Out-of-sample, cost-aware, merge-gated

Quant research

Research decisions where promotion depends on held-out evidence, not on in-sample fit.

Run it locally Module reference Walk the loop

Evidence

What this module supplies.

Distributed training merges candidate parameters across shards, then evaluation compares the candidate against the baseline on the same sealed out-of-sample cases. Promotion requires beating the floor after deflation; in-sample fit alone can never promote a candidate.

01Immutable point-in-time market and research corpus
02Artifact and parameter lineage per shard
03Development and out-of-sample case identities kept disjoint

Gates

What gates every run.

Gates are evaluated before a decision is emitted, not at execution. A run that cannot clear one ends in a reasoned abstention or a rejected candidate instead of an effect.

gate 01Out-of-sample coverage floor with abstentions counted
gate 02Deflated performance floor after selection cost
gate 03Capital-risk admission before any position intent

Limits

What this module never does.

These are contract lines rather than caveats: the module has no path that requests them, in any configuration, on any surface.

neverNo broker, exchange, or capital path is contacted
neverNo promotion from in-sample fit alone
neverNo synthetic digest standing in for real lineage

LimitProduct-wide, in every module

· No live decision, order, wager, or external effect exists in the shipped adapter.
· No custody of funds, accounts, or customer assets, and no counterparty reach-through.
· No guaranteed profit, no promised strike rate, and no performance claim sold as fact.
· No action when evidence, authority, replay, settlement, or effect identity is incomplete.

Worked sample run

The record this module produces.

An illustrative sample from this module, with the charts reading the numbers the run reports. The values are samples, and no production run exists yet.

sample dataCross-sectional equity signal — 400 names, 30-day horizon

decision clock 09 Feb 2026, 21:05 UTCAdmitted act

Four training shards merged, the candidate cleared the out-of-sample floor after deflation, and the resulting position intent was simulated.

belief

0.00belief density1.00

p = 0.58interval 0.51–0.64interval width 13%

Recorded gradient model v1.9.0 — Per-name scores aggregated to a single portfolio-level belief with interval.

believed against observed, sealed holdout

believed (x) against observed frequency (y) · dashed line is perfect calibration

evidence admitted

Point-in-time pricesImmutable market corpus · sealed digest
observed 09 Feb 2026, 20:00 UTC · available 09 Feb 2026, 20:00 UTC400 names across 30 sessions; corporate actions applied as known at each date.
Research featuresFeature store snapshot
observed 09 Feb 2026, 19:30 UTC · available 09 Feb 2026, 19:31 UTCFifteen features, each carrying its own availability stamp.
Holdout sealEvaluation ledger
observed 09 Feb 2026, 20:40 UTC · available 09 Feb 2026, 20:40 UTCDevelopment and holdout case identities proved disjoint before scoring.

mandate checks

Objective: Research decisions with a hard capital-risk ceiling

Mandate presentSigned, unexpired authoritypass
OOS coverageat least 0.80 of holdout cases answeredpass
Deflated floorMedian bps above floor after selection costpass
Capital riskWorst-case loss inside declared ceilingpass
Permitted effectSimulation adapter onlypass

authority expires 30 Jun 2026, 00:00 UTC

decision

DecisionAdmitted under mandate

The merged candidate beat the baseline out of sample, coverage cleared the floor, and the position intent stayed inside the capital-risk ceiling.

effect and settlement

adapter: simulation
adapter calls: 1 · idempotent
idempotency key: quant-2026-02-09-h30-candidate-b7
effect receipt: sha256:c40e8a11…3f70

settlement: Horizon closed: realised return recorded against the simulated intent.
evaluation: Evaluation receipt with deflated score and coverage.
learning: Learning receipt: features aged past their half-life were the main error term.

run fingerprint sha256:9db2170c…4e18replay deterministiclive:false

Illustrative worked example. Field names match a real run receipt; the values are samples, and no production run exists yet.

Quant researchAdmitted act

Cross-sectional equity signal — 400 names, 30-day horizon

09 Feb 2026

Horizon closed: realised return recorded against the simulated intent.

Sample runs · illustrative records, not production
Decision	Outcome	Settled	Run
Cross-sectional equity signal — 400 names, 30-day horizon09 Feb 2026	Admitted act	Horizon closed: realised return recorded against the simulated intent.	open

Next steps

Keep going from here.

Read the contract, run the loop, then read what the run recorded.