 (32%).
- Hit rate at 1R **41.0%** (W/L/close 184/205/60), mean R -0.051, $ at 1 NQ -49,484; halves 47.4% (n=211) / 35.3% (n=238).
- By side: call 42.1% (n=228, R -0.047), put 39.8% (n=221, R -0.056).
- By regime: long 35.9% (n=64, R -0.183), neutral 43.8% (n=178, R +0.058), short 40.1% (n=207, R -0.104).
- By touch time: after 10:00 42.1% (n=316, R -0.006), 09:30–10:00 38.3% (n=133, R -0.159).
- Per-year hit rate: 2023 43% (n=35), 2024 49% (n=142), 2025 37% (n=156), 2026 37% (n=116).
→ **REFUTED** (rule: ≥57% in both halves, long-γ ≥ short-γ).
## Robustness and closing notes
- **Book identity.** On all 24 overlap sessions the live 09:10 book equals the prior-close book (net GEX by strike r = 1.00). The hourly forward archive adds intraday wall migration only; the pre-open wall map is fully reconstructable from the dated endpoint. History is stored in `fp-data/gex/history/YYYY-MM-DD.json.gz` (QQQ + NDX strike GEX, QQQ regular close; 730 days, 31 holidays absent).
- **Look-ahead correction.** Before the walk correction Part D printed 62–64 % at 1R with mean R +0.24 — entirely an artefact of crediting the touch bar's opposite extreme. Corrected: 42 %. The live grader (`lib/scoreBrief.ts` walkPlan) had the same flaw for LIMIT fills; fixed and deployed 2026-09-09, all sessions since 2026-07-29 regraded. <redacted: live shadow-ledger tally after regrade; the desk's live record, withheld under the 2026-09-10 disclosure rule>.
- **Costs** were not needed for the verdict (gross expectancy already negative). Under Part D the median stop is 13 pt; 17 % of trades still stop on the touch minute and the median hold is 1 minute — the wall is not a level NQ respects at 1-minute granularity with a 13-pt stop, and widening the stop to the doctrine stop (D2) did not recover it (41 %, both halves < 57 %).
- **P10 cage (descriptive):** RTH close inside [put wall, call wall] 68 % overall; long-γ 77 %, neutral 72 %, short-γ 59 %. The cage is a useful *range prior* (where the close lands), not an entry signal.
- **NDX vs QQQ regime agreement 36 %** — the NDX spot is approximated (QQQ × 41.0); treat as unreliable and do not use it.

## Verdicts
| Rule | n | Result | Verdict |
|---|---|---|---|
| P8 long-γ fade | 20 triggers | −0.285 R/trigger | REFUTED |
| P9 short-γ break | 136 triggers | −0.134 R/trigger, p=0.96 | REFUTED |
| Part D first-touch fade (13-pt stop) | 449 touches | 42.1 % at 1R, −0.158 R | REFUTED |
| Part D2 first-touch fade (doctrine stop) | 449 touches | 41.0 % at 1R, −0.051 R | REFUTED |
| P10 cage held | 693 sessions | 68 % | descriptive only |

The wall-fade line is **closed** on this sample (pre-registered in D2). Gamma walls remain what the record already showed: a range/cage prior and context for the read, not a mechanical entry. Any future gamma entry rule needs a new pre-registration with a different mechanism (e.g. wall + confirmation on 5m close, or wall migration from the hourly archive), not another stop/target variant.