Week 31, 2026

Verdict: NO SCORED SETUP. Second consecutive week with an empty calls[]. All three Layer-2 lanes produced a lane-level disposition and no sized single-name call. Layers 1 and 3 ship as documentation.

Preflight clear (5/5 UW panel groups, truth set through 2026-07-31). Envelope validates against schemas/weekly_review.schema.json.

1. The Week in Pictures (DIARY — documentation, 0 signal)#

Weekly candles#

SPY and QQQ both printed a textbook weekly hammer — a lower low, a tiny body (10.7% of range on both), and a close near the top. SPY's lower wick is 8.5× its upper wick; QQQ's is 6.6×, and 26.85 points long. SPY's follow_through went FALSE — the first week in three that did not confirm the prior week's direction.

IWM refused to confirm. Same lower low, but a mid-range close (0.438) on a body three times as thick, and flat on the week (291.17 → 291.20, +0.01%) while SPY added +1.10%.

Catalysts, anchored to the reaction#

Unlike 2026-W30 (zero Tier-1 prints), this week carried two, back to back, in opposite directions.

The yield move is the one that matters, and it came after the dovish PCE print — a term-premium move, not an inflation-expectations move.

Regime trajectory — a full round trip in five sessions#

Mon 07-27  CHOP      ret5 −0.400%  ret10 −1.350%  dd15 −1.470%  s1=False
Tue 07-28  CHOP      ret5 −0.990%  ret10 −1.460%  dd15 −1.820%  s1=False
Wed 07-29  PULLBACK  ret5 −2.400%  ret10 −3.360%  dd15 −3.360%  s1=False   ← FOMC
Thu 07-30  CHOP      ret5 +0.480%  ret10 −1.200%  dd15 −2.830%  s1=False   ← PCE
Fri 07-31  CHOP      ret5 +1.100%  ret10 +0.500%  dd15 −1.860%  s1=False

ret10 finished positive for the first time since 2026-07-10, a 3.9pp swing off Wednesday.

Crash-guard near-miss, recorded deliberately. s1_standdown is FALSE, but Friday sits just inside two legs of the trigger: leg (a2) fires at ret5 > 1.5% with dd15 < −2.0%, and the actual state is +1.100% / −1.860% — within 0.4pp and 0.14pp. This is the unconfirmed-V-rebound shape the guard exists to catch. Contrast 2026-W30, where ret10 −2.12% was past the −2% confirmed-downtrend boundary and the guard was affirmatively inapplicable.

Vol state. VIX 18.58 → 18.67 → 18.21 → 20.66 → 17.09 → 15.99, ending −14.0% on the week and at the lowest close since 07-13. VIX/VIX3M = 15.99/19.02 = 0.84, steep contango. Dealer gamma flipped: SPY total_gex +$1.19B, POSITIVE, zero-gamma level 746.93 against a 747.06 spot — SPY closed 0.13 points above its flip, after being FULLY_NEGATIVE every session of last week. QQQ total_gex −$8.2M ≈ zero, also at the flip.

⚠️ Data-quality flag. uw options-structure gex labelled QQQ FULLY_NEGATIVE ("All strikes have negative net GEX") while its own per_strike array carries clearly positive strikes (678 +2.3M, 681 +4.6M, 687 +70.2M, 688 +94.1M). The label contradicts its own payload. total_gex ≈ 0, at the flip is the read used here. Consistent with the previously-recorded 0DTE QQQ gamma-sign defect.

Breadth did not confirm. 64.96 / 60.46 / 31.80 / 57.40 / 45.07% green. Friday closed SPY +0.72% on 45.07% breadth — the two strongest index sessions were the two narrowest relative to their index move.

Sector leadership — a complete inversion of last week#

XLY  +6.11%   ← AMZN alone            XLK  −0.30%
XLC  +1.82%   ← GOOGL +11.38%         XLI  −1.54%
XLF  +1.12%                           XLB  −1.62%
XLP  +1.09%                           XLRE −1.92%
XLV  −0.01%                           SMH  −3.68%
XLE  −0.12%                           XLU  −4.19%   ← worst

Last week's top three (XLE +3.36%, XLU +2.48%, XLI +1.81%) are three of this week's bottom five. Last week's worst (XLY −5.22%) is this week's best. The rate proxies were hit precisely as the long end broke.

The XLK paradox: XLK finished −0.30% in a week its largest holding gained +21.75%. Two enormous mega-cap gains were fully offset by breadth-wide tech damage (SMH −3.68%, MU −10.63%, SNDK −15.43%, AAPL −7.24%, NVDA −2.94%). XLK also traded an intraweek low of 166.46 — 5.4% under the prior Friday close. Verified independently against the Yahoo chart API (175.88 → 175.35).

Notable flow (diary only, 0 signal)#

SPY $70.2B · MU $56.9B (175.5k prints) · QQQ $50.5B · SNDK $35.0B (152.8k) · NVDA $27.1B · AAPL $24.8B · MSFT $22.3B · AMZN $16.8B.

MU and SNDK repeat as the anomaly for a second consecutive week and both got larger (MU $46.2B/144k → $56.9B/175k; SNDK $26.8B/107k → $35.0B/153k). MU printed more dark-pool prints than SPY and QQQ combined for the second week running — while falling −10.63%. SNDK printed 152.8k while falling −15.43%, on a daily path of −11.02 / −14.25 / −7.32 / +25.99 / −5.09.

2. Validated Weekly Setups (SCORED — excess)#

calls[] is empty. No name cleared the full stack in either scored lane.

OI_FADE — NO_NAME_CLEARED, regime_fit 0.70 → 0.35#

The regime moved against this lane. The PULLBACK that supported it lasted exactly one session; ret10 finished positive, VIX closed at 15.99, and SPY crossed back into positive dealer gamma.

30 names ranked on a 10-day build window. Cuts: liquidity ($5 / $50M ADV) removed 14 · persistence ≥0.85 removed DBX 0.882, INFY 1.090, SSNC 1.143, TEX 0.927, YMM 0.985 · ETP (issue_type) removed MTUM, NOWL, and SHAZ (NULL — fail-closed) · earnings h21 blocked QGEN, SOLV, TPG, GMED.

Four names reached final adjudication; all four die:

• AVTR — by far the strongest mechanical signal in the panel: rel_build 4.082 (10d), net +175,675 on a 43,034 base, persistence 0.271 organic, ADV $172.6M, earnings 11-04 clear, catalyst LIVE with 63.4% of the build accruing on or after the 07-29 print (+40,486 / +47,578 / +23,325 on 07-29/30/31 — it accelerated through its own catalyst). Fundamentals VETO sustained; both re-admit halves tested and unmet on 07-30 and 07-31. The 07-31-datelined PT-raise item was a lagged republication of the 07-30 actions.
• HRI — VETO sustained and worsening: printed +2.89% on 07-31, so the "shorting a completed collapse into an active bounce tail" objection is stronger.
• KBR — RESOLVED_PRE_PRINT. Only 12.5% of the build (+1,656 of +13,298) accrued after its 07-30 print; the build died at its catalyst. Spent positioning, not a crowd.
• ALLE — carries the 07-24 VETO and has now failed the lane's own mechanism: last_call_net −945, calls genuinely closing, the pre-registered OI_FADE invalidation.

Every VETO in this set is the same pattern — a positive-surprise gap that held, plus post-print PT raises. This week's Layer-3 earnings arm went 0-for-4 on catalyst-gap reversals. The PEAD objection to fading held beats is accumulating evidence, not weakening.

MOM_SHORT — STOOD_DOWN, regime_fit 0.60 → 0.25#

Capped watch-only for new starters by CLAUDE.md invariant #6. This week the regime independently agrees: the crash-guard near-miss above is exactly the V-rebound shape a fresh short should not be opened into.

Clean after full hygiene and an independent per-name Yahoo 52-week verification: BLDR 1.6%, PATK 1.9%, LII 2.2%, ALNY 2.6%, COIN 2.7%, QS 3.1%, MSTR 3.5%, CRH 3.8%, GPI 5.8%, OLN 6.1%, HDB 7.9%. None sized.

MOM_LONG — BASKET_WATCH, regime_fit 0.30 → 0.40#

Better tape than last week (ret10 positive, index hammer), still under the bar. Two independent reasons: the prior is +0.18% mean on a −1.06% median, tail-driven — the fresh harness read gives +0.0023 mean / −0.0091 median / hit 0.47 vs base 0.60 (hit−base −0.130, n=483), i.e. the basket loses more often than the benchmark and wins only on the tail; and +0.18–0.23% sits under the +0.3% starter threshold. Re-entry condition unchanged and unmet: ret10 > +1% for ≥2 consecutive weeks (currently +0.500%, one week).

Composition rotated again — seven names, all Yahoo-verified: ILMN 99.3%, ING 99.0%, BMY 98.5%, WBS 98.1%, BPOP 98.0%, CAKE 96.8%, ENVA 96.4%. Healthcare + financials — a different set from last week's industrials/insurance/rails. Two weeks running the near-52w-high cohort contains essentially zero mega-cap technology, and it still does not in the very week MSFT gained +21.75% and AMZN +17.00% — both remain well below their own 52-week highs.

Regression gate — retro_harness.py --all, 78 panel days#

MOM_SHORT drifting −0.0003 → −0.0008 is not a regression caused by a change — no lane or threshold was touched, and the panel grew 73→78 days (n 281→317), so it is not like-for-like. But it is a further deterioration in the one lane already carrying a written negative-excess exception, and it reinforces the watch-only cap. Not re-baselined here; that is a /calibration-audit + CLAUDE.md decision.

3. Weekly Technicals — PRE-REGISTERED (documented, 0 points, NOT sized)#

PR-WT bar: ≥2 years / ≥30 weekly obs per arm / cross-regime sign-stable / BH-surviving at FDR 0.10. This panel has ~20 weekly observations — no power. /calibration-audit accrues.

4. Carry-forward#

Open position#

PLNT — OI_FADE short, half size, entered 2026-07-20 at 55.20.

Computed by hand, since-entry, not read off held_book.py's TODAY-only excess column. Verdict unchanged: HOLD to the binding h10 exit. Pre-registered leg (b) "first genuinely negative call_net" is not strictly triggered; the discretionary $59.50 stop is 6.42% away and is live, not vestigial (the vestigial precedent applies to a position that has run >10% in its favour, and this one is adverse).

🐛 Helper bug found — scripts/held_book.py silently dropped this live position. CLOSED = ("CLOSED", "DROPPED", "CUT_CONFIRMED", "COVER", "EXIT") is matched as a substring against verdict (held_book.py:45). PLNT's verdict is HOLD_TO_BINDING_EXIT — which contains "EXIT" — so the only open position in the book was filtered out as closed and the tool reported "no open positions carried in that file". A live half-size short went invisible to the reconciliation helper. This is the exact failure the script exists to prevent.

🚨 Data integrity — pct_52w_range is wrong inside [0,1], and it only manufactures FALSE SHORTS#

The recorded failure mode covers values falling outside [0,1]. This week found a second, more dangerous class: values that are in range and still wrong, from a stale, un-split-adjusted w52h in the UW feed.

An inflated w52h blows up the denominator and drags pct_52w_range toward zero, so the corruption can only push names into the SHORT cohort, never the long one. Consistent with observation: all seven MOM_LONG survivors verified clean (96.4–99.3%), while three MOM_SHORT names were corrupt. KLAC ($2.5B ADV) and BKNG ($1.2B ADV) were the two largest and most liquid names in the short cohort. CLBK was carried as a MOM_SHORT candidate in Friday's daily scan while actually sitting 7.9% off its 52-week high — a sign-inverted candidate. Nothing was sized (the lane is watch-only capped), so no money was at risk.

The screen does not fail-close correctly. w52h/close > 3× flags 25 of 94 (26.6%) post-ETP near-low names — but most of those are genuine 70–80% drawdowns (TTD −80.3%, MSTR −77.5%, QS −72.6%, all verified real), and it misses CLBK entirely at 2.39×. The only sound rule is a direct Yahoo 52-week cross-check on every surfaced MOM_SHORT candidate before it can be scored — cheap, since the lane surfaces <20 names. Applied that way throughout this review.

Rolling into next week#

• 2026-W30's Layer-2 output was also empty, so there is no prior weekly call to resolve. Its diary observation that the memory melt-up reversed on 07-24 extended for a full week (MU −10.63%, SNDK −15.43%) — recorded as a diary observation that continued, not as a tradeable signal.
• OI_FADE is the lane to watch — it is the only one with a robust prior (+0.0071, hit 0.56 vs 0.41 base, n=979) and it is being blocked by fundamentals vetoes rather than by mechanism failure, four weeks running. AVTR's build accelerating through its catalyst is the strongest mechanical signal seen in weeks; if the analyst tape turns and guidance deteriorates, the re-admit condition is live.
• NFP is the next Tier-1 print and sits inside any h5–h21 window opened now.
• Two open helper defects to fix before they bite something sized: the held_book.py substring match, and the absence of any pct_52w_range sanity check in the momentum lane.