Platform

A deterministic engine, a governed surface, an action layer.

Build: the engine turns any table with a yes/no outcome into a calibrated, validated model — automatically. Prove: the Model tab and evidence drawer show the receipts. Act: the action layer delivers approved actions and measures their lift. All three run on one deterministic core.

Why this account is ranked first run 3b3dc851…
outcome bound · renewal_lost · out-of-time split
leak-guard · 2 columns quarantined, reasons recorded
top-decile lift · 2.96× · holdout, not in-sample
calibration · coefficient 1.01 · refit every cycle
action staged · awaiting approval
engine-verified · replayable byte-for-byte

Every claim on this page traces to a row like one of these. That’s the product.

The six guarantees

Six things the engine does that you never have to configure.

build · no knobs

Point it at any table.

Declare the yes/no outcome column and the entity. The engine handles discovery, leak-guard, calibration, and out-of-time validation — there are no model or hyperparameter knobs to expose, so there is nothing to mis-tune.

accuracy that holds

Continuous, drift-aware recalibration.

The calibration coefficient refits every cycle and drift checks watch the series — so the 0.92 your team acts on in March still means 0.92 in September.

leak-guard

Leaked columns are quarantined and named.

Columns that encode the outcome after the fact are excluded automatically, with the column and reason recorded in the response and the audit bundle. “94% accurate in training, random in production” never happens to you.

counterfactual levers

Every finding carries the action to take.

Per-entity minimal feature changes most associated with a better outcome, always labeled as model associations, never causal guarantees.

determinism

Byte-identical replay.

Identical inputs produce identical outputs to the byte, and every response carries the engine version and core hash. Re-run any cycle and diff the envelopes.

honest-empty

It tells you when not to trust it.

Below a validated lift of 1.5 on out-of-time holdout, the engine returns a structured no with recorded reasons — and the run is never billed.

Determinism

Run it twice, get the same answer.

When we re-ran the Fable-5-trained GBM on identical data, 94 of the top 2,000 flagged customers changed between runs. Hunter-Seeker’s reruns changed zero — and every response carries the engine fingerprint that produced it.

# same inputs, twice
$ hs replay --run 3b3dc851 --diff

envelope   identical (sha256 match)
entities   2000/2000 unchanged
factors    4/4 unchanged
verdict    byte-identical
Get started

See it run on your own table.

Platform · Hunter-Seeker