Aug 12, 2026

Regime & Verdict#

• Regime: UPTREND · ret5 +0.350% · ret10 +5.900% · dd15 +0.000% · VIX 14.55 (LOW tercile, contango) · breadth 51.9% green vs 33.9% bullish flow (~18pt divergence) · directional_tradable=true · s1_standdown=false
• Bottom line: no directional edge tonight. Book stays FLAT for a third session. Two OI_FADE names (COO, NRG) cleared the full mechanical stack and are held at WATCH — each overdetermined by independently-sufficient gates. Vol book stands aside on 0DTE.

The label is right; the story behind it is not what it looks like#

The mechanical read is UPTREND and it is correct. The qualitative overlay is the opposite of a breakout, and tonight a Phase-A read got this backwards in a way worth recording:

• dd15 = 0.000% does not mean "at the highs." Per scripts/_regime.py:13 it is min(last 15 closes) / close_10_sessions_back - 1 — explicitly not a peak-to-trough drawdown, with a docstring warning that a peak-based re-derivation "silently changes which arm of the crash guard fires." It reads 0 here because the window low (729.46, 07-29) is the 10-session anchor. So ret10 +5.9% is a trough-anchored rebound measurement, not a continuation measurement.
• Today is not a fresh high. 772.49 is the 3rd-highest of the last 15 closes, −0.10% under the 08-07 max (773.26). Not a fresh intraday high either (774.87 vs 776.85 on 08-05).
• The CPI print was faded. Opened 774.71 (the day's high), closed 772.49 at 34% of the day's range, −0.29% from the open. Up +0.25% vs the prior close, but the gap was given back.
• Four-session stall: 773.26 → 773.03 → 770.56 → 772.49.

So: the V-rebound off the 07-29 trough completed by 08-07 and the tape has gone nowhere for four sessions, fading a benign macro print, on ~18pt price/flow breadth divergence, defensive rotation, SPY dealer short gamma (spot below the 779.01 flip), with VIX at the bottom of its 90-day range. Treated as regime-transitional, same as 08-11. The P0.6 out-of-regime half-cap binds — non-binding in practice only because nothing reached a sizing band.

The label will flip on its own inside the horizon#

ret10 is anchored on the 07-29 trough. As that rolls out of the window, flat price alone re-labels the tape:

Anything entered tonight at h10 is scored under a regime that will not survive its own holding period.

Directional Book (excess-scored)#

EMPTY — zero new starters, zero open positions carried. Prior book confirmed genuinely flat (conviction_2026-08-11.json records a positive zero-count, not the held-book EXIT-substring bug).

Why the lane prior is discounted tonight — a standing methodological gap#

The live lane ranks candidates by rel_build (build relative to the name's own base). The harness that produced the +0.59% mean / n=1087 baseline selects top-15 by raw oi_net_5d with only liquidity/earnings/ETP gates — no persistence, catalyst_split, or rel_build filter at all (retro_harness.py:82-99).

All five candidates pass every harness gate. But on tonight's raw-ranked list of 1,359 gated names they sit at:

The top-15 cutoff tonight is 19,000 (TTD). (The lane's 5-day sum and features' oi_net_5d are the same metric exactly 5× apart — mean vs sum — so ranking is preserved. This is a selection-rule difference, not a data bug.) The harness's own top-15 is dominated by mega-cap and crypto-beta names (SPCX, MSTR, PCG, DKNG, META, AAPL, MARA, TSLA, NFLX, AMZN…) — the "raw = mega-cap beta short" artifact the relative screen exists to avoid.

Consequence: the +0.59% prior is being borrowed for a differently-ranked population, not measured for this selection rule. NRG (rank 29) retains most of the claim; COO (rank 192, ~13× outside the measured population) is effectively unvalidated for this selection rule — sufficient on its own to cap it at WATCH. This is a standing gap between the live lane and the instrument that grades it, not a tonight-only issue.

Candidate detail#

• COO $76.73 · $-ADV $149.3M · Common Stock · Healthcare/Medical Instruments. 5d net call-OI build +5,497, rel_build 1.115, persistence 0.733 (organic; positive net three straight sessions — the distributed shape is genuine). Earnings gate PASS (next print 09-09). catalyst_split NO_CATALYST_IN_WINDOW — the 08-26 date was a stale preliminary estimate; the current field has been stable at 09-09 since 08-05. 3rd consecutive WATCH night (08-10, 08-11, 08-12).
  • Invalidation corrected tonight. Last night carried "~$80.50." Verified against OHLC: COO's max high since 06-01 is $77.42; the $84 area is from February 2026, six months stale. An $80.50 stop sat ~4.9% above spot with no anchor in recent structure. Operative level is the confirmed $77.42 (08-11 high), ~0.9% above spot.
  • Fundamentals CAUTION (−1), flagged ELEVATED / borderline-VETO. 2 of 3 mechanical legs contradict the short with hard first-party evidence: a 4-for-4 EPS beat streak and insider net buying (MSPR +40.33, real dollar purchases). Only growth confirms (−42.61% YoY EPS on 63× P/E). D/E 0.3041 — genuinely low, and here the leverage_flag is correct. The UBS Neutral $75 PT carried from 08-10/08-11 was not re-verified tonight; treat as unconfirmed.
• NRG $120.66 · $-ADV $379.1M · Common Stock · Utilities/IPP. 5d net +49,909, rel_build 0.604 (down from 0.835 — mechanical, the large 08-05 session rolled out of the trailing window; not an unwind), persistence 0.419 — the cleanest, most-distributed build tonight. Earnings gate PASS (next print 11-05). catalyst_split: 08-04 print confirmed; shipped >= 74.2% vs strict > 81.3% — the strict boundary raises the post-catalyst share, so the known print-day defect strengthens this read (the opposite of EWTX). News tape tonight: "Red-Hot Power Stocks Are Losing Steam" — sector rotation out, supports the fade. Invalidation $124.71 confirmed (08-05 swing high), ~3.4% above spot. 2nd WATCH night.
  • Fundamentals CAUTION (−1), but a structurally weaker contradiction than COO's. Insider selling (MSPR −60.91) and a two-quarter miss trend both confirm the short; only growth contradicts. The loud-looking 62–75% analyst-target gap is sell-side opinion whose targets were just cut post-gap (Scotiabank $226→$211, Evercore $215→$195) — bulls marking down conviction, against insiders selling into the same catalyst stack. D/E 9.78 is genuinely HIGH — the finnhub_enrich leverage_flag "low" is the known bug reproducing.
  • This inverts last night's asymmetry note, which read NRG's CAUTION as the louder of the two. On first-party evidence COO is the more dangerous short; NRG's contradiction is a single soft leg amplified by a sell-side narrative that is itself losing conviction.

Killed upstream#

• USFD (persistence 1.053), NXPI (1.022), UTHR (1.057) — all exceed the 0.85 ceiling that oi_build.py:145 itself flags as "likely a single-day block, not organic building." The lane had ranked USFD #1 and NXPI #2 by rel_build; demoted. Name-specific notes: USFD's price has risen every session since its print (crowd validated, not faded, no confirmed resistance above spot); NXPI's single 08-07 block is of unconfirmed origin (possibly an institutional roll); UTHR is already −5.0% off its 08-07 post-print high, so much of the fade may be realized.
• EWTX — CUT. Repeat of the 2026-08-10 boundary artifact: shipped >= reads 85.3% LIVE, strict > reads 15.6% — the build is almost entirely the print day. Do not re-propose.

Vol Book (non-directional)#

• 0DTE-VRP — STAND ASIDE, both SPY and QQQ. Two independent vetoes.
  • The LOW-VIX conditional net is negative once cost is charged: SPY −0.095%/day, QQQ −0.037%/day (gross 0.005% / 0.063% less 0.10% round-trip). The pooled headline (+0.13% / +0.25%) is not the governing number — VIX 14.55 puts us squarely in the LOW tercile, which is exactly the pooling trap.
  • Gamma gate vetoes independently: SPY dealers short gamma (spot 772.82 < flip 779.01), a trend-acceleration regime this sleeve's unsampled left tail cannot survive. QQQ sits at its flip (724.04 vs 723.64) with the CLI and the local flip level disagreeing — regime not reliably known, so fail closed.
• Earnings IV-crush — 2 structural qualifiers, neither sized.
  • TPR — IV-rank 87.7, front IV 165.6% (2DTE) vs back-month 52.9%, backwardation 3.13. Defined-risk iron fly, wings set beyond the ~12.2% raw-IV move rather than the 8.4% headline (TPR gaps hard).
  • LOW — IV-rank 83.2, backwardation 1.16 (weakest of the survivors), tightest execution. Size smaller.
  • QFIN / MRCY excluded — 33 and 41 front-expiry contracts, spreads would eat the crush edge.
  • No per-name net-of-cost backtest exists for this lane. Structural qualification only; the vol-short tail cap requires a net-of-cost quote, so neither sizes tonight. Net expectancy must be verified against the live iron-fly credit at entry.

Watch / Stood-down#

• MOM_SHORT — BASKET_WATCH, 10 names: APP, ACM, RBA, GME, ROL, AMRZ, VICI, NKE, WING, STLA. All Yahoo-verified as genuine near-52w-lows (the pct_52w_range artifact breaks both ways and the in-range class manufactures false shorts), all liquidity/earnings/ETP gated. s1_standdown=false, so the crash guard is not what caps this — invariant #6 does (baseline −0.0128, n=803, knowingly negative). GPI and HDB excluded on prior blocking verdicts.
• MOM_LONG — BASKET_WATCH, 50 names (top-10 spot-verified). Baseline −0.0108, n=1066. DURABLE-N bar unmet; basket-only, never sizes.
• S2 (advisory, +0.09%, n=1123) — 15 names fully news-verified, out of 80 that cleared liquidity/earnings: WFC, WDAY, FICO, SU, APO, CG, DBX, FISV, HBM, LBRDK, RY, MAR, TOST, NSC, MDLZ. AVB VETOED (AVB-EQR merger shareholder vote same day) and BAH VETOED (Treasury contract cancellation, −1 day). NTRA/RBLX/ROKU flagged. The other 60 names stay excluded, fail-closed — last night this lane shipped 80 names with an incomplete news gate; tonight it was cut to what could actually be verified.
• S4 (advisory, +0.35%, n=1145) — 27 names post-gate, led by PPL, EFX, JXN, IMAX, COKE, LEN, IRM. Separate orthogonal ivrank_chg_5d h3 tilt (10 names: BSP, TBIL, ASO, WKC, VOD, DVY, NI, CHWY, KR, EWH) — reported alongside, never summed with the PCR signal.

Regression gate (retro_harness.py --all, run tonight)#

No lane or threshold changed tonight — but the truth set was rebuilt and extended (83 → 86 panel days), so the gate was run to see where the baselines now sit.

Verdict: NOT a regression. Four of five lanes sit marginally below their recorded baseline, but zero code changed tonight — the only input that moved is the panel, which grew three days with n rising across every lane. That is the textbook data-drift case the gate explicitly does not treat as a failure. The gate exists to catch code changes that degrade a lane; there was none to catch.

But OI_FADE's softening is material to tonight's call. −34% relative (+0.59% → +0.39%) on 45 new observations is the largest move of any lane. The lane's structure is intact — hit-minus-base is essentially unchanged (+0.172 vs the recorded +0.17) and the median stays positive (+0.0081 vs +0.0095) — so this reads as the mean softening, not the edge inverting. The consequence compounds with the selection-rule finding above: the prior cited for COO and NRG was already being borrowed from a differently-ranked population, and now the prior itself measures ~a third lower on its own basis. Citing "+0.59%" tonight was generous. This strengthens, not weakens, the decision to hold both at WATCH.

Risk#

• Event stack inside the h10 window (08-13 → 08-27): PPI 08-13, Retail Sales 08-14, FOMC minutes 08-19. CPI cleared today, but two Tier-1 prints land within two sessions — independently sufficient for −1 tier on both OI_FADE names, exactly as on 08-11.
• Correlation: COO/NRG verified as distinct singletons — measured pairwise correlation 0.127 (n=141 paired days) against the 0.70 threshold, computed rather than inferred from sector labels. Both had full price history, so this is a real pass, not the absentee-reads-as-pass trap.
• Tail: dealer short gamma + bottom-tercile VIX is a larger-realized-move environment, which raises the cost of being wrong on either short. Noted as a risk, not invented as a sizing gate — OI_FADE has no named tail cap in the ruleset.
• Half-cap, no auto-full. Nothing sized tonight at any level; neither name would have exceeded starter even absent every downgrade.

Orchestrator notes / action items#

1. Phase-A dd15 misread — recorded so it is not repeated. The regime agent read dd15 = 0.000% as "today is the 15-day high" and built a "fundamentally-anchored breakout, genuine trend-continuation" conclusion on it, twice asserting a fresh high. OHLC says 3rd-highest close, −0.10% below the 08-07 max. This is the exact failure the _regime.py docstring warns about. The corrected read (trough-anchored rebound, stalled, faded the print) points the opposite way and matches the agent's own breadth/rotation caveats. Any figure a lane hands back that carries a directional conclusion needs arbitration against the source definition, not just against regime_check.py's numbers — the numbers here were right; the interpretation was not, and --claim-label/--claim-ret10 would not have caught it.
2. Lane/harness selection-rule mismatch (standing). Live lane ranks by rel_build; the harness baseline is top-15 by raw oi_net_5d. Candidates at raw rank 29–192 of 1,359 are borrowing a prior measured on a different population. This is the ranking-metric sibling of the known "patch the lane, not the harness" defect. Worth a maintenance decision: either add the live lane's persistence/catalyst_split/rel_build screens to the harness and re-baseline, or record explicitly that the live lane's prior is inherited rather than measured. Any change must clear retro_harness.py --all first.
3. COO invalidation was stale, not merely loose. "~$80.50" had no anchor in six months of price. Carried references should be re-verified against OHLC before reuse — the corrected $77.42 sits ~0.9% above spot, which is a materially different trade.
4. finnhub_enrich.py leverage_flag bug reproduced again on NRG (D/E 9.78 reported "low"); it read correctly on COO (D/E 0.30). So it is not uniformly broken — it mislabels the high end specifically. Still needs a fix; until then override by hand on any high-D/E name.
5. fz CLI exposes no Recom/Target Price/Short Float fields — confirmed again by direct probe on both names, in both --agent and full --json modes. Analyst targets cited tonight came from Finnhub news headlines and are attributed as such.
6. Next /calibration-audit due 2026-08-15 per cadence. 41 open rows (17 OI_FADE, 12 MOM_SHORT, 11 S4) mature 08-10→08-21; the 08-03→08-07 crash-guard block matures 08-18→08-22 — the first cohort spanning non-overlapping windows. Count distinct exit-days, not rows.
7. Watch tomorrow: whether COO's build stays distributed (persistence has run 0.659 → 0.733) or tips into single-day-block territory, and whether NRG's rel_build keeps decaying mechanically as older sessions roll out.