01
What replay does
Replay re-runs the same sealed inputs through the same model version and recomputes the belief and the decision. It is a recomputation, not a re-read of the recorded answer.
Proof surface
A verified outcome is a property of one run.
Verified is not a slide, a screenshot, or a backtest summary. It is seven conditions that have to hold on the same run, and a receipt that shows each of them. This surface shows the conditions, the fields, and what replay does.
live:falsesimulation-onlyno custodyabstain is a result
Not liveNo hosted service, no live effect, no customer data.
The operator entry point and the console run locally against sealed corpora and checked-in fixtures. Effects are simulated, there is no custody of funds or accounts, and there is no released hosted service. Read more on trust and posture.
- One receipt per run, replayable from the same inputs.
- Abstentions settle and stay in coverage.
- Rejected candidates never plan and never reach an adapter.
Verified outcome
Seven conditions, one run, all required.
Each condition is checked against the same run identity. A run that misses one is not counted as verified, however well the other six hold.
- 01Evidence carries observation time, availability time, and provenance.
- 02Replay over the same inputs and model version is deterministic.
- 03The belief exposes uncertainty and source lineage.
- 04The mandate admits or refuses without an override.
- 05The run ends in a typed act or abstention receipt.
- 06The outcome resolves and binds to that exact run.
- 07Settlement emits evaluation and learning receipts.
the count
The published count is weekly verified outcomes: completed runs that satisfy all seven conditions inside the rolling window. It is not calls, orders, bets, stakes, or profit.
abstentions
A reasoned abstention that satisfies the same conditions is a verified outcome. Abstention rates stay visible because a system that declines correctly is measurable.
rejections
A candidate rejected at the promotion gate is not counted as a verified outcome, and it leaves the count unchanged. Nothing was planned and no adapter was called.
Receipt anatomy
Open a receipt field and read what it binds.
Pick a sample run, then open any field. Each one says what it carries and why the product keeps it, and you can copy the whole record as JSON.
sample data
Sample runs
Inputs
Outcome
Integrity
decision
outcomeWhat it carries
Admitted under mandate
Why it matters
Exactly one typed outcome per run: an admitted act, a reasoned abstention, or a rejection that never became a decision. Abstention carries its reason and is a complete result.
- Why it ended this wayThe pre-off window had not closed, the quote was still reachable, and the stake stayed inside both the mandate bound and the ruin policy.
Sample record for sport and racing · field names follow a recorded run receipt, values are illustrative.
{
"run": {
"id": "run-sport-0714",
"domain": "sport",
"market": "Ascot 15:40 — 8-runner handicap",
"decision_at": "2026-03-14T15:32:00Z",
"outcome": "act"
},
"evidence": [
{
"label": "Field and draw",
"observed_at": "2026-03-14T14:05:00Z",
"available_at": "2026-03-14T14:06:00Z",
"source": "Declared racing corpus · sealed digest",Show the complete receipt (125 lines)
{
"run": {
"id": "run-sport-0714",
"domain": "sport",
"market": "Ascot 15:40 — 8-runner handicap",
"decision_at": "2026-03-14T15:32:00Z",
"outcome": "act"
},
"evidence": [
{
"label": "Field and draw",
"observed_at": "2026-03-14T14:05:00Z",
"available_at": "2026-03-14T14:06:00Z",
"source": "Declared racing corpus · sealed digest",
"detail": "8 declared runners, 2 late non-runners excluded before inference."
},
{
"label": "Pre-off quotes",
"observed_at": "2026-03-14T15:20:00Z",
"available_at": "2026-03-14T15:20:30Z",
"source": "Quote snapshot · per-runner availability time",
"detail": "Best reachable price per runner; stale quotes past the window refused."
},
{
"label": "Going and conditions",
"observed_at": "2026-03-14T13:40:00Z",
"available_at": "2026-03-14T13:41:00Z",
"source": "Course bulletin",
"detail": "Going updated twice; only the reading available at decision time was used."
}
],
"belief": {
"probability": 0.62,
"interval_low": 0.54,
"interval_high": 0.69,
"model": "Ranked statistical model",
"version": "v3.4.1",
"note": "Ranked belief across the field, renormalised after non-runners.",
"calibration": [
{
"caseId": "cal-0.50",
"believed": 0.5,
"observed": 0.497,
"sampleSize": 420
},
{
"caseId": "cal-0.60",
"believed": 0.6,
"observed": 0.594,
"sampleSize": 388
},
{
"caseId": "cal-0.70",
"believed": 0.7,
"observed": 0.706,
"sampleSize": 301
},
{
"caseId": "cal-0.80",
"believed": 0.8,
"observed": 0.792,
"sampleSize": 214
},
{
"caseId": "cal-0.90",
"believed": 0.9,
"observed": 0.884,
"sampleSize": 96
}
]
},
"mandate": {
"objective": "Paper-first racing decisions under a fixed stake ceiling",
"expires_at": "2026-04-01T00:00:00Z",
"checks": [
{
"label": "Mandate present",
"requirement": "Signed, unexpired authority",
"result": "pass"
},
{
"label": "Confidence floor",
"requirement": "at least 0.58 after interval shrink",
"result": "pass"
},
{
"label": "Stake bound",
"requirement": "1.5% of declared bankroll or less",
"result": "pass"
},
{
"label": "Ruin policy",
"requirement": "Worst-case drawdown inside policy",
"result": "pass"
},
{
"label": "Permitted effect",
"requirement": "Simulation adapter only",
"result": "pass"
}
]
},
"decision": {
"kind": "act",
"reason": "Admitted under mandate",
"detail": "The pre-off window had not closed, the quote was still reachable, and the stake stayed inside both the mandate bound and the ruin policy."
},
"effect": {
"adapter": "simulation",
"idempotencyKey": "sport-2026-03-14-1540-runner-4",
"receipt": "sha256:6b31c2f7…a9d4",
"calls": 1
},
"settlement": {
"result": "Observed outcome: 1 winner; simulated stake settled at the quoted price.",
"resolvedAt": "2026-03-14T15:58:00Z",
"evaluation": "Evaluation receipt bound to run and effect identity.",
"learning": "Learning receipt: belief stood above the floor; price was the binding constraint."
},
"integrity": {
"runFingerprint": "sha256:1f0a94be…7c22",
"replay": "deterministic",
"live": false
}
}Illustrative worked example. Field names match a real run receipt; the values are samples, and no production run exists yet.
Replay
Same inputs. Same fingerprint. Zero adapter calls.
Replay is the check that separates a recorded decision from a described one. It is deterministic by construction, and it cannot reach an effect.
02
What it proves
That the recorded decision is the decision those inputs produce. The run fingerprint names the exact inputs, model version, mandate, and decision, so two runs can be compared field by field.
03
What it never does
Replay performs zero adapter calls. It cannot place an order, take a bet, or move funds, and a replay that disagrees with the record stops the run before anything downstream happens.
How a retry differs from a replay
A retry presents the same idempotency key and reads the original effect receipt instead of acting twice. A retry that presents a known key with a different requested effect fails closed, and the original receipt stands. Replay is a different check again: it recomputes the decision and performs no adapter call at all.
What a refusing boundary looks like instead
A fact was published after the decision time. Refused before inference. No belief is produced and nothing downstream runs. You can watch that boundary run in the quickstart, and every named refusal is listed in the refusal reference.
Read the method, then run one.
The verification surface lists the checks a run must pass and what would falsify each claim. The quickstart takes you from one command to one receipt on your own machine.