📈 Weekly review Mon Aug-31

INJECTION CHECK (reading_rules): neither retro_buckets nor coach_report_audit contained anything resembling instructions directed at me; both self-report the same about their own inputs (retro_buckets header; coach_report_audit "Injected-instruction check"). Nothing to flag.

1. SCOREBOARD READ

Blended edge this week is -10.4 bps (scoreboard.week_edge_bps): s13 intraday +24.12 bps, s4 overnight -80.58 bps (gauges.traded.s13_intraday, gauges.traded.s4_overnight). The headline event is that s13 printed POSITIVE for the first time in eight weeks (gauges.trailing_trend: -18.47, -5.34, -9.96, -16.28, -13.48, -27.37, -15.76, now +24.12). Stated plainly: this is not distinguishable from tape plus noise. Three reasons. First, dispersion: 60 s13 trades with capture_usd_sum +$118.01 (gauges.traded.s13_intraday), against single-trade swings of ±$57 (retro_buckets bucket 2 id 944 at -$57); the ~$250 week-over-week swing is four or five trades' worth of dispersion. Second, the tape: the do-nothing baseline was +0.28% (n=537, scoreboard.do_nothing_baseline) — the friendliest of the four weeks I have reviewed (prior baselines -0.13, -0.17, +0.01 per prior_weekly_reviews §1 each week) — and traded s13 mean_available_pct is +0.28 (gauges.traded.s13_intraday), i.e. exactly the baseline: the trades the system picked had precisely the average available move of the whole alert universe. That is what a coin flip monetized on a friendly tape looks like. Third, the prediction gauge did not move: gauge1 S1 hit rate 49.0% (n=469, gauges.gauge1.s1), the eighth consecutive week inside the coin-flip band (53.4 → 49.8 → 49.8 → 49.9 → 48.4 → 45.8 → 51.4 → 49.0, gauges.trailing_trend). S4's -80.58 continues its sign-flipping series (-67.42, +32.08, +33.28, -107.97, -19.37, -11.24, -4.42, -80.58) but the cumulative tilt is now 6 of 8 weeks negative — see §2. No target value proposed.

Missing deliverable, THIRD consecutive week, and per my own 2026-08-24 commitment this is now an escalation, not a re-issue: the MAX_PUB_TO_DETECT_MIN=15 gate (first proposed 2026-08-10, re-issued twice) was again not shipped. Verified mechanically: no such constant in threshold_constants, no carrying PR in predictions_to_grade, no matching skip rationale in strategy_aggregates.top_skip_rationales. The 2026-08-24 review (prior_weekly_reviews §5) stated that on a third non-ship the honest deliverable is: the pilot cannot answer its own north-star question for S1-S3 — "signal never existed" vs "signal alive at publish, dead by the ~12.6-minute median detection delay" (article_funnel.latency_pub_to_detect_secs median 756.0s) — because its designated intervention channel does not execute. That is now the finding. It is corroborated on the coach side: coach_report_audit §3 documents the coach's two soundest recommendations of the week (S4 give-back stop, 2 days; S4 shadow mode, 2 days with the cited cost of inaction growing in the interval) each repeated verbatim and never actioned, while every one of the 41 genuine clauses in predictions_to_grade that DID ship is live. The pipeline ships code; it does not ship the code its own measurement organs ask for. One incidental confirmation of why the unshipped gate's grading metric had to be baseline-relative: the bet's success metric (traded s13 mean_available_pct ≥ +0.15) was "achieved" this week at +0.28 with the gate never built — driven entirely by the baseline drifting to +0.28. Absolute mean_available_pct is tape-contaminated; any future version of that bet must grade the spread over the do-nothing baseline.

2. CHAIN DIAGNOSIS

Link 1 — Prediction (all alerts, gauge1). S1: 49.0% hit rate (n=469, binomial SE ~2.3pp), mean_available_pct +0.09 against mean_abs_available_pct 0.87 (gauges.gauge1.s1) — eighth consecutive coin-flip week. S3: 52.6% (n=171, SE ~3.8pp) — coin flip (gauges.gauge1.s3). S2: 60.0% (n=35, SE ~8.3pp, gauges.gauge1.s2) — S2's first above-50 week, but only ~1.2 SE above coin flip, and pooled across the four reviewed weeks S2 sits at roughly 58/146 ≈ 40% (8/36, 6/24, 23/51 per prior_weekly_reviews §2 each week, plus 21/35 now) — one good thin week does not clear a pooled deficit. Materiality calibration (gauges.gauge1.materiality_calibration): 60-69 → 48.5% (n=342), 70-79 → 59.4% (n=32, thin), 80-89 → 47.0% (n=83), 90-100 → 50.0% (n=12). Fourth consecutive week: flat and non-monotone everywhere the sample is adequate; materiality does not rank prediction skill.

Link 2 — Transit. Priced-at-detection 21.7% (208/959, gauges.priced_at_detection). On the traded cohort transit is again not the leak — it added money on both books: s13 transit_usd_sum +$15.56, s4 transit_usd_sum +$93.58 (gauges.traded). Blended pub→detect median 756.0s ≈ 12.6m; per provider, alpaca 234.7s vs eventregistry 761.8s (article_funnel.by_provider) — a 3.2x gap, stable in the 2-3.5x range of the last four weeks. No per-ticker transit-leak field exists in per_ticker (#87 still not live), so that read remains unavailable.

Link 3 — Selection. Split reading this week, and it is the crux. s13: available_usd_sum +$133.57, mean_available_pct +0.28 (gauges.traded.s13_intraday) — positive for the first time in four reviews, but exactly equal to the +0.28% do-nothing baseline (scoreboard.do_nothing_baseline): zero selection alpha, the tape rose and the picks rose with it. s4: available_usd_sum -$100.36, mean_available_pct -0.42 (gauges.traded.s4_overnight) — 0.70pp WORSE than doing nothing, on a friendly tape. Attribution agrees: S4 selection_pnl -214.03 out of S4's -$193.94 total, with market_pnl +23.94 (attribution.decomposition.per_strategy.S4) — S4 rode a helpful tape and anti-selected its way through it, at a 33.3% win rate (10/30, scoreboard.per_strategy S4). Per-ticker dispersion (per_ticker rows): worst QCOM -250.71, TSLA -131.24 (with 67.0% priced rate — TSLA's news arrives dead), AAPL -96.43; best AMD +85.94, TSM +61.67, MSFT +39.54. Surprise rate 0.0% on all 15 tickers. The two-week priced_block pattern I flagged 08-24 (blocked cohort making money) did NOT extend: this week the PRICED block saved money — blocked_pnl_pct_sum -17.22, missed_wins 90 vs avoided_losses 104 (gauges.declined.priced_block). Sign bounces; dropped.

Link 4 — Capture. Exit slippage small and positive: +$20.15 over 60 exits, mean +$0.34 (attribution.exit_slippage). MFE-realized remains degenerate (S1 +10.30, S4 -5.00, attribution.mfe_realized — fractions outside [0,1] are unusable except as a near-zero-denominator symptom). Timing_pnl -30.25 total, S1 -35.34 (attribution.decomposition) — by far the smallest of the four reviewed weeks (-311.13, -162.64, -120.57, now -30.25). The entry-side timing leak that carried the last two bucket promotions shrank markedly this week.

SINGLE WEAKEST LINK: prediction, fourth consecutive week — but the conviction now concentrates on S4's overnight-drift thesis specifically, because it is the one directional call the data can convict without the unshipped latency gate. Convicting numbers: S4 win rate 33.3% on 30 closed trades (scoreboard.per_strategy); mean available move of -0.42% against a +0.28% do-nothing baseline (gauges.traded.s4_overnight, scoreboard.do_nothing_baseline) — a -0.70pp selection spread; selection_pnl -214.03 on a week when S4's market_pnl was +23.94 (attribution.decomposition.per_strategy.S4); 6 of 8 trailing weeks negative (gauges.trailing_trend); and retro_buckets bucket 2 (-$212.91, 10 trades, "flatly wrong directional signal") cross-checks against all of it. Meanwhile S1-S3's coin flip (gauges.gauge1) produced a tape-level week — no evidence of edge, no evidence of a new leak — and its publish-vs-detection confound remains blocked by non-shipping, not by data (§1). Timing/exits are exonerated this week by their own numbers (timing -30.25, slippage +$20.15).

3. PREDICTION GRADES

All genuine clauses carry deploy_status "live" — a single-instant read at the week's last session (2026-09-04), cited per reading_rules; where that gives no assurance the change ran for the cited numbers, I stay at not-yet-measurable. Backfilled clauses (#108 and below) are attribution material only, not graded. Oscillation guard: no grade below reverses a prior week's grade.

GRADE #253 cost: not-yet-measurable — deploy_status live; no shadow_sessions.elapsed_ms or page-count field in any mechanical section. Fourth week at this status. GRADE #242 signal: not-yet-measurable — deploy_status live; no mechanical section counts counter-move S4 entries or exposes s4_alignment_skips (coach_report_audit's 09-01 reference to those skips is untrusted, one hop removed). GRADE #241 cost: not-yet-measurable — deploy_status live; no page-count data in any mechanical section (the 7,520 baseline lives only in the clause text). GRADE #228 freshness: not-yet-measurable — deploy_status live; no stale-at-fill cancel count anywhere mechanical (gauges.traded transit_queued_unknown_n=0 measures a different thing). GRADE #226 freshness: not-yet-measurable — deploy_status live; no absorbed-move counterfactual ledger in gauges.declined or strategy_aggregates. Noting without grading: retro_buckets bucket 4 (untrusted) still describes chased in-direction entries this week (ids 861, 886, 942). GRADE #222 signal: not-yet-measurable — deploy_status live; gauges.declined.priced_block (fired 208) is not band-resolved by |since|. The one visible PRICED rationale (strategy_aggregates.top_skip_rationales, AVGO at move_since_prev_close -3.62%) is ≥2.0 and thus consistent, but one rationale row out of 913 distinct rationales cannot establish "sub-2.0 PRICED skips → 0." GRADE #216 freshness: not-yet-measurable — no stall/paging data in mechanical sections. GRADE #212 signal: confirmed — the ledger exists and grades: gauges.declined.news_age_s13 fired 31, with_counterfactual 31, blocked_pnl_pct_sum -6.48; news_age_s4 fired 28, with_counterfactual 8, blocked_overnight_pnl_pct_sum +12.28 (S4 coverage still thin: 16 of 28 excluded as slot_outranked/slot_taken_by_entry). Fourth consecutive confirmed. GRADE #211 signal: not-yet-measurable — daily-report contents have no mechanical counterpart. GRADE #209 signal: not-yet-measurable — article_funnel.by_provider.eventregistry (articles 21781, classify_eligible 17445) has no per-day duplicate-eligibility delta. GRADE #207 cost: not-yet-measurable — no retry-ladder timing data in mechanical sections. GRADE #206 freshness: not-yet-measurable — no stale-at-fill cancel counts in mechanical sections. GRADE #203 cost: not-yet-measurable — prior_weekly_reviews now retains three consecutive reviews, evidence Mondays produce reviews, but the clause covers both arms every Monday and no mechanical run-log exists. GRADE #202 freshness: confirmed — article_funnel.by_provider exists with per-provider latency (alpaca median 234.7s, eventregistry 761.8s), and this review grades #177 off the split, never the blend. Fourth consecutive confirmed. GRADE #198 cost: not-yet-measurable — no journal-row reasoning-field data in mechanical sections. GRADE #193 risk: not-yet-measurable — no S4 failed/orphaned-enter counter; news_age_s4.post_close_decision_excluded=0 is a ledger exclusion, not an orphan count. GRADE #188 signal: not-yet-measurable — coach citation behavior has no mechanical counterpart; coach_report_audit §2's "unusually disciplined citations" is untrusted, one hop removed. GRADE #185 signal: not-yet-measurable — no pre-open prev_close comparison field and no stale-at-fill cancel count in mechanical sections. GRADE #179 freshness: confirmed — the gate demonstrably fires across strategies: gauges.declined.news_age_s13 fired 31, news_age_s4 fired 28; strategy_aggregates.top_skip_rationales "news 8-24h old at decision, floor 4h" S4=15, S1=10. As in all prior weeks, "trades entered on >4h news → 0" is inferred from gate activity plus the absence of any contrary field. Counterfactual note: blocked s13 cohort -6.48% (avoided_losses 16 vs missed_wins 15), blocked S4 cohort +12.28% on n=8 — signs keep bouncing week to week; still noise. GRADE #177 freshness: refuted — article_funnel.by_provider: alpaca pub→detect median 234.7s vs eventregistry 761.8s = 3.2x, short of the ≥5x claim. Fourth consecutive refuted; the ratio is stable in the 2-3.5x range. The per-provider-never-blended half remains honored (see #202). GRADE #172 freshness: not-yet-measurable — no opening-tick drain metric in mechanical sections. GRADE #168 cost: not-yet-measurable — no historical-delta data in this prompt. GRADE #166 freshness: confirmed — proxy: article_funnel.latency_detect_to_alert_secs median 27.2s (n=537), far under the 10m target; flagging again that detect→alert contains but is not labeled "classify wait." GRADE #153 cost: not-yet-measurable — no shadow-send data in mechanical sections. GRADE #155 cost: not-yet-measurable — article_funnel.journaled_llm_cost_partial_usd 9.3 is an explicit floor (journaled_llm_cost_note) with no per-call baseline split. GRADE #151 cost: not-yet-measurable — no audit fire/no-fire flip data in mechanical sections. GRADE #147 risk: not-yet-measurable — no mechanical tripwire counter; retro_buckets and coach_report_audit both report no injection this week (untrusted). GRADE #146 risk: not-yet-measurable — same class as #147; no mechanical measurement. GRADE #143 signal: confirmed — scoreboard.per_strategy: S2 1 trade +$4.30 + S3 6 trades +$29.15 = +$33.45 combined, the first positive combined week and well inside "toward ≥ -$30" from ≈ -$108, with the floor visibly firing (gauges.declined.materiality_floor fired 96; strategy_aggregates.top_skip_rationales "materiality 65 below floor 75": S3 47+32). Fourth consecutive confirmed. Standing caveat: the floor caps volume, not quality — its blocked cohort would have MADE +7.75% this week (missed_wins 53 vs avoided_losses 40, gauges.declined.materiality_floor), and pooled S2 hit rate is still ~40% (§2). GRADE #140 cost: not-yet-measurable — no timeout/page data in mechanical sections. GRADE #139 cost: not-yet-measurable — no error-envelope data in mechanical sections. GRADE #138 cost: not-yet-measurable — no auditor fire-count in mechanical sections. GRADE #137 freshness: confirmed — publish→alert ≈ 783.2s ≈ 13.1m (article_funnel: pub→detect median 756.0s + detect→alert 27.2s), stable versus 13.4m/13.4m/12.9m in prior weeks and consistent with the honest-baseline claim; ER, 98% of articles, sits at 761.8s per by_provider. GRADE #134 cost: not-yet-measurable — no page-storm/WARNING data in mechanical sections. GRADE #131 cost: not-yet-measurable — scaffolding-citation rate has no mechanical counterpart. GRADE #130 signal: not-yet-measurable — no delimiter-breakout counter in mechanical sections. GRADE #128 cost: not-yet-measurable — no 400/self-heal data in mechanical sections. GRADE #126 cost: not-yet-measurable — no prompt_chars measurement in this prompt. GRADE #121 cost: not-yet-measurable — no timeout-page or shadow-row counts in mechanical sections. GRADE #120 risk: confirmed — measurable half: gauges.declined.tape_gate S1 fired 65, with_counterfactual 63, missed_wins 19 vs avoided_losses 44, blocked_pnl_pct_sum -22.2% (≈ -0.35% per blocked trade — clearly loss-avoiding, better than the "≈ 0" the clause promised, and consistent with 08-24's -0.27%/trade). The "worst-day S1 loss ↓" half remains unassessed (no worst-day field). GRADE #117 signal: confirmed — proxy: processed/classify_eligible = 17818/(373+17445) = 100.0% (article_funnel + by_provider). Second consecutive clean week after the 08-17 outage-week dip, which cements last week's rule-following restore; standing caveat that this proxy cannot isolate the capped-tick tail specifically.

4. BUCKET PROMOTION

Newly promoted: retro_buckets bucket 2 — S4's overnight-drift thesis simply wrong (10 trades, -$212.91). Mechanical cross-check per reading_rules before repeating the LLM summary: scoreboard.per_strategy S4 shows -$193.94 at 33.3% win rate; attribution.decomposition.per_strategy.S4 shows selection_pnl -214.03 against market_pnl +23.94; gauges.traded.s4_overnight shows mean_available_pct -0.42 against a +0.28% do-nothing baseline. All four numbers convict the same thing the retros say in words. Recurrence across weeks: S4 drift-failure was retro_buckets bucket 3 in the 08-17 review (-$133.88, prior_weekly_reviews §4) and S4's weekly edge is negative 6 of 8 trailing weeks (gauges.trailing_trend). The mechanical rule it deserves is the coach's own twice-repeated, never-shipped recommendation (coach_report_audit §3 finding 3b): S4 shadow mode. That is this week's §5 bet.

Standing promotion, weakened: the stale/chased-entry pattern (promoted 08-17, re-promoted 08-24). It still appears — retro_buckets bucket 4, 11 trades, -$76.32 — and remains unshipped (no fill-gap cancel or recap-timestamp-anchor PR in predictions_to_grade). But its mechanical corroboration shrank sharply: timing_pnl -30.25 this week versus -120.57 / -162.64 / -311.13 in the three prior weeks (attribution.decomposition; prior_weekly_reviews §2 each week), and exit slippage is again benign (+$20.15, attribution.exit_slippage). One week of shrinkage on a friendly tape does not close it; it stays promoted but is no longer the most expensive open defect — S4 is.

Previously promoted, still absent: the "short run over by a violent adverse rally" tail (promoted 08-10, downgraded to dormant 08-24). Third consecutive week with no such bucket; the worst single trade this week is ~-$57 (retro_buckets bucket 2, id 944). Remains dormant, reopen on recurrence.

Observed, not promoted: the netting mechanism touched 10 trades this week (retro_buckets bucket 8 plus its cross-note), and coach_report_audit §3 calls it the week's most persistent unrecommended recurring pattern — but its measured net cost this week is +$6.42 and no mechanical section isolates it; too thin. Also explicitly not promoted again: bucket 5 (exits riding winners back to losses, -$51.81) — the three prior reviews' reasoning stands without oscillation: while gauge1 S1 reads 49.0% coin flip, exit redesign optimizes the monetization of noise.

5. ONE BIGGER BET

Target: prediction, the weakest link named in (2), at its one convictable point — S4's overnight directional thesis. Per my 2026-08-24 commitment (prior_weekly_reviews §5), I am NOT re-issuing the MAX_PUB_TO_DETECT_MIN gate a fourth time; its triple non-shipment is escalated in §1 as a governance finding, and the S1-S3 publish-vs-detection question is formally recorded as blocked by non-action. This bet deliberately coincides with the coach's twice-pushed, never-actioned shadow-mode recommendation (coach_report_audit §3 finding 3b) — if it fails to ship too, the reviewer channel (3 misses) and the coach channel (2+ misses) will have independently evidenced the same execution defect on the same class of ask.

• 2026-09-07 owner: S4 goes shadow — S4 entries journal as qty-0 with every gate, verdict, slot-resolution, and exit rule unchanged, and the shadow cohort continues to grade weekly in gauges.traded.s4_overnight (available/transit/capture as today) — expect: signal, the shadow S4 cohort's selection spread (s4_overnight.mean_available_pct minus scoreboard.do_nothing_baseline.mean_pct; this week -0.42 - 0.28 = -0.70pp on n=30) stays ≤ 0 for two consecutive weeks, confirming the overnight-drift thesis selects worse than doing nothing and retiring S4 as a calibrated no-signal conclusion; if instead the shadow spread prints ≥ +0.2pp in either week, the thesis has real drift and the leak is monetization (exit/entry), reopening retro_buckets bucket 5's exit question for S4 specifically.

Grading notes for next week: the metric is the baseline-relative spread, not raw mean_available_pct — §1 showed raw available moved +0.42pp in one week purely on tape drift, so an absolute threshold is tape-contaminated and ungradeable. n≈30/week is thin; that is why the retire branch requires two consecutive weeks, and why the live-P&L consequence (removing a -$193.94/week drag, scoreboard.per_strategy S4) is a side effect, never the graded metric — no edge bps target is proposed or implied. If this bet is unshipped next week, do not re-issue it either: at that point the correct weekly deliverable is a standalone finding that the pilot's intervention channel is inoperative across both requesting organs, with the §1 and coach_report_audit §3 ledgers as evidence.

1. SCOREBOARD READ

This week’s blended edge was -10.4 bps (scoreboard.week_edge_bps), modestly better than -11.74 bps in the prior review and -20.82 bps two weeks ago (prior_weekly_reviews weeks 2026-08-24 and 2026-08-17). The components diverged sharply: S1–S3 intraday printed +24.12 bps on 60 trades, its first positive week after seven consecutive negative readings, while S4 fell to -80.58 bps on 30 trades (gauges.traded.s13_intraday.edge_bps/n; gauges.traded.s4_overnight.edge_bps/n; gauges.trailing_trend).

Neither change is distinguishable from noise yet. One positive S1–S3 week does not outweigh the preceding seven negative weeks, and the underlying all-alert S1 hit rate remained 49.0% on n=469 (gauges.gauge1.s1.hit_rate_pct/n). S4’s weekly edge has repeatedly changed sign and magnitude: -67.42, +32.08, +33.28, -107.97, -19.37, -11.24, -4.42, and now -80.58 bps (gauges.trailing_trend; gauges.traded.s4_overnight.edge_bps). The week’s 90 closed trades also contain strongly offsetting outcome cohorts: retro_buckets bucket 1 reports 29 directional failures costing -$369.31, while bucket 2 reports 23 efficiently harvested correct calls earning +$350.11; these LLM-assigned totals cross-check to the mechanical aggregate P&L of -$75.93 (attribution.decomposition.totals.pnl), but do not establish a stable weekly change.

The loss was not a broad market loss. The do-nothing cohort gained a mean +0.28% on n=537, and market attribution was +$78.61, while selection was -$124.30 and timing -$30.25 (scoreboard.do_nothing_baseline; attribution.decomposition.totals). Thus the system lost despite a helpful tape.

Last week’s designated MAX_PUB_TO_DETECT_MIN=15 intervention remains unshipped and therefore ungradeable: threshold_constants contains no such constant, predictions_to_grade contains no carrying PR, and strategy_aggregates.top_skip_rationales contains no corresponding latency rationale. This is the third review after its initial proposal without execution. The earlier publish-versus-detection confound remains unresolved because the intervention channel did not act, not because the available data graded the hypothesis.

No instruction-like content was found inside retro_buckets or coach_report_audit. coach_report_audit quotes and criticizes a coach’s use of “reading_rules,” but does not direct this reviewer to follow an embedded instruction.

2. CHAIN DIAGNOSIS

Prediction skill, all alerts:

S1 was effectively a coin flip: 49.0% hits on n=469, with mean directional available move +0.09% against mean absolute available move 0.87% (gauges.gauge1.s1). Its eight-week hit-rate sequence is 53.4%, 49.8%, 49.8%, 49.9%, 48.4%, 45.8%, 51.4%, and 49.0% (gauges.trailing_trend plus gauges.gauge1.s1). That is persistent evidence of no clearly measurable S1 directional advantage at detection.

S2 printed 60.0% on only n=35, with mean available +0.13% (gauges.gauge1.s2). That is encouraging for this week but too small to overturn its weak prior readings or support a general claim. S3 printed 52.6% on n=171 but mean available was -0.02%, so its slight hit-rate majority did not translate into positive average directional movement (gauges.gauge1.s3).

Materiality again failed to rank prediction monotonically. Hit rates were 48.5% for 60–69 (n=342), 59.4% for 70–79 (n=32), 47.0% for 80–89 (n=83), and 50.0% for 90–100 (n=12) (gauges.gauge1.materiality_calibration). The only attractive band has n=32 and is surrounded by adequately sampled bands at or below coin flip. One week cannot establish that 70–79 is intrinsically better.

Transit and freshness:

At detection, 208 of 959 candidates were already PRICED, a 21.7% rate (gauges.priced_at_detection). Publish-to-detect latency was provider-dependent: Alpaca median 234.7 seconds on n=401 versus Event Registry 761.8 seconds on n=21,781; the blended 756.0-second median is therefore mix-sensitive and should not be used as the provider comparison (article_funnel.by_provider; article_funnel.latency_pub_to_detect_note). Detect-to-alert added a median 27.2 seconds (article_funnel.latency_detect_to_alert_secs).

Transit was a small leak for intraday trades but a much larger one for S4. S1–S3 incurred $15.56 of transit loss, mean 0.04%, before capturing $118.01 and +24.12 bps on 60 trades (gauges.traded.s13_intraday). S4 incurred $93.58 of transit loss, mean 0.37%, converting -$100.36 available at detection into -$193.94 captured P&L and -80.58 bps (gauges.traded.s4_overnight). Transit materially worsened S4, but it did not create the original negative opportunity: S4 was already pointed at a -0.42% mean available move at detection.

There is still no per-ticker transit-leak field in per_ticker, and no #87 clause appears in predictions_to_grade, so whether transit leakage dropped by ticker is unsupported this week. The ticker-level detection-latency medians are not a substitute for decision-to-entry transit leakage.

Selection:

Selection is where this week’s mechanical loss is concentrated. Total selection attribution was -$124.30, versus timing -$30.25 (attribution.decomposition.totals). S1 selection was +$53.04, S3 +$36.61, and the one S2 trade +$0.08, but S4 selection was -$214.03 on 30 trades (attribution.decomposition.per_strategy). S4 won only 10 of 30 trades, or 33.3%, and lost -$193.94 (scoreboard.per_strategy S4). Its mean available move was already -0.42% at detection (gauges.traded.s4_overnight), demonstrating wrong-side selection before entry and exit effects.

This is not merely a one-session artifact. The previous review also found S4 selection at -$145.24 despite +$139.07 of market attribution (prior_weekly_reviews week 2026-08-24, chain diagnosis). The exact weekly magnitude remains noisy, but two consecutive weeks mechanically localize substantial losses to S4 selection.

Ticker outcomes were highly dispersed. QCOM was -250.71 bps with a 32.1% PRICED rate; TSLA was -131.24 bps with a 67.0% PRICED rate; AAPL was -96.43 bps with a 24.6% PRICED rate. Conversely AMD was +85.94 bps at 27.8% PRICED, TSM +61.67 bps at 11.6%, MSFT +39.54 bps at 17.9%, and NVDA +30.05 bps at 19.3% (per_ticker rows). No monotonic relationship between PRICED rate, latency, volatility, and ticker edge is visible this week. For example, high-PRICED TSLA was deeply negative, while low-PRICED TSM was positive; QCOM was worse than both despite an intermediate PRICED rate. ASML had no week_edge_bps and cannot be ranked.

Every ticker reported a 0.0% surprise rate over five eligible sessions (per_ticker.surprise_rate_pct/surprise_eligible_sessions). Whatever that field intends to capture, it supplied no cross-sectional discriminator this week.

Capture:

The reported intraday exit-slippage total was $20.15 across 60 exits, or $0.34 per exit, with no missing-close, post-close, or cross-day exclusions (attribution.exit_slippage). That is small relative to the -$124.30 selection loss and -$193.94 S4 P&L.

MFE-realized fractions are not interpretable as ordinary capture ratios this week: S1 was 10.3044 and S4 -4.9994, both outside the usual [0,1] range; S2 was 0.8007 on one trade and S3 0.3981 on six trades (attribution.mfe_realized). The extreme values indicate denominator/sign degeneracy, so they cannot support a claim that S1 captured ten times its available excursion or that S4 captured negative five times it. The reliable capture evidence is instead the decomposition and exit-slippage fields: timing cost -$30.25, exit slippage $20.15, versus selection cost -$124.30 (attribution.decomposition.totals; attribution.exit_slippage).

SINGLE WEAKEST LINK: selection, specifically S4 directional selection.

Prediction remains structurally weak across the broad S1 alert pool, but this week’s largest convicted pipeline failure is the subset the system selected for S4: 33.3% wins, mean available -0.42%, available P&L -$100.36 at detection, selection attribution -$214.03, and realized P&L -$193.94 on 30 trades (scoreboard.per_strategy S4; gauges.traded.s4_overnight; attribution.decomposition.per_strategy.S4). Transit then worsened those selections by $93.58, but it did not cause the initial wrong-side call. This is also the second consecutive week with strongly negative S4 selection attribution (prior_weekly_reviews week 2026-08-24). Evidence remains too thin to claim S4 is generally incapable of prediction, but it is sufficient to identify S4 selection as the current weakest link and to stop spending even paper capital while measuring it cleanly.

3. PREDICTION GRADES

Oscillation guard: five grades change from prior reviews. #179, #166, #137, and #117 move from confirmed to not-yet-measurable because the prior reviews used proxies or inferred zero violations without the clause’s actual outcome field; this is a measurement correction, not new contrary evidence. #120 moves from confirmed to refuted because this week’s directly measured blocked P&L sum is materially different from its “≈0” clause. All genuine clauses report deploy_status=live in predictions_to_grade, but that is only the final-session manifest snapshot.

GRADE #253 cost: not-yet-measurable — predictions_to_grade #253 deploy_status=live; no mechanical section reports shadow_sessions.elapsed_ms or new-page counts.

GRADE #242 signal: not-yet-measurable — predictions_to_grade #242 deploy_status=live; no mechanical field counts counter-move S4 entries or s4_alignment_skips.

GRADE #241 cost: not-yet-measurable — predictions_to_grade #241 deploy_status=live; no mechanical page-count field measures outage-day ⚠️ pages.

GRADE #228 freshness: not-yet-measurable — predictions_to_grade #228 deploy_status=live; no mechanical field counts “stale at fill” cancels or deferred retries.

GRADE #226 freshness: not-yet-measurable — predictions_to_grade #226 deploy_status=live; gauges.declined and strategy_aggregates contain no absorbed-move ledger or count of in-direction 1.5–2.0% entries.

GRADE #222 signal: not-yet-measurable — predictions_to_grade #222 deploy_status=live; gauges.declined.priced_block reports 208 fires but does not split them below versus above 2.0%, so sub-2.0 PRICED skips are uncountable.

GRADE #216 freshness: not-yet-measurable — predictions_to_grade #216 deploy_status=live; no mechanical stall-page or stale-quote page count is supplied.

GRADE #212 signal: confirmed — predictions_to_grade #212 deploy_status=live; gauges.declined.news_age_s13 reports 31/31 graded fires and -6.48 blocked P&L percentage points, while news_age_s4 reports 28 fires and 8 counterfactuals with +12.28 blocked overnight P&L percentage points.

GRADE #211 signal: not-yet-measurable — predictions_to_grade #211 deploy_status=live; no mechanical field counts closed trades absent from every daily report.

GRADE #209 signal: not-yet-measurable — predictions_to_grade #209 deploy_status=live; article_funnel.by_provider.eventregistry reports 17,445 classify-eligible articles but no ER duplicate-eligibility increment per day.

GRADE #207 cost: not-yet-measurable — predictions_to_grade #207 deploy_status=live; no mechanical retry-ladder timing or attempt-budget field is present.

GRADE #206 freshness: not-yet-measurable — predictions_to_grade #206 deploy_status=live; no mechanical field counts sub-2.0 stale-at-fill cancels.

GRADE #203 cost: not-yet-measurable — predictions_to_grade #203 deploy_status=live; retained prior_weekly_reviews show reviews exist, but no mechanical Monday run log measures Mondays ending with zero reviews across both arms.

GRADE #202 freshness: confirmed — predictions_to_grade #202 deploy_status=live; article_funnel.by_provider separately reports Alpaca and Event Registry counts and latency, and this review grades #177 from those provider rows rather than the blended median.

GRADE #198 cost: not-yet-measurable — predictions_to_grade #198 deploy_status=live; no mechanical journal-field audit reports NULL reasoning counts by role.

GRADE #193 risk: not-yet-measurable — predictions_to_grade #193 deploy_status=live; no mechanical field counts S4 post-close failed or orphaned entries.

GRADE #188 signal: not-yet-measurable — predictions_to_grade #188 deploy_status=live; no mechanical field measures coach recommendations citing named aggregate rows, and coach_report_audit is one-hop LLM output rather than the required mechanical outcome.

GRADE #185 signal: not-yet-measurable — predictions_to_grade #185 deploy_status=live; no pre-open prev_close comparison or stale-at-fill cancel count is provided mechanically.

GRADE #179 freshness: not-yet-measurable — predictions_to_grade #179 deploy_status=live; gauges.declined.news_age_s13 fired 31 times and news_age_s4 fired 28 times, proving gate activity, but no mechanical entered-trade age field establishes that trades on news older than four hours were exactly zero. REVERSAL from prior confirmed grades due to correcting an absence-of-violations inference.

GRADE #177 freshness: refuted — predictions_to_grade #177 deploy_status=live; article_funnel.by_provider gives Alpaca median 234.7 seconds versus Event Registry 761.8 seconds, only 3.25x faster rather than at least 5x. The provider split exists, but the numerical speed claim fails.

GRADE #172 freshness: not-yet-measurable — predictions_to_grade #172 deploy_status=live; no opening-tick drain-duration field is supplied.

GRADE #168 cost: not-yet-measurable — predictions_to_grade #168 deploy_status=live; no mechanical comparison over the referenced 422 historical trades is present.

GRADE #166 freshness: not-yet-measurable — predictions_to_grade #166 deploy_status=live; article_funnel.latency_detect_to_alert_secs median is 27.2 seconds, but the clause targets alerting-article classify wait specifically, which is not separately measured. REVERSAL from prior confirmed proxy grades.

GRADE #153 cost: not-yet-measurable — predictions_to_grade #153 deploy_status=live; no mechanical restart-day duplicate-shadow-send count exists.

GRADE #155 cost: not-yet-measurable — predictions_to_grade #155 deploy_status=live; article_funnel.journaled_llm_cost_partial_usd is $9.30 but is explicitly a partial floor and contains no weekly call-count or daily-spend comparison.

GRADE #151 cost: not-yet-measurable — predictions_to_grade #151 deploy_status=live; no mechanical late-exit audit fire/no-fire comparison is present.

GRADE #147 risk: not-yet-measurable — predictions_to_grade #147 deploy_status=live; no mechanical prompt-tripwire count measures live tags reaching a prompt.

GRADE #146 risk: not-yet-measurable — predictions_to_grade #146 deploy_status=live; no mechanical hostile-versus-benign rationale test outcome is supplied.

GRADE #143 signal: confirmed — predictions_to_grade #143 deploy_status=live; scoreboard.per_strategy reports S2 +$4.30 on one trade and S3 +$29.15 on six trades, +$33.45 combined versus the clause’s approximately -$108 baseline, while gauges.declined.materiality_floor fired 96 times.

GRADE #140 cost: not-yet-measurable — predictions_to_grade #140 deploy_status=live; no mechanical shadow-coach timeout or watchdog-kill count is supplied.

GRADE #139 cost: not-yet-measurable — predictions_to_grade #139 deploy_status=live; no mechanical malformed-envelope error-name field is present.

GRADE #138 cost: not-yet-measurable — predictions_to_grade #138 deploy_status=live; no mechanical auditor fire count or no-call count is supplied.

GRADE #137 freshness: not-yet-measurable — predictions_to_grade #137 deploy_status=live; article_funnel separately reports publish-to-detect median 756.0 seconds and detect-to-alert median 27.2 seconds, but it does not report the clause’s publish-to-alert median, and separate medians cannot be added into that metric. REVERSAL from prior confirmed proxy grades.

GRADE #134 cost: not-yet-measurable — predictions_to_grade #134 deploy_status=live; no mechanical page-storm or per-attempt WARNING count is present.

GRADE #131 cost: not-yet-measurable — predictions_to_grade #131 deploy_status=live; no mechanical scaffolding-as-fact citation rate is supplied.

GRADE #130 signal: not-yet-measurable — predictions_to_grade #130 deploy_status=live; no mechanical delimiter-breakout or clean-article regression test is present.

GRADE #128 cost: not-yet-measurable — predictions_to_grade #128 deploy_status=live; no mechanical mandatory-reasoning 400, self-heal, page, cost, or latency field is supplied.

GRADE #126 cost: not-yet-measurable — predictions_to_grade #126 deploy_status=live; no coach prompt_chars or report-quality measurement appears in the mechanical sections.

GRADE #121 cost: not-yet-measurable — predictions_to_grade #121 deploy_status=live; no mechanical coach-timeout page or shadow-row-per-day count is present.

GRADE #120 risk: refuted — predictions_to_grade #120 deploy_status=live; gauges.declined.tape_gate.by_strategy S1 reports 63 counterfactuals, 19 missed wins, 44 avoided losses, and blocked_pnl_pct_sum -22.20, materially different from the clause’s net gated P&L “≈0”; no worst-day field measures the other half. REVERSAL from prior confirmed grades on the directly measured numerical clause.

GRADE #117 signal: not-yet-measurable — predictions_to_grade #117 deploy_status=live; article_funnel shows processed 17,818 and provider classify_eligible totaling 17,818, but that 100% aggregate does not isolate the clause’s capped-tick tail rate. REVERSAL from prior confirmed proxy grades.

4. BUCKET PROMOTION

Promote “Directional thesis failed,” retro_buckets bucket 1, to the stable and costly pattern that now deserves a mechanical capital gate. Stage A assigns 29 trades and -$369.31 to this bucket. That LLM assignment is mechanically corroborated at the aggregate level by selection attribution of -$124.30 overall and, most importantly, -$214.03 for S4, together with S4’s 33.3% win rate and -0.42% mean available move at detection (attribution.decomposition.totals/per_strategy.S4; scoreboard.per_strategy S4; gauges.traded.s4_overnight). The previous review also measured S4 selection at -$145.24 (prior_weekly_reviews week 2026-08-24). The exact bucket membership remains one-hop LLM interpretation, but the recurring S4 selection failure is mechanical. The appropriate response is not another tuned exit; it is a capital gate that preserves shadow measurement, implemented as the sole bigger bet below.

The previously promoted stale/chased-entry pattern did not disappear. It appears again as retro_buckets bucket 3, “Late entry into a priced-in move,” with 11 trades and -$69.80. Mechanical timing attribution remains negative at -$30.25, and S4 transit alone cost $93.58 (attribution.decomposition.totals.timing_pnl; gauges.traded.s4_overnight.transit_usd_sum). It is therefore still recurring, though this week it is less costly than directional selection and does not receive a separate intervention.

The previously promoted “short run over by a violent adverse rally” tail pattern stopped appearing again. No corresponding bucket exists in retro_buckets this week, and the taxonomy’s 90 trades are fully allocated without such a category (retro_buckets coverage check). This is now a third calm week after its 2026-08-10 promotion (prior_weekly_reviews weeks 2026-08-17 and 2026-08-24). Because the proposed event gate was never shipped, the absence means dormant, not fixed; reopen it only if the tail recurs.

5. EXACTLY ONE BIGGER BET

The prior publish-to-detect gate is not re-issued a fourth time. Its repeated non-execution is a governance defect, and continuing to restate it would not produce evidence. This week’s sole intervention instead targets the currently weakest mechanical link: S4 selection.

2026-09-07 owner: place S4 behind a qty-0 capital gate for the full week while preserving its candidate ranking, freshness/alignment gates, hypothetical close fill, and next-session-close outcome in a mechanically reported shadow ledger — expect: signal, deployed S4 entries fall from 30 to 0 while at least 20 shadow S4 selections are graded; shadow directional hit rate remains ≤40% and shadow mean_available_pct remains ≤-0.20% versus this week’s 33.3% win rate and -0.42% mean_available baseline, confirming that the current S4 selector has no measurable short-horizon directional edge; a hit rate ≥55% with mean_available_pct >0 would refute that diagnosis and justify investigating monetization before restoring capital (scoreboard.per_strategy S4.closed/win_rate_pct; gauges.traded.s4_overnight.n/mean_available_pct; attribution.decomposition.per_strategy.S4.selection_pnl).