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 estimate168 DB tables · SQLite WAL7,444 passed in coverage run · 2026-08-12✓ green · 2026-08-12 14:38 · 6cd9da5a last Python push gate101 Telegram commandsLLM: Gemini 3.5 Flash · Gemma 2:9b fallback17 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.
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.
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.
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.