04

HERMES

live
deep-dive · HERMES · sub-second event-driven microservice · skeleton 2026-04-25back to overview
command link · unreachable
backend did not respond. every panel below reflects the last cached render, not live state.fetch failed

Hermes · sub-second event-driven microservice

skeleton — pre-regs locked 2026-04-25 · paper Phase 1A targets 05-01

live P&L—no resolutions yet
statusskeletonno event loop yet
strategies2FORGE·H · VANGUARD
bets opened00 resolved
host serviceofflineunknown

strategies

Phase 1 backtest gated by 04-30 Tier 1 audit
FORGE·H
Sub-second merge-arb (Phase 1A)

Port of Prophecy FORGE to event-driven raw asyncio. Reacts to Polymarket WebSocket orderbook deltas in 200-500ms. BUY YES + BUY NO same conditionId when sum < $1.

Evidence

Prophecy FORGE: $686 lifetime / 4,206 trades at 5-min cadence. Expected +30-50% capture vs slow path.

Live hit

—

n=0

Live P&L

$0.00

warming

Penny scale ($1/trade × 100-300 pairs) at first live · no live orders until Phase 3A (target 05-13)default · $1/leg
VANGUARD
METAR trajectory front-running (Phase 1B)

When NOAA AviationWeather publishes a fresh METAR, scanner re-evaluates active buckets via TrajectoryModel v1 and BUYs YES on cheap mispriced buckets. Edge = lag between raw METAR publish and Wunderground display update.

Evidence

3 witnesses: HondaCivic2 +28pp edge n=103, Joe/Henry pair +13.6pp n=305, WeatherTraderBot +6.7pp · single-regime sample

Live hit

—

n=0

Live P&L

$0.00

warming

Statistically thinner than Forge — gated behind Forge proof · penny scale onlydefault · $1/bet

phasing roadmap

per 05_MASTER_PLAN_v2.md
  • Phase 0 · now → 04-30Wunderground per-city lag, Seoul rounding rule, NOAA cache fallback, both pre-regs locked, latency baseline.
  • Phase 1A · 05-01 → 05-04hermes_forge backtest + /quant-risk + /quant-researcher + /quant-validator.
  • Phase 1B · 05-04 → 05-08hermes_vanguard backtest (independent validator dispatch).
  • Phase 2A · 05-08 → 05-13hermes_forge skeleton + shadow mode.
  • Phase 3A · 05-13 → 05-20hermes_forge penny-scale live ($1/trade × 100-300 pairs).
  • Phase 2B + 3B · 05-20 → 05-31hermes_vanguard skeleton + penny live.
  • Phase 4 · 05-31 → 06-14Production scale ramp ($5 → $20 → $50 / trade).

about this generation

Why a separate microservice?Hermes runs raw asyncio + persistent WebSocket + event-driven scanners. APScheduler (used by weatherbot) doesn't compose with this model. Separate process = crash isolation, independent restart, independent logs — same DB, same git repo.

Why two strategies? They share one event loop and one execution layer. hermes_forge is the validated mechanism (de-risks the microservice infrastructure). hermes_vanguard is the speculative bet (METAR trajectory front-running) that rides on the debugged platform.

Latency target: 200-500ms METAR-publish → order-submitted. Polygon block time (2s) is the floor — Hermes aims to be FIRST in the next block, not faster than physics.

Skeleton at: /opt/weatherbot/src/weatherbot_hermes/. Master plan: docs/strategies/foresight/05_MASTER_PLAN_v2.md.