arb_audit_log (Postgres) at
detection time, then reconciled to resolution outcome the instant
the market closes. The aggregate is queryable via
GET /v1/arb/audit/aggregate
or browsable at the public audit page.
Four tiers
| Tier | Trigger | Direction | What it claims |
|---|---|---|---|
logical | yes_ask + no_ask < $1.00 | BUY_BOTH | Math-guaranteed profit at resolution. |
endgame | model says near-certain + dual-oracle agreement, in final 0–120 s | BUY_YES or BUY_NO | High empirical hit-rate from book-lag in the last seconds. |
stable | probability-arb on the maker-zone book (fill ≥ $0.30) | BUY_YES or BUY_NO | Where the maker bot is parked — usually well-calibrated. |
stale | probability-arb on a thin / deep book (fill < $0.30) | BUY_YES or BUY_NO | Deep mispricings, often already arbed by HFT. |
/v1/arb/live endpoint returns currently-
firing opportunities across all four tiers in real time.
Audit record per detection
When the detector emits an opportunity, the following is logged before any user sees it:| Column | Type | Notes |
|---|---|---|
detected_at | timestamptz | Wall-clock at detection. |
market_id | text | The CLOB market hash. |
ticker / event_type | text | BTC / 5m, etc. |
tier | text | logical, endgame, stable, stale. |
direction | text | BUY_YES, BUY_NO, BUY_BOTH. |
fill_price | numeric | Real ask we’d hit. |
model_yes_prob | numeric | Our log-normal model’s P(YES). |
model_ev_per_share | numeric | Net EV after entry fees. |
est_fee_per_share | numeric | Polymarket 2026 taker fee. |
resolution_at | timestamptz | When the market will resolve. |
events.resolution_outcome and
fills in:
| Column | Type | Notes |
|---|---|---|
resolved_at | timestamptz | Reconciliation time. |
resolved_outcome | text | yes_won / no_won. |
realized_pnl_per_share | numeric | payoff − fill_price − est_fee, computed correctly for each direction (BUY_YES, BUY_NO, BUY_BOTH). |
/v1/arb/audit/aggregate
is unfiltered count(*), avg(realized_pnl_per_share), etc.
Calibration
The aggregate returns bothavg_model_ev and avg_realized per tier
so you can compare. A model whose avg_realized ≈ avg_model_ev is
well-calibrated; the gap is the engine’s honest blind-spot.
For BTC/ETH/SOL Up/Down markets across our entire log, the probability
arb (stable + stale) tiers have realised ~zero edge — Polymarket’s
maker bots are well-calibrated for crypto. The logical tier is
positive by construction, and the endgame tier has shown ~95 %
empirical win rate from book lag in the final seconds.
Read more on the public audit page —
same numbers, no signin required.