GME Agent Team · roster from AGENTS.md (2026-08-12) · chips computed 2026-08-12 17:22

The Desk, on one page

Eleven external feeds flow into one scheduler, fan out across 28 agents in four lanes, land in one SQLite database, and surface on Telegram. Exactly one path — the highlighted one — is scored, paper-traded, and feeds back into what the system acts on.

28 agents (AGENTS.md headings) ~177 scheduled jobs · documented estimate 168 DB tables · SQLite WAL 7,444 passed in coverage run · 2026-08-12 ✓ green · 2026-08-12 14:38 · 6cd9da5a last Python push gate 101 Telegram commands LLM: Gemini 3.5 Flash · Gemma 2:9b fallback 17 circuit breakers

System map

Left to right: where data comes from, who consumes it, and where humans see the result. Every lane reads and writes agent_memory.db; only the Signal Desks lane passes through the gate pipeline.

SOURCES TradingView WS Finnhub news SEC EDGAR FINRA short vol Polygon / Massive LSE API NYSE halts CSV TwitterAPI.io CourtListener GameStop IR RSS Google Trends 17 breakers ORCHESTRATOR · APSCHEDULER · ~177 JOBS SIGNAL DESKS — scored & paper-traded Pattern Intraday · 5mTrendy · 4h Futurist · 2h (shadow)Scalper · 1m open-range OptionsFlow · 2h (underpowered) every emit carries 0.0–1.0 confidence CONTEXT & ANALYSIS Valerie · 5mChatty · 5m Newsie · 30mGeoRisk · 1h FlowDetect · 30mSynthesis · 5m (shared brief) blackboard facts only — verdict posts refused in code WATCHERS — silence-as-signal IntelWatch · 30mTroveIntelPush · 30m HaltMonitor · 1mMergerWatch · 4h 24/7 EconCalendar · 30m 24/7Social · daily 08:15 Form4DatasetWatch · wkTroveIntelStory · 12:30 a push always means a state transition — the default is silence META & AUDIT — detection-only Historian · 09:15+16:45CTO · daily crons Delorian · 08:45+FriAgentX · Sun weekly Boss · 09:00Briefing · 10:00 verdicts are deterministic — the LLM only narrates signal_gate EMIT / SHADOW signal_manager alert + paper trade agent_memory.db SQLite WAL · 168 tables nightly 03:00 backup signal_scores · agent_logs … read / write TP/SL outcomes → signal_scores calibration ×mult · LIVE 2026-07-27 SURFACES Telegram — primary UI 101 commands · public channel + owner DM Discord (routed mirrors) OBS overlays (read-only) /obs/floor · /obs/telegram · /obs/ledger IBKR (live 2026-07-24) small size · first fills routed 2026-07-29 EMIT only notify() direct — no gate polled
The highlighted path is the only one with consequences: a signal-desk emit passes signal_gate, opens an automatic paper trade, gets scored in signal_scores, and the calibration multiplier flows back into the confidence the system acts on. Everything else — watchers, meta, context — pushes cards or context with no position and no score.

The learning loop

The part that makes this a learning system rather than a notification firehose. One loop is closed and live; three siblings still run detection-only.

agent emits confidence 0.0–1.0 signal_gate rejected agents → SHADOW paper trade TP/SL checked every 5m signal_scores win-rate per agent calibration · 10m per-agent multiplier EMIT acting_confidence() — the multiplier gates severity & stored confidence · LIVE since 2026-07-27 · auto-revert armed cross-examiner dock · SHADOW only (records, never applies) Detection-only siblings still waiting on powered verdicts: cross-exam dock · overlay loop · regime loop (applied=0).
Promotion was gated: 30/30 sustained HELPS days plus a forward out-of-sample replication before the multiplier was allowed to touch live confidence. If either stage flips to HURTS it reverts to display-only with no grace period.

The hypothesis engine

The part that makes this a research desk, not a prediction machine. Every claim — agent skill, factor edge, structural pattern — runs a pre-registered gauntlet with frozen bars before it touches live behaviour. The graveyard is public. The falsified outnumber the accepted 5:1, and that ratio is the proof the rigour is real.

ideate quantpedia · operator · agent observation pre-register H₀ + H₁ frozen promote / kill bars set decision date frozen before any data runs forward-test clock walk-forward observations no peeks (NO PEEKS policy) n accrues per arm ~48 clocks currently live adjudicator decision date fires verdict computed by the frozen pure function — not a human call accepted ✓ 38 confirmed edges rejected — tombstone 214 public falsifications underpowered → stays open HELPS HURTS UNDERPOWERED SHADOW STAIRCASE — detection before action calibration loop (LIVE · applied=1) · cross-exam dock · overlay loop · regime loop · HMM regime — each gated by a 2-stage autowire: 30/30 sustained + forward-OOS replication before touching live behaviour, auto-revert on HURTS
Every edge on the desk ran this gauntlet. The calibration multiplier — the only self-healing loop currently live — was promoted only after 30/30 sustained HELPS days plus a forward-OOS replication, and auto-reverts on HURTS with no grace period. The three detection-only siblings (cross-exam, overlay, regime) follow the same staircase but have not yet cleared both stages. α-wealth 0.094 (solvent) — the multiple-testing budget; spending it on dead ends is the cost of looking honestly.
SELECTED TOMBSTONES (the desk keeps its falsified edges public)
✗ Pattern daily — 22.7% on n=21, significantly anti-correlated. The timing question closed.
✗ Cheap-price premium — t=+2.79 collapsed to t=−1.24 under PIT pricing. Back-adjusted prices were a future-winner detector.
✗ Midday-overlay — +13.4pp in-sample reversed to −10.2pp out-of-sample. The winner's curse, measured.
✗ 15 VRP conditioners — vol regime, event dates, tenor, drawdown, dispersion: all dead. Every conditioner used information the market can also see.
✗ Cross-sectional VRP harvest — premium is real (+0.02 vp/yr, ρ=0.77) but cost eats it at retail (8.22% round-trip in the meme tier).
✗ DV v4 mechanism — the grid claim holds (t=3.12); the "DR rescues v2" story doesn't (ρ=−0.222).
✓ Schema-spread calibration — the one accepted row from the VRP arc. cbbo-1m overstates trade-time spreads 4.09×.
✓ DV/DR v3b — +9.6pp top-to-bottom on 1,012 names. The desk's strongest measured edge.
214 tombstones total · 38 accepted · 48 on forward-test clocks · α-wealth 0.094 (solvent) · the falsified outnumber the accepted 5:1

Honesty notes. Live reads at page load: agents from AGENTS.md headings, tables from sqlite_master, commands from the COMMANDS registry, and breakers from distinct get_breaker() names in source. The test total is the last coverage run, not CI. The push-gate chip is written only when the local pre-push gate actually runs; unverified means no recorded Python push gate on this checkout. The ~177 jobs figure is a documented estimate — 158 static add_job sites plus loops; not independently recounted. The map and roster are a dated snapshot of AGENTS.md (see the header), so trust make status / make ledger / the live overlays when they disagree.

The roster

All 28 agents from AGENTS.md, grouped by lane. Status chips are the ledger's verdicts, not aspirations — this desk keeps its falsified agents on the books, silenced, accumulating evidence.

Signal desks

The only lane whose output is scored, paper-traded, and calibrated.

📐 Pattern Intraday 5 min

Triangle/wedge breakouts, confidence floor 0.80, repeats deduplicated. Daily Pattern was falsified (22.7% on n=21) and unscheduled 2026-06-19.

borderline emit · 56% n=241
📈 Trendy 4h + 20:00 EOD

Deterministic UP/DOWN/SIDEWAYS verdict; structural edges enrich the narration only.

live
🔮 Futurist 2h

1h/4h forecasts with the structural-edges payload. Baseline falsified (46.2% on n=130); midday-window overlay under pre-registered forward test.

shadow · rehab in progress
⏱️ Scalper 1 min · 09:46–10:59

Opening-range touch-and-turn, one alert per session, liquidity and gap gates.

live
🌊 OptionsFlow 2h

Chain-positioning interpreter. 35% on n=20 — below coin-flip but too thin to call falsified; accumulating sample, scoring decoupled from broadcast.

underpowered

Context & analysis

Feeds the shared picture; makes no scored calls.

✅ Valerie 5 min

OHLC sanity checks — stops every downstream agent reasoning on garbage data.

live
💬 Chatty 5 min

Stream colour between real signals. Window-gated after burning ~288 LLM calls a night narrating the prior close.

live
📰 Newsie 30 min

Finnhub headlines → deterministic composite sentiment + one grounded narrative line.

live
🌍 GeoRisk 1h

Geopolitical risk scoring with search grounding.

live
🧠 Synthesis 5 min

One-line consensus every agent re-reads next cycle; pushes only on a headline transition.

live
🎰 FlowDetect 30 min

LSE prints ≥ $50k flagged BLOCK / SWEEP / UNUSUAL. Deliberately makes no directional claim — that hypothesis already failed.

detection-only

Watchers

Silence-as-signal: a card always means a state transition. Most run 24/7 because their best alerts land off-hours.

🛰️ IntelWatch 30 min

Massive structural surfaces — insider buys, short-vol stress line, deal cadence, ±2σ stretch. Hosts the only tool-use pilot (off by default).

detection-only
🎰 TroveIntelPush 30 min

Fires once per structural transition: "newly appeared" and "⌛ concluded". Cards are a transition log, not a state board.

detection-only
🧵 TroveIntelStory 12:30 ET Mon–Fri

One narrated story a day over the desk's own output. A tripped honesty gate withholds the story — never softens it.

detection-only
🎯 SetupWatch 30 min

Three-filter setup candidates (substance + agreement + AXTI shape) on newness. Research intel, not a signal — card says "passed the filters", not "buy".

detection-only
🛑 HaltMonitor 1 min

NYSE halt feed diff. ~390 polls a day, a handful of pings a year — the ping is the event.

live
🤝 MergerWatch 4h · 24/7

Form 425 deal communications only, under the target's CIK. Base rate ~0/decade, so any push is high-signal.

detection-only
📅 EconCalendar 30 min · 24/7

Tier-1 macro releases (CPI, NFP, FOMC, PCE) with actual vs consensus surprise.

detection-only
🔍 AttentionWatch 30 min · 24/7

Fires when GME search interest or the retail fuel composite crosses 70 — with a 65 re-arm band so flapping at the line can't re-announce.

detection-only
🐦 Social daily 08:15

Cohen / DFV / Gill / Burry watch. Critical accounts burst immediately; Burry posts push on any keyword hit.

live
🗂️ Form4DatasetWatch Mon weekly

Probes the SEC quarterly insider dataset; ingests and propagates pillars when a new quarter lands.

detection-only

Meta & audit

The desk watching itself. Verdicts are computed in Python; the LLM only narrates. None of these move a weight.

🏛️ Historian 09:15 + 16:45

T+35 FTD settlement clocks, lesson relevance, agent accuracy trends.

detection-only
🛡️ CTO daily crons

DV score with its own forward record stamped on every card, plus the end-of-day honest-assayer audit of the other agents.

detection-only
🚗 Delorian 08:45 + Fri 16:55

Replays yesterday: ceiling (perfect hindsight) vs capture (what the calls actually caught). Audits graduated lessons out-of-sample.

detection-only
📊 AgentX Sun 09:15 weekly

Five-dimension maturity assessment with a goal-substitution flag — "is the desk's time flowing to the right dimension?"

detection-only
👔 Boss 09:00 / 10:00

Daily mission frame; reviews the previous day.

live
📋 Briefing 10:00 + /brief

The polished human-readable digest, with day-of-week character.

live
📐 Calibrator 10min + hourly

Scores predictions the moment their horizon passes; fires only on an accuracy-state transition (NORMAL/DEGRADED/INVERTED). Feeds the calibration loop.

detection-only

System cycles (not agents, still load-bearing)

Plumbing and un-headlined watchers the orchestrator also runs — the gap between "28 agents" and "~177 jobs".

SEC filing monitor · 24/7 burst IRPress · 30m 24/7 LegalWatch · 4h 24/7 WarrantWatch · hourly IndexWatch · S&P eligibility voice forwarder · 1m calibration · 10m paper-trade checker · 5m tick aggregator · 5m bond yields · 21:15 Monday options · 08:30 standup · 11:00 + 16:00 DB backup · 03:00