the corpus is the part of the instrument you own

Every number gaige emits is a property of a corpus, a scorer, and an operating point. The scorer is public. The operating point is arithmetic. The corpus is yours, and it is the only part nobody else can own for you.

WHY OWN IT

A vendor’s calibration corpus encodes a vendor’s objectives. When an organization measures on someone else’s material, it inherits someone else’s idea of what normal looks like, silently. Owning the corpus is how an organization’s vision and objectives become operational: the material you curate is the standard your instruments answer to.

The corpus is policycuration = objectives

What you admit into the corpus defines what your instrument reports tomorrow. Curation decisions are policy decisions: owning them is the difference between measuring your standard and renting someone else’s.

Provenance or nothingpinned, hashed, seeded

Slices pin to a revision whilst the corpus file carries its sha256, and sampling starts from a seed that the receipt writes down. gaige already holds public benchmarks to this standard, and your own material deserves nothing less.

An asset, compoundingor debasing

A provenance-clean corpus appreciates: every receipt built on it strengthens the next. One un-vetted admission debases every number downstream, and nobody can tell you when it happened unless you kept receipts.

VETTING NEW MATERIAL

Before new material enters the corpus, measure what it does to the instrument. The whole flow runs with today’s commands; nothing here is roadmap:

1 · PINbefore anything runs

Hash and version the corpus as it stands. Nothing unpinned enters, nothing unpinned is measured against.

2 · CALIBRATEthe baseline

gaige run on the current corpus: AUROC with intervals, thresholds stated both ways, subgroup rates, full fingerprint. This receipt is what “our corpus, unchanged” looks like.

3 · VETthe candidate slice

Score the candidate material with the same instrument, as its own labeled slice. Compare score distributions against the baseline; stratify: aggregates hide, and the receipts show where the candidate differs, with intervals.

4 · WATCHafter admission

Register runs as a series; gaige series watch applies per-interval conformal alarm bounds. movement inside the measured variance bound is not evidence of drift, and the reports say so.

5 · REFUSEwhen evidence is thin

When the candidate slice is too small to support the conclusion you want, the honest output is a refusal, and that is what gaige ships: “A tighter guarantee than your data supports is not a guarantee.”

DIVERGENCE VS DRIFT

Two different questions hide inside “did it change?” They take different receipts, so gaige keeps them apart:

Drift“did it change over time?”

Movement against your own history: same corpus, same fingerprinted instrument, runs repeated on a schedule. gaige answers with registered series, replicate-measured variance bounds, and per-interval conformal alarms, and the fingerprint separates “the model changed” from “my measuring pipeline changed.”

Divergence“does it match the standard?”

Movement against your chosen baseline: candidate material scored by the same instrument that calibrated your trusted corpus, distributions compared with intervals, stratified so the report shows where. History does not decide this one; the standard you curated does.

Same machinery, two reference points: what you saw before, versus what you chose to trust. The machinery is also scorer-agnostic: anything that maps text to a score can be calibrated, compared, and watched. The shipped scorers are AI-text detectors, so the same receipts that measure divergence also carry calibrated indications of AI involvement: indications with disclosed error rates, never verdicts.

DISCONNECTED OPERATION

Built for denied and disconnected environments. Scoring runs on local models once the weights are staged; analysis re-derives every number from stored scores with no model, no GPU, and no network at all; receipts are plain files that cross an air gap and replay bit-identically on the other side. Nothing phones home; there is nothing to phone.

Local-first is also how the corpus stays protected. Nothing gets uploaded to a vendor to be scored. No third party ever holds the material that defines your standard, and the only copies in existence are the ones you decided to keep.

The honest boundary: first-time model staging needs a connection, or weights carried in by hand. After that, the instrument is yours.

DIVERGENCE, HONESTLY

Your vision and objectives live in the corpus you curated. What gaige adds is the receipt: whether new material shifts what your instrument reports: by how much, with what interval, and whether the evidence is sufficient to say anything at all.

The framing is deliberate. gaige can tell you that scores on new material sit somewhere your trusted corpus never sat, and it can put an interval on how far out they sit. Whether that gap means the material is off-mission is a question about your mission, not about the scores, so gaige leaves the call with you, and hands you receipts sturdy enough to survive whoever eventually questions it.

No alignment verdicts. gaige reports measured divergence with intervals. “On-vision” and “off-vision” are human judgments made on top of receipts, never emitted by the tool.
No silent standards. If the corpus changed, the receipts changed with a recorded reason, or the change didn’t happen.

QUEUED

gaige admit is the planned single-verb form of the admission check: candidate slice against corpus baseline, divergence with intervals, refusal floors. It is queued, not shipped. Today you compose the same flow from run, analyze, and series watch, exactly as above.