Is the simulator honest? Checked against real trades

CASE FILE · CALIBRATION-ARB085-FILL-MODEL

SIMULATOR CALIBRATIONEVIDENCE RECORD2026-05-19 → 05-26 · 199 real arb events · 352 fill legs

93 / 93

STRICT FILL CHECK MATCHED REAL TRADES

Without a depth/freshness check, taker-at-ask filled 81.7% on Polymarket and 69.2% on Kalshi, even with a 2-cent aggressive offset. With depth and freshness verified: 100.0%. The depth-aware fill model's core design is what real execution demands.

ARB-085: THE OPERATOR'S LIVE CROSS-VENUE ARB BOT (REAL MONEY, REAL FILLS)
POLYMARKET + KALSHI BTC 15M UP/DOWN

PROOFLINE

SIMULATOR CALIBRATION

RECORD SEALED

352 REAL FILLS MATCHEDNAIVE MODEL REFUTEDSTRICT CHECK 93/93

WHAT HAPPENED?

Before trusting simulated trades, Proofline checked its own fill simulator against 352 real trades from a live trading bot. The naive assumption (a shown price always fills) failed often. The strict check Proofline uses (is there enough available, and is the quote fresh?) matched reality every time it was applied. That is why Proofline refuses to simulate a trade it can't support with market data.

WHAT WAS TESTED?

Is the simulator honest? Checked against real trades: Polymarket + Kalshi BTC 15m up/down · 2026-05-19 → 05-26 · 199 real arb events · 352 fill legs.

WHY TRUST IT?

The comparison used 352 real trade attempts from a live bot, matched one-for-one to exchange-confirmed records, not to the bot's own notes.

WHAT IS STILL UNKNOWN?

These numbers come from one bot, one venue pair, small trade sizes, and one week of trading. They justify the fill checker's design; they do not guarantee every simulated fill.

EVIDENCE · EXPAND TO INSPECT

PLAIN STORY FIRST · PROOF UNDERNEATH

Real fill legs from a live trading bot, matched one-for-one to exchange-confirmed records, not to the bot's own notes.

The naive assumption (a shown price always fills), measured against exchange-confirmed real trades.

The availability + freshness check Proofline's simulator uses, measured against the same real trades.

Average difference between the price the bot saw and the price the exchange confirmed.

✓Real trades matched to exchange records
✗Naive price-always-fills assumption: failed 1-in-5 (PM) to 1-in-3 (Kalshi)
✓Strict availability + freshness check: 93 of 93

CSV legs matched to log evidence352 / 352 (0 unmatched)

Price-only firing, PM (pre-gate)81.7% filled (94/115)

Depth+freshness-verified firing, PM100.0% filled (93/93)

Kalshi instant execution at ask+2c69.2% (162/234)

PM price error vs seen ask (n=187)mean +0.0003, |mean| 0.9c

FOK kills at first price21 (retries at +4c filled only 25%)

Kalshi venue-confirmed fill prices were never recorded by the bot; price error is not computable on that leg (disclosed, not estimated).
Fees were not recorded anywhere in the source data, so fee error is not calibratable.
Fill model v0.1 has no snapshot-freshness threshold yet; the 100% regime was verified at <1s quote age. This gap is disclosed on every replay that inherits it.

TECHNICAL PROOF · FOR THOSE WHO WANT IT

NEVER SHOWN BEFORE THE STORY

FINGERPRINTS (HASHES)

No hashes on this record kind.

EVIDENCE FILES · PROOFLINE REPOSITORY

compiler-benchmark/calibration/arb085_calibration_report.jsonSEALED

compiler-benchmark/calibration/arb085_calibration_report.htmlSEALED

compiler-benchmark/calibration/arb085_attempts.jsonSEALED

compiler-benchmark/calibration/arb085_leg_matches.jsonSEALED

Written by the Proofline engine; the site only displays them. See the method for how a record is reproduced.

ENGINE RECORD

CALIBRATIONREAL FILLS (EXCHANGE-CONFIRMED)

Run statusCOMPLETED

P&L DETAIL

P&LNot a P&L artifact; this record calibrates execution assumptions

VERIFIED UNDER THE DATA AND ASSUMPTIONS SHOWN. NO GUARANTEE OF FUTURE RESULTS.