A dossier an
inspector can rebuild
A Context of Use and seven credibility steps, with the model and prompt pinned for each stage.
Declared, derived,
or pending
Declare a Context of Use. Grade the model risk. Then Antevra indexes what it recorded about one run or one review: which artifacts were produced, which model answered each call, and what a named person declared about its adequacy. Every item is content-addressed by hash.
Built from records the engine was already keeping for honesty’s sake.
Dossier · STUDY-A · seven steps
The question this run answersdeclared
The decision it influences, and what is out of scopedeclared
Model risk: influence × consequence, no input of its ownderived
Which artifacts will be the evidencedeclared
Each planned artifact found, hashed, locatedderived
Every model claim backed by the ledgerderived
A named person’s determination, append-onlyblocked · unbacked_model_claim
A page with no button
that clears a blocker
A blocked dossier lists every blocker with the artifact and locator it concerns. “There is no override, and no button on this page can clear them.” Adequacy stays pending, and “adequate” is not offered while the dossier is blocked.
Every blocker, listed
Not just the first. The evidence table names each artifact by hash and says NO where the ledger is missing.
Seven steps and a Context of Use
The question of interest and the decision it influences, model risk derived from influence and consequence, a declared plan, and incomplete until someone records a determination.
Reconstruction
One row per model call with the model that answered and the hashes of the prompt and output. Where a fixture answered instead of a model, the page says so rather than naming a model that did not run.
The command line returns a non-zero exit code on a blocked dossier, so a script cannot miss it.
AI is in the filing;
the record often is not
Teams have model-generated analyses in their work and no systematic record of what was asked, what answered, and on what basis anyone believed it.
ICH E6(R3) (opens in a new tab) makes computerized-system validation a sponsor accountability: tools pinned by version, inputs archived, outputs reconstructable. Assurance is shaped to produce that record.
The model on record is the one that answered
Before the dossier could be built, the record had to be corrected: some places wrote a requested tier or a pre-resolved id as “the model”, including on a cache hit when no model ran at all. Now the id that reaches any record comes from the reply a backend returned, captured at the one seam every call passes through. A stage that did not run is skipped and names no model.
Every evidence item that makes a model claim carries ledger_backed. A dossier with an unbacked claim is blocked, and the blocker names itself: unbacked_model_claim or unverifiable_artifact.
An index over existing
artifacts
The run and the review already produce versioned artifacts that carry their provenance, and the server already content-addresses every stored file. The dossier harvests them, hashes them, attaches each to a step, and determines the state: pure, with no override.
Declared or derived, never assumed
Each step is declared by a person, derived from the artifacts, or pending. The model-risk grade is derived from the two declared grades and has no input of its own.
The reconstruction view
One row per model call: the model id from the ledger, the prompt hash, the token counts, opening the archived bytes read-only.
Determinations are append-only
An adequacy determination is a named person’s declaration with a timestamp. It is never edited; a later one is added beside it. A blocked dossier accepts only “not adequate”.
Three renderings of one record
JSON is canonical and content-hashed; Markdown and HTML are renderings of it, and the HTML carries no script and no external reference.
Hashes in the record, bytes in the archive
Every evidence item is a kind, a SHA-256 and a locator. The state is decided in order: any item whose model claim is not ledger-backed, or whose hash is unknown, blocks. Then incomplete, if a planned kind has no evidence or no determination exists. Complete otherwise. An empty plan leaves plan_executed pending, not vacuously satisfied.
- DeclareQuestion of interest, decision influenced, the two grades, what is out of scope.Person
- HarvestThe run directory or the review, read as it is; nothing re-generated.Engine
- BackEach model claim checked against the invocation ledger.Rules
- DetermineBlocked, incomplete or complete: pure, in that order.Rules
- SignAdequate or not, by a named person, appended.Person
- RenderJSON canonical; Markdown and HTML derived; archived bytes readable.Engine
Recorded,
not reconstructed
| Who | Today | With Assurance |
|---|---|---|
| Quality · computerized-system validation | Validates the AI tooling with a procedure document and a screenshot, and hopes the version in the screenshot is the version that ran. | Opens the reconstruction view for the run in question: one row per model call, the answering model from the ledger, the prompt hash, and the archived bytes. |
| Regulatory affairs | Assembles a credibility narrative after the fact, from whatever the team remembers about what was asked. | Declares the Context of Use up front, and receives the seven steps, each marked declared, derived or pending, with every artifact hashed. |
| IT validation · audit | Asks the vendor what the model was asked and is shown a log the vendor can edit. | Reads an append-only, content-addressed archive whose only join key is the hash already in the record. |
| Program leadership | Signs off on model-derived evidence on the strength of a slide. | Signs an adequacy determination against a dossier that says, in fixed text, exactly what it does not establish. |
What it does not do
“An index of evidence is not an approval.”
Fixed text on every dossier: it does not establish that the output is clinically valid, that the declared Context of Use is appropriate, or that a regulator will accept it.
- Adequacy is a person’s declarationThe dossier records it and never makes it.
- Reports rendered on demand are marked as not citedThe design-risk report and the eval report are built on read and not stored, so a dossier says exactly that rather than hashing a file that does not exist.
- No “adequate” button on a blocked dossierNot disabled: absent. A test asserts the option is not in the page.
- It records what Antevra didA model used elsewhere in a program is outside the ledger and outside the dossier.