A timing-based bot: buy the favorite 9 minutes in

CASE FILE · REPLAY-LEGACY-EMA-FAVORITE-15M

HISTORICAL REPLAYVERIFIED · ADAPTED300 windows · depth era (from 2026-06-10) · 685,018 events

NEGATIVE

NET −$55.99 · 300 WINDOWS · TIMING VERIFIED · NO % BASIS

Window-relative and reproducible: every entry landed inside the intended 9-minute gate and every window had a recorded opening reference, and the strategy still lost. The engine models the timing and the movement faithfully; the loss is the strategy's.

KALSHI 15M EMA-FAVORITE EDGE (THE C3 CLOSE BOT) — RECOVERED FROM SOURCE
KALSHI BTC 15-MINUTE UP/DOWN (FAVORITE SIDE, EITHER UP OR DOWN)

PROOFLINE

HISTORICAL REPLAY

RECORD SEALED

161 ENTRIES · ALL IN 9-MIN GATEMOVEMENT FROM OPEN · RECORDED300 SETTLED · 0 OPENREPLAY BYTE-IDENTICAL

WHAT HAPPENED?

This is the operator's own 15-minute close bot. Nine minutes into each Bitcoin window it buys the market's favorite side (priced 70–92¢) only if the price has moved that way since the window opened and is on the right side of its moving average, then holds to the result. Proofline recovered the exact rules from the bot's source, measured the movement from each window's opening price and the exact time of each entry, and replayed it on 300 real windows. Every one of the 161 trades happened inside the intended 9-minute entry window; all 300 windows had a real opening reference price. The bot lost $55.99 net after $38.77 in fees.

WHAT WAS TESTED?

A timing-based bot: buy the favorite 9 minutes in: Kalshi BTC 15-minute up/down (favorite side, either up or down) · 300 windows · depth era (from 2026-06-10) · 685,018 events.

WHY TRUST IT?

The entry time (seconds since the window opened) and the opening reference price are recorded in the source data, not guessed; every trade was checked against real order-book depth, every window settled to a real Kalshi result, and rerunning the sealed record reproduces the exact same report.

WHAT IS STILL UNKNOWN?

This does not mean the strategy works, and it is not a plain Verified Record. Fills are modeled as taker-at-ask (the bot rests a maker at the mid, which cannot be depth-verified); sizing is a fixed $20 (the bot uses fractional Kelly capped at the same 2%); and the bot's trading-hours session filter is not applied. The core timing + movement + moving-average signal is faithful. See the disclosures.

EVIDENCE · EXPAND TO INSPECT

PLAIN STORY FIRST · PROOF UNDERNEATH

Verbatim from the sealed record. The technical proof below lists the artifacts behind it.

How many times the rules fired during the tested period.

Verbatim from the sealed record. The technical proof below lists the artifacts behind it.

Positions still awaiting a final result when the record ended. Anything pending stays pending; results after the recording stopped are never filled in.

The sealed spec re-run on the c3-window-relative DatasetVersion produces a byte-identical report_hash (6e893d65…) across independent runs; the ledger chain (325ca2bc…) is stable regardless of engine code_version.

✓Strategy rules recovered from source
✓Entry timing (seconds into the window)
✓Movement measured from the window's opening reference
✓Whether trades could go through (both sides)
✓Fees
✓Final market results
✓Profit/loss
✓Reproducibility

Events replayed685,018

Entry evaluations1,173,421

Intents161 (85 up + 76 down)

Entries inside the 540–660s gate / outside161 / 0

Rejected for missing timing or opening price0

Fills161 (147 depth-verified, 14 partial, 0 unverified)

Settlements applied300

Positions closed / open161 / 0

This is a Verified ADAPTED Record. Execution is modeled as taker-at-ask with depth verification; the bot rests a maker at the mid and escalates to taker, and maker fills cannot be depth-verified from the tape. This is the honest execution model, and it is stricter than the bot's assumed mid fill.
Sizing is a fixed $20 per trade (= 2% of a $1000 bankroll); the bot uses fractional Kelly capped at the same 2%, which only ever reduces size below it. No percentage return is stated: a fixed dollar size has no declared capital basis, so a percentage would be invented.
The bot's trading-hours session filter (trade 04:00–20:00 UTC, skip the 12:30–14:00 macro window) is not modeled: the replay engine has no hour-of-day gate. The core signal (favorite + movement-from-open + EMA + timing + price band) is faithful.
Movement is measured from each window's recorded opening reference price (window_open_price), a market-input fact — not a strategy return. The −$55.99 is the strategy's dollar outcome, not a percentage.

TECHNICAL PROOF · FOR THOSE WHO WANT IT

NEVER SHOWN BEFORE THE STORY

FINGERPRINTS (HASHES)

SEMANTIC SPEC FINGERPRINT

c98f8194b3057d747b3ed4c7cccb2e0f1256e7baee47d7733a653ac9f76a18b9

SPEC HASH

5008b19c43d1a5a550676b207e8719df81b89ba0e284088acba3f1006f89dd81

REPORT (TWO-SIDED V0.2)

6e893d656aa87409d7a0234a1e99bc8c2fc5ce087dd2c7d2a796f543e841b7e4

LEDGER CHAIN

325ca2bc07b08ff5eecb195194b1225ec97333155c42ca21f4bd22601116bf28

DATASETVERSION (WINDOW-RELATIVE, NORMALIZED)

574f95c0a31f090a44e19faa9ff9943e6f817b6cb6644a3c0d939f6598457951

REPRODUCIBILITY

EVIDENCE FILES · PROOFLINE REPOSITORY

compiler-benchmark/legacy/ema-favorite-15m/spec.jsonSEALED

compiler-benchmark/legacy/ema-favorite-15m/report_ema_favorite.jsonSEALED

compiler-benchmark/legacy/ema-favorite-15m/M8_NUMBERS.jsonSEALED

compiler-benchmark/replays/c3-window-relative/dataset/manifest.jsonSEALED

compiler-benchmark/replays/c3-window-relative/coverage.jsonSEALED

docs/integration/FIRST_WINDOW_RELATIVE_STRATEGY_SELECTION.mdSEALED

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

ENGINE RECORD

REPLAYREPLAY ONLYLEGACY IMPORTWINDOW-RELATIVETWO-SIDEDP&L: COMPUTABLEBYTE-REPRODUCIBLE

Trust levelREPLAY ONLY

Run statusCOMPLETED

P&L statusCOMPUTABLE

P&L DETAIL

Net P&L−$55.99

Realized P&L−$17.22

Fees$38.77

Max drawdown−$139.04

Two-sided spec on the Kalshi up book + its exact binary-complement down book. secs_elapsed = recorded t_into; window_open_price recorded from the bot's trades log; window_delta_pct computed from recorded spot + open; EMA = the bot's recorded 100-EMA. Semantic hash c98f8194….

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