…failures

Publisher could not resolve a slug — the planner embeds schedule_slot nested
in post_outline (a JSON string under output_key), but Publisher read top-level
state['schedule_slot'] (never set) and outline['slug'] (wrong nesting), so
slug='' and publish_to_firestore('') returned error while /generate still
returned 200 — blog_posts stayed empty.

- Parse the writer's YAML front matter (authoritative: slug/title/description/
  keywords/cta; schedule row lacks description) as primary metadata source;
  strip it from the stored markdown body so posts render cleanly.
- Loud ERROR log on unresolved slug; /generate now 500s on error/rejected so
  failures surface and Cloud Scheduler retries.

Verified: real /generate published blog_posts/why-uoa-is-mostly-noise.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Front-matter parsing is the primary metadata source but the LLM occasionally
emits a body whose front-matter block isn't at position 0, so it fails to
parse; combined with the nested-in-outline schedule_slot (not in top-level
state), slug resolved to '' and publish 500'd. Add a last-resort read from
tools.fetch_next_schedule_slot() — in the cron drain the first pending row is
the current post; manual retries resolve slug from state['slug'] earlier and
never reach this path. Uses the existing plain tool fn (no ADK ToolContext).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… score-led ranking

Three owner-directed signal-quality changes to the STRICT selection path,
driven by the first leak-free realized-option-PnL backtest (1,375 fills;
backfilled full 3-day option minute bars + replayed the exact +80/-60/trail
bracket). No trader-mechanics change; every tradeability gate retained.

- Remove `V/OI > 2` conviction gate. Realized PnL: dropped ~55-63% of real
  option winners for precision lift statistically <= 0 (90% CI [-0.061,-0.001]),
  not fillability-confounded, stable across chronological halves. Folklore gate
  that was the main cause of picker-slate starvation (~2 candidates/day).
- Re-rank STRICT ORDER BY from directional-V/OI-DESC to overnight_score-led
  (overnight_score, recommended_oi, spread, ticker) -- now identical to FALLBACK.
  V/OI is a poor filter and a poor ranking key. Supersedes the 2026-05-01 primary.
- Widen moneyness cap 0.10 -> 0.13 (STRICT only; FALLBACK_MONEYNESS_MAX decoupled
  and pinned at 0.10 -- prior `= MONEYNESS_MAX` would have leaked deep-OTM onto
  low-conviction skip days). Realized PnL: 10-13% increment +8.9% (CI
  [+.014,+.163]); the (0.14,0.15] bin is toxic so cap stops at 0.13. Mechanism
  correction, not a literature reversal: the deep-OTM-cliff lit is hold-to-expiry,
  our 3-day bracket is not that trade. Thin/single-regime -- reversible, monitored.

All three cleared gammarips-review. Decision docs:
docs/DECISIONS/2026-06-02-voi-gate-relaxation-proposal.md,
docs/DECISIONS/2026-06-02-moneyness-cap-widen-to-13.md.

Research (backtesting_and_research/): exit-design study FALSIFIED the deep-research
"drop the -60% premium stop" claim (removing it = zero EV change, just fatter
tails -- the wick-out is a hold-to-expiry artifact); sweep/ISO detection (H20)
parked, blocked on Polygon data tier. Brief: H16-H21. Docs synced (CLAUDE.md,
TRADING-STRATEGY.md, CHEAT-SHEET.md, INTELLIGENCE_BRIEF.md, NEXT_SESSION_PROMPT.md).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Captures the 2026-06-03 sector-persistence + webapp internal-linking work:
enrichment now writes sector/industry to overnight_signals_enriched + the
Firestore signal doc; the webapp ranks same-sector related signals. The columns
are NULLABLE and NON-GATING (read by no gate/WHERE/ranking).

- DATA-CONTRACTS.md: sector/industry metadata columns entry
- DECISIONS/2026-06-03-sector-persistence-and-webapp-internal-linking.md (new)

The enrichment code hunks for this ship in the following commit (same file as
the FRED work, unavoidably entangled in one working tree).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…e gate

Captures what shipped live 2026-06-03 (enrichment rev 00039, signal-notifier
rev 00031) after a FRED VXVCLS/VIXCLS outage NULLed VIX3M for an entire
scan_date and wiped the signal slate (no trade that day).

- enrichment-trigger: _fetch_fred_csv retry (3x/30s, backoff) for VIX3M + VIX;
  on VIX3M failure, bounded carry-forward of the last non-null vix3m_at_enrich
  from BigQuery (<=7 calendar days, strictly < scan_date, else fail-closed).
  Bundled here with the parallel sector/industry persistence hunks (same file;
  NON-GATING).
- signal-notifier: live-VIX fallback FRED -> Stooq -> Yahoo with two-source
  corroboration (1.5 vol-pt agreement), plausibility bound, d<today guard, and
  vix_source provenance on todays_pick.
- docs: TRADING-STRATEGY regime-gate note; DECISIONS/2026-06-03-vix3m-fred-
  retry-and-carry-forward.md.
- research: premium_affordability_study.py (EV-neutral premium-cap finding).
- .gitignore: ignore the regenerable realized_label.pkl artifact.

gammarips-review: GO on both change sets (no lookahead — carry-forward reads
strictly < scan_date; sector is gating-inert; ALTER is idempotent).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Pre-existing SEO tooling, committed to clean the working tree. Read-only
analytics helpers (Google Search Console + GA4) under scripts/seo/ and the
gammarips-seo subagent definition. No secrets; .venv and __pycache__ ignored.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ades

Publishes the per-trade V5.4 cohort to Firestore ledger_trades/{scan_date}_{ticker}
for the public scorecard ledger table. Same cohort filter + fixed-dollar sizing
as cohort_stats/current, so the table rows and aggregate tiles can never disagree.

- compute_and_write_ledger_trades(): reads closed V5.4 forward_paper_ledger rows
  (V5_4_AGENT_RANKER, entry >= LIVE_COHORT_START_DATE, realized), parses the OCC
  contract (_parse_occ_contract) into option_type/strike/expiration, writes one
  idempotent Firestore doc per trade. Non-fatal; never raises into the email path.
- Called alongside compute_and_write_cohort_stats in run_notifier + /refresh_stats.
- DATA-CONTRACTS.md: ledger_trades collection schema.

Live (signal-notifier rev 00032); /refresh_stats backfilled all 6 closed trades.
gammarips-review: SAFE TO DEPLOY (read-only publish; cohort/sizing parity verified;
no gate/trader impact).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Inject a curated, leakage-safe forensic case-memory into the V5.4 Picker so it
reasons by analogy over past option winners/losers. Direct fenced-block
injection (the house pattern), NOT ADK MemoryService (session-recall RAG).

- build_case_memory.py (read-only): join realized_label.pkl (FILLED option
  outcome + underlying path) with overnight_signals_enriched (greeks/IV/catalyst)
  on (recommended_contract, scan_date); overlay 6 matched live ledger closes.
  Emits case_memory/{bull.md 846, bear.md 529, exemplars.md ~50 curated,
  build_manifest.json, case_index.parquet}.
- Outcome keyed on realized_ret>0 (option PnL) NOT is_win (stock direction) —
  they disagree 44.2%. WHY is deterministic option physics (theta drag / delta
  capture / inferred IV residual); no LLM-authored cause.
- quant.md: 12 hand-authored ledger-independent priors (Q1-Q12).
- Wire: {case_memory_block} in _build_picker_instruction; picker_v5.md adds a
  "how to use case memory" section; Dockerfile ships case_memory/;
  PICKER_PROMPT_VERSION=5.
- Fail CLOSED if v5 ships without the block (no silent v4 degrade);
  RankResponse.case_memory_bytes for observability; deploy.sh preflight assert.

Audited by gammarips-review (SHIP-WITH-FIXES, all fixed). Same-ticker
outcome-import documented as accepted+bounded. Advisory/non-gating; owner
waived N>=15/30-day-OOS ceremony, leakage was not waived.

Decision: docs/DECISIONS/2026-06-03-picker-case-memory.md

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The 2026-06-03 fix required TWO public sources (Stooq+Yahoo) to corroborate
within 1.5 vol-pts before trusting a live VIX when FRED is down. FRED was still
timing out on 2026-06-04; only Yahoo answered (16.06) so the rule rejected it
and the regime gate fail-closed — wiping scan 2026-06-03 (4 candidates, vix3m
carry-forward 18.66, plainly contango).

Drop the corroboration requirement: use the best public source that answers;
when both answer take the MAX (conservative for the one-sided vix_now>vix3m=>skip
gate — a low-biased source can't manufacture a false trade). Single source is
sufficient; only a total blackout fail-closes. Remove VIX_FALLBACK_TOLERANCE.

Lookahead-safe: every fallback bar still passes _vix_date_ok (d<=scan_date AND
d<today). gammarips-review: GO. Deployed signal-notifier-00033-sjr; re-ran scan
2026-06-03 -> BBWI BULLISH emailed before the 10:00 ET entry.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…e fix

The 2026-06-02..04 "FRED outage" was self-inflicted. Our fredgraph.csv requests
carried no start date, so FRED serialized each series' full history back to 1990
on every call — a payload that now exceeds the read timeout every morning. The
retries couldn't help (each re-requests the same giant dump) and the carry-forward
/ two-source / Yahoo fallbacks only fired because the primary always timed out.

Proven by live probe: bare ?id=VIXCLS times out at 30s deterministically;
?id=VIXCLS&cosd=<date> returns HTTP 200 in ~2s. Same for VXVCLS.

Bound every FRED CSV fetch with cosd (start date) = scan_date/target_date − N days:
- enrichment-trigger: _fred_csv_url() helper + FRED_CSV_LOOKBACK_DAYS=45 (VXVCLS
  gate + VIXCLS context); unparseable-date falls back to unbounded URL.
- signal-notifier: cosd inlined in fetch_vix_close (live VIX leg, 45d).
- forward-paper-trader: _fetch_vix_daily_fred(target_date) bounded 60d, 15s→30s
  timeout (telemetry-only, non-blocking).

Parse helpers unchanged — still take the latest close on/before scan_date, so cosd
only moves the window START earlier (lookahead-safe). gammarips-review: GO on all
three services. Retry/carry-forward/source-fallback retained as defense-in-depth.

Revs: enrichment-trigger-00041-trm, signal-notifier-00034-ds9,
forward-paper-trader-00037-t65.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…e (judge_v6)

Replace the two-stage V5.4 ranker (per-candidate Scorer fanout + Picker) with a
single memory-aware judge: one structured gemini-3.1-pro-preview call scores every
gate-cleared candidate AND selects the pick. Rename service+dir+code signal-ranker
→ signal-judge (BQ table signal_ranker_runs + Firestore v5_4_* keys intentionally
kept — migration/webapp landmines, no payoff).

Why: post-gate slates are mostly ≤5 candidates, so the Scorer's top-5 cut was a
no-op ~80% of days; ITM/earnings/spread rules were triple-encoded (gates + scorer +
picker). A 13-slate replay (workflow) showed the single judge agreed 9/13 with the
2-stage baseline and was structurally sounder 4-to-1 on divergences (each was the
judge rejecting a two-label-trap the 2-stage took). gammarips-review: SHIP. Live
gemini smoke on the 2026-06-03 slate reproduced the BBWI pick with anti-anchoring.

- judge_v6.md: trusts upstream gates, anti-anchoring ("score as if only candidate"),
  per-candidate verdict array, mass-leakage skip, deterministic composite/tiebreak.
- JudgeOutput/PerCandidateVerdict schemas; ScorerOutput/PickerOutput kept for replay.
- run_judge: leakage-assert all candidates → ONE structured call → JUDGE_MAX_ATTEMPTS=3
  bounded retry (replaces the lost MIN_SCORER_SUCCESS_FRAC partial-failure tolerance).
- persist_run: one row per verdict; judge mirrored into both scorer/picker REQUIRED
  columns at version=6 → BQ DDL unchanged, cohort cleanly separable.
- Wire contract preserved → signal-notifier needs no logic change (only repointed
  SIGNAL_RANKER_URL → SIGNAL_JUDGE_URL at the new service).
- Docs: TRADING-STRATEGY, CLAUDE.md + GEMINI.MD, CHEAT-SHEET, decision note.

Deployed: signal-judge-00001-4kn (DRY_RUN=false), signal-notifier-00035-bvh repointed.
Old signal-ranker kept live as rollback. 32 unit tests pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

V/OI>2 was removed as a gate 2026-06-02 (no realized-PnL selection value).
Update the gate list the blog writer cites + spread 10%→8% so future posts
don't describe V/OI as a gate.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ame + V/OI removal

Durable-handoff pass over the authoritative/living docs (historical DECISIONS/
EXEC-PLANS left as the record):

- MODELS.md: registry now shows gemini-3.1-pro-preview as the single signal-judge
  judge_v6 (scores + selects in one call); gemini-3.5-flash no longer powers any
  judge stage (Scorer removed). Env-swap table + cohort-attribution note updated
  (segment by scorer_prompt_version: 5=two-stage, 6=judge_v6).
- ARCHITECTURE.md: added signal-judge + signal-notifier service entries (the map
  previously jumped enrichment → trader); fixed the data-flow gate stack.
- DATA-CONTRACTS.md: policy-contract gates corrected (spread 0.08, moneyness
  5-13%, V/OI gate removed) + judge/rename note; moneyness field 5-15% → 5-13%.
- GLOSSARY.md: added signal-judge row; signal-notifier row de-stale'd.
- TRADING-STRATEGY.md / CHEAT-SHEET.md: residual Scorer/Picker/signal-ranker
  mentions → single judge_v6 / signal-judge; SLO + fallback + skip-reason copy.
- TESTING.md: enrichment spread 0.10 → 0.08.

Kept deliberately: signal_ranker_runs table name, Firestore v5_4_* keys, the
V5_4_AGENT_RANKER policy_version label. Not touched: the broader Agent-Arena
"5-model debate" narrative (separate webapp/marketing pass).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ignals

signal-judge now runs a randomized bracket tournament (tournament_v1, version 7)
instead of the single judge_v6 call, and signal-notifier ungates the candidate
pool so every enriched signal gets a chance.

Why: the gated single judge over-fit junk (picked score-2 BBWI over $100M-flow
names) and the selection/liquidity gates were choking real winners on stale
scan-time OI (the sweep only becomes OI the next morning; we enter at 10:00 and
ride the build). Features barely separate winners from losers (EV ~flat), so a
rigid scorer can't extract edge that isn't there — a simple prompt over the full
pool at least picks sensible, report-aligned trades.

signal-judge (app/agent.py): 3 independent brackets; each shuffles the pool,
batches of <=10, top-2 advance (top-1 in the final batch) until one remains
(94->20->4->1); consensus across the 3 = pick (3/3 high, 2/3 medium, 1/3 low).
Dead-simple prompt + daily report + per-contract JSON. No memory, no rubric.
assert_no_leakage on every candidate; fail-closed on no winner.

signal-notifier (main.py): candidate query ungated (moneyness/OI/vol/DTE/V-OI
removed, LIMIT 200) + rich feature columns added (no outcome cols); active-days
liquidity gate BYPASSED (also drops ~N Polygon calls). KEPT: no-earnings-in-hold
and regime fail-closed (safety, not selection). /rank + service timeouts bumped
for the longer tournament.

persist: tournament has no rubric scores -> advancement proxy (round reached ->
1-10) into the REQUIRED columns, version 7, finalists+winner+runner_up only. BQ
DDL unchanged; cohort separable (5=two-stage, 6=judge_v6, 7=tournament).

gammarips-review (leakage): SHIP — no forward column can reach the model (60-col
allowlist + assert_no_leakage catches all 10 outcome cols). Deployed signal-judge
+ signal-notifier-00036-c6t; live /rank on 94 candidates verified (MSFT, medium,
version=7 rows persisted). See docs/DECISIONS/2026-06-04-bracket-tournament.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

V5.4 (gated single-judge) was a dud — 13 live closes, avg 0.0%. Owner-directed
clean break: wipe the ledger and start the tournament cohort fresh as V6.

- forward_paper_ledger TRUNCATED (13 V5.4 rows; dumped to .scratch as insurance).
- policy_version V5_4_AGENT_RANKER -> V6_TOURNAMENT across all WRITE sites
  (forward-paper-trader POLICY_VERSION constant; signal-notifier todays_pick +
  ledger_trades writes) and all READ-FILTERS (signal-notifier cohort_stats + 14d
  summary, win-tracker, x-poster, blog-generator) so the app surfaces V6 not blank.
- LIVE_COHORT_START_DATE -> 2026-06-04; cohort_stats/current refreshed -> V6, 0 trades.
- webapp cohort-stats default -> V6 (committed separately in gammarips-webapp).

Deployed: forward-paper-trader-00038-fd5, signal-notifier-00038-7wx.
win-tracker/x-poster/blog code updated but NOT redeployed (harmless — V5.4 gone,
no V6 closes yet; deploy before first V6 close surfaces).

CLAUDE.md policy summary rewritten for V6. See
docs/DECISIONS/2026-06-04-bracket-tournament.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

_best_contract scored contracts with a `vol/max(oi,1)` term that REWARDED low
open interest as "unusual" — so it picked the swept lottery strike (OKTA $127
6/12: OI 5, ~35% live spread, untradeable) over the standard liquid strike (OKTA
$130: OI 48, fillable). Unusual flow is the right signal for the NAME+direction;
using it to pick the CONTRACT hands you the strike you can't fill.

Rewrite the score: open interest PRIMARY (5x, accumulates, can't be faked by one
sweep), volume secondary (2x), snapshot spread tertiary (1.5x — it read 0.5% on
that 35%-wide OKTA strike, so it's noisy), keep sweet-spot delta + gamma (de-
emphasized 20x->8x) + theta penalty. Flips OKTA to $130.

No hard OI floor (scan OI is stale; a fresh sweep that builds reads ~0 — don't
reject those). Live fillability check at pick time is the next, separate layer.

Deployed overnight-scanner-00010-jk2; takes effect next overnight scan.
See docs/DECISIONS/2026-06-04-contract-selection-liquidity.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Adversarial multi-agent audit (every finding re-verified vs code+BQ data) found
16 silent bugs. 13 fixed; 3 deferred (need PIT data/schema). Trigger: OKTA $127
untradeable-ghost pick.

CRITICAL #1 (root cause): polygon_client stopped substituting day low/high for a
missing bid/ask — that produced fake/0% spreads on ~43% of picks (718/1815 = 0.0),
defeating the spread gate and feeding the judge "0.5% spread" ghosts. Missing quote
-> NULL spread; chosen contract now carries a real spread. Enrichment spread gate
loosened 0.08->0.30 (was filtering fake 0s; real spreads are wider).

HIGH #2: divergence flip resolves BEFORE conviction sub-scores (was scoring flipped
names on the abandoned side -> ~87% suppressed below MIN_SCORE). #5: judge prompt
strips still-stale volume/OI/V-OI (keeps now-real spread). #8: technicals window
bounded to scan_date not date.today() (was leaking next-day bars = lookahead).
#6/#7/#10: signal-notifier fallback moneyness band restored + dead gates/docs cleaned.

MED #9/#12/#13: trader symmetric slippage + stale-TIMEOUT + late-fill guards (removed
EV-optimism). #11: empty judge batch re-queues instead of dropping 10 candidates.
#14: in_top_5 is a real top-5. LOW #15: overnight_score fail-safe default. #16:
greeks store raw with None preserved (delta no longer NULL->0.0).

Deferred (separate pass): #3 OI frozen-snapshot, #4 volume frozen-cumulative (need
Polygon flat files / day-bars), #15-full (schema add). See
docs/DECISIONS/2026-06-04-pipeline-bug-fixes.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Make the scaffold tell the truth — no V5.4/judge_v6/Scorer-Picker/case-memory as
the live description anywhere; historical DECISIONS/EXEC-PLANS left as the record.

- CLAUDE.md + .gemini/GEMINI.MD (lockstep): policy summary = V6 tournament; spread
  gate 8%->0.30; blog-generator marked DEPLOYED; bug-hunt summary added.
- CHEAT-SHEET.md: retitled "V6 Tournament"; gate list + pick section rewritten.
- docs/TRADING-STRATEGY.md: Status/regime/filter-stack/policy_version -> V6.
- docs/MODELS.md: registry -> tournament_v1 judge; no Scorer/Picker; cohort 5/6/7.
- docs/ARCHITECTURE.md + GLOSSARY.md: signal-judge/notifier/trader + spread 0.30.
- docs/DATA-CONTRACTS.md: V6 policy contract + field-quality caveats + new cols.
- NEXT_SESSION_PROMPT.md: bug-hunt handoff + the has_contract WATCH.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…he-data lesson

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The blog-generator's BQ filters were already V6_TOURNAMENT, but its
prompt/forbidden-list NARRATIVE was two eras stale — it would re-draft
V5.4 / Agent Arena / "gate stack" / "5-13% OTM moneyness" / "9:00 AM"
content on the next weekly cron (Mon 05:00 ET) and on any /generate regen.

- agent.py (writer forbidden-list): current floor is score>=4, alert is
  7:30 AM ET; add Agent Arena / Scorer / Picker / gate stack / 5-13% OTM
  moneyness to the retired list.
- fast_api_app.py: newsletter retired-concepts list (add V5.4 / Arena /
  Scorer / Picker / gate stack / moneyness); reddit lead_style + algotrading
  mod_traps drop "gate stack" → tournament/bracket; reddit topic list rewritten
  to the V6 enrichment bar + 2 safety rails + "no selection gates" (was the
  full V5.4 moneyness/spread/V-OI gate list).
- tools.py / voice_rules.py: stale V5.4/V3-V4 comments + the X-poster
  "V5.4 cohort >= 30 closes" do-not rule → V6.

Shared voice_rules.py is vendored into x-poster too, so this also de-stales
X copy. Not deployed — engine code only.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Alert time moved to 07:30 ET (was 9:00 AM). Update the blog_schedule slot
title_candidate + keyword so /generate re-drafts an honest title; the live
blog_schedule/current row was updated to match. The /9am slug stays (URL
artifact only). Post regenerated clean on V6 (reviewer 10.0).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>