EVERYTHING WE'VE SHIPPED.

All re-arm checklist items now live (trading stays PAUSED until owner re-arms): - #1416 news dampen 0.5x + #1285 flow down-weight 0.8x (both env-tunable; inverted/overweighted per lift data) - #1439 concentration gate: one position per underlying (parseOccUnderlying + hook 7b) — the 07-07 double-AAPL class closed - Spread-cost gate: entry skipped when live spread > 8% (ALPACA_MAX_ENTRY_SPREAD_PCT) — 43% round-trip tax measured vs ~5% modeled - #1398 refit part 1: cf>=3 excluded from tradeable (90d: 46% WR/−6.2% vs cf=2 57%/+10.5% n=391); TRADEABLE_MAX_CONFLUENCE env - (Thu) dead-zone ordering bug fix — gate re-validates FINAL edge after insider adj

Tests: 1316 Vercel + 208 price-stream. Both deployed; VPS restarted, auto-trade still DISABLED. Paper this week: +$23k (39W/12L) — heavily untradeable big-ticket; tradeable funnel currently ~0/day (strict floors + caps + null-bucket LEAPs — watch post-refit). Re-arm: owner decision at Thu check-in — needs 2-3 days of post-fix tradeable-cohort paper evidence + the funnel producing picks.

Three losses in three live trades since the \$500 deposit (SPY −\$46, AAPL 315C −\$75 dead-thesis, AAPL 312.5C −\$147 gap-through-stop at open). Equity \$1,186.14, lifetime trading P&L −\$1,313.86 (−52.6% on \$2,500). Each loss expressed a KNOWN unfixed defect: flat scorer (#1398 band), same-ticker concentration (#1439), 20-30% carry gray zone, and a measured 43% round-trip spread tax vs ~5% modeled (effective-spread telemetry, 2nd reading).

Owner call: pause live entries, use the week for calibration. ALPACA_AUTO_TRADE=false deployed + verified ("Alpaca auto-trade: DISABLED"). Paper full speed; tracking/settle/alarms unchanged; deposits stay. Re-arm checklist in CLAUDE.md. Calibration queue: #1398 → #1285 → #1416 → #1439 → spread-cost gate (+ #1281 gate build; #1446 Kalshi NO-side bug also queued).

Also: #1446 filed — Kalshi qualifying pick (MIL ML, gap 8.9%, WON) never fired; suspect NO-side ticker handling + ephemeral decision logs. Phone-notification diagnostic page sent (no delivery errors logged server-side; awaiting owner confirm on app-side).

• Deposit #6 landed: lifetime $2,500, equity $1,408.39 (CSD-verified). Found + fixed OwnerDeposit table at $500 (deposit-poll 30d lookback never backfilled Apr/May) — backfilled all 6 rows w/ alpaca: idempotency tags; expenses.json updated. Trading P&L −$1,091.61 (−43.7%). Cap → ~$352.
• Cicero 0DTE conviction filter shipped (paper, #1392): DAY_TRADE requires vol/OI≥3 + premium≥$25k per his Jul-1 read. His playbook mined from archives (10:30 gate ✓ aligned, GEX = no feed, event-defer + delta-stops queued). His signal publisher fixed their side — first JSON signals inbound → #1408.
• Issue-hygiene deep-dive (subagent, full board): 51→33 open. 19 closed w/ evidence; calibration cluster (#1305/#1306/#1286/#1337) consolidated into #1281 = the auto-inversion-gate epic. #1435 filed (backup alarm + tier-brief). Top queue: #1398 → #1285 → #1281 → #1416 (all "make the score honest") → #1367.
• Strategy/weights verification for owner: 6/7 scouts firing (EARNINGS_IV seasonal-watch #1434), 12/17 components alive, 5 zeros explained (2 known-dead pipes #1408, 1 retired, 2 conditional).

Wednesday: engine hunts at $352 cap · carry-guard live at 30% · $1k milestone ping expected.

#1363 CLOSED — the #1 loss mechanism has an active counter. Owner approved 30% per carry-guard-eval evidence. runCarryGuard (was shadow-only) force-closes positions ≥30% underwater at 3:45 ET via the proven limit-escalation exit path (CARRY_GUARD_CUT). Safety: OCC-matched plays only (matchPositionToPlay — never double-close, fail-open on unmatched), CLOSE_FAILED pages owner, JSONL log continues w/ mode field. 205 tests pass. Deployed + armed: CARRY_GUARD_LIVE=true, CARRY_GUARD_THRESHOLD_PCT=0.30. Revert = env flip.

• GH triage: 13 closed w/ evidence (#1363 arc complete, #1410 umbrella, #1365 docs-regen done 07-02, #1383 feed alive, #1390 already-implemented, #1417 host false-alarm, #1418/#1428-#1431 dupes → #1422/#1408, #1432 FIXED — cursor untracked/machine-local, #1433 dupe → #1402). 63→50 open.
• Owner DCA confirmed: +$500 this week → cap ~$352/contract on land.

• Scorecard live-pnl bug (owner-caught): weekly ping said "live +$2,287" — was sum(pnl) over ALL settled (paper carries pnl too). Truth: 1 trade/−$46. Fixed both call sites (played=true filter, "LIVE: N trade(s)" label), deployed.
• PC artifact backlog merged: 265 commits (pure artifacts — sweeps 07-01→04, loss-ledgers, lift-audit +42 rows, contexts). Half the "dead slots" were delivery failure, not production failure. #1423 closed. PC still owns: scheduler restart for slots dead since ~07-04 (#1422 fault A).
• #1422 doctor page storm FIXED (fault B): Discord escalation backoff 30m→1h→2h→4h→8h cap per (key,day), both channels; GH/ntfy legs unchanged. Behavioral-tested.
• carry-guard-eval.mjs BUILT + RUN (60d, n=303 carried): guard WINS at every threshold — +$15.1k @15%, +$15.9k @30% (6:1 save ratio). Live shadow log empty (live flat every 3:45 ET). Owner decision pending: enable live at 30% / 15% / keep SHADOW. Evidence on #1363.
• #1412 LEAP_SCORE re-fit SHIPPED (paper-only): affine compression 25+0.35×raw → 43-58 band (raw sum hit 90+ on a 45% WR book; 80-89 was anti-predictive). Gate+ranking unchanged. 27 tests.
• Sunday sweep logged (data/sweeps/2026-07-05.md): eq $908.39, 0DTE first data 1W/4L −$527 (n=5), sentinels caught all PC slot deaths (#1420-#1426).

Next: #1409 Kalshi expectancy gate · carry-guard owner call · watch live entries at $239 cap.

Owner: Discord floods bury #live-trades — phone is now the trade feed. Built + deployed + owner-confirmed all 3 priority tiers on-device (urgent/high/default test notifications received).

• Alpaca (VPS sendTradeAlert): all 9 event types mirror to ntfy Priority:high (entries, fills, partials, cancels, exits, 0DTE, hard stop, EOD sweep) with price + P&L.
• Kalshi (Vercel notifyOwner helper, new src/lib/notifications/ntfy.ts): order placement (weather/MLB/NBA scanners), fill confirmations (fill-reconcile), ALL played settlements (notifyKalshiSettlement = single interception), deposits (deposit-poll w/ new-caps context).
• New alert classes: EOD digest 16:15 ET (equity + overnight-carry list — #1 loss mechanism gets a nightly exposure line; silence gets a heartbeat) · loss-streak pause now PAGES (was ops-noise, same mis-tier as loss-cap) · tier-crossing milestones ($1k/$1.5k/…, day-deduped, crossedTiers tested) · weekly scorecard TL;DR ping (Sunday, cross-sector, fires even on quiet Kalshi weeks) · decision-needed pattern (first: weather unpause GATE MET page in kalshi-scorecard).
• 201 price-stream tests pass; both tsc clean; VPS + Vercel deployed; 3-tier test delivered.

First live trade in 23 days (Jul 2, post-auto-scale): SPY $751C AGG_SWING, $227 stake, DEAD_THESIS_CUT −20% same-day (−$46) — option collapsed to $0.13 by close, cut beat model stop by $42, exit broker-confirmed (alpacaExitOrderId set — integrity pipeline held on first live test). Cut-vs-hold counterfactual auto-captured (modelPnl vs pnl on same row); review rule at n≈20 cut events.

Kalshi honest read (owner asked): played MLB 25 @ 76% WR but −$0.22 real (fair-price favorites, no edge — #1330 signature); 17/25 NULL pnl (#1348); book idle since 5/24 (gates correctly blocking); model +$23.57/288 ≈ 8¢/pick = noise. Wallet ≠ model (commingled). Fixes queued in #1409.

Owner directive: full-project deep dive on Fable — website, model/data learning loop, comms, agent team + my adds (autonomy/alerting, money truth). Six parallel audit agents, all read-only/evidence-based. Report: data/audits/2026-07-01-fable-deep-dive.md. Issues #1403-#1412 filed; #1321/#1388/#1404 closed same night. Owner decisions A-D recorded in the report.

Headline findings: (1) DB live P&L was $749 rosier than Alpaca truth — 3 phantom exits incl. the #1284 orphan still booked +$220 (actual −$325); real live record 33% WR / −$1,045.47. (2) Public site badged paper +$37k as "Live · Alpaca" + scrolled fake ticker prices. (3) Root defect across all tracks: silent death — dead pipes (cicero ingest never fed once, macro-events 59d stale, Nostradamus 41% of agent-comms = 0 scoring effect) sat unalarmed for weeks. (4) ~5h/day agent compute produced unread Discord posts; learning contract inverted ~May 2 (findings→receipts). (5) #1321 edgeScore inversion RESOLVED post-June-fixes (was 60% measurement artifact); the real live inversion is LEAP_SCORE anti-predictive at edge 70+ → #1412.

Shipped same night (all verified): - #1403 CLOSED-path: backfill-live-exits.mjs (dry-run→apply; DB now −$1,041.00 = Alpaca exact) + nightly settled-fills audit on VPS (settled-audit.ts, alert-only, regression tests). 198 tests pass. - #1405 pager stack ($0/mo): pageOwner() → ntfy.sh urgent (owner escalation, CLOSE_FAILED, STARTUP_FAIL, MAIN_LOOP_ERROR, RECONCILE_GAP, daily-loss-cap hit reclassified, NEW −20%/day drawdown-watch.ts) + healthchecks.io dead-man pings (VPS 5-min + PC doctor cycle) + cron-job.org onFailure/onDisable enabled on 8 critical jobs (was 0/65) + Vercel safety-watchdog ntfy leg. Test page delivered. Owner installs ntfy app + creates healthchecks account (pending). - #1404 CLOSED — site honesty: MODEL (PAPER, 1-CONTRACT) labeling + LIVE (REALIZED) row + disclosure on public track record; fake tickers → STRATEGY_STATUS-derived brand facts; strategy counts derived (ACTIVE_STRATEGY_COUNT); canonical nav (was noindexed /brand-preview/*); MLB-as-NBA ledger fallback fixed. Deployed to Vercel prod, live-verified. - 22 shelf-ware timeslots disabled in config/timeslots.json (Onion paused, Lana daily-wrap only — owner decision A); issue-or-NO-FINDING mandate on weekend deep-research. - autonomy-doctor checks 9-10 — lift-audit/loss-ledger/macro-events staleness alarms (>48h); doctor P0s page via ntfy (breaks outbox circularity); end-of-cycle healthchecks ping = first off-machine PC watcher. - exitReason populated at settle + backfilled on all 788 settled plays (334 STOP / 204 T05 / 134 T1 / 41 EXPIRY / ...). - settled-audit + dead-man timers were armed inside the streaming-window branch (would never arm overnight / false-alarm the dead-man) — moved before the main loop, redeployed. - full cohort decomposition (n=781); closed with reopen trigger.

Root-caused session confusion: Mac had a 15-day-old zombie event-emitter replaying a stale June-16 boot backlog, because session-start.sh started the emitter on every machine and pre-compact.sh recovery state told every session to "resume event-wait.py" — both predate the Mac/PC split.

• `~/.agent/instance-role` marker (machine-local, not in git): "mac" = interactive, "pc"/missing = autonomy. Missing-marker default preserves PC behavior — PC needs no change.
• session-start.sh: skips emitter start on role=mac.
• pre-compact.sh: role-aware recovery state — Mac version says interactive session, do NOT drain timeslot events, Discord read-only.
• CLAUDE.md: new "Dual-Instance Roles (Mac vs PC)" section — the split now lives in the repo, not just memory files.
• Killed the stale Mac emitter + pidfile. Verified: hook skips emitter, breath file writes Mac variant.