Get started

Three-provenance capability eligibility

Source

Capabilities and runtime speed constraints narrow which candidates may serve a workload before cost per successful unit is weighed, and each constraint is bound to the provenance class that can honestly answer it.

Why provenance classes exist

Not every fact about a model has the same standing. A provider or public catalog listing may declare that a model accepts tools or structured_outputs, while a customer's own schema still fails in 14% of calls. Conversely, a third-party speed monitor measures real tokens per second across tens of thousands of routed requests, which is stronger than a brochure and still different from the customer's own workload traffic.

Treating all three as interchangeable either admits vendor claims as if they were verified evaluations or forces operators to omit real eligibility requirements from the decision record. Vitna records all three and holds each to its own admissibility boundary:

Provenance classSourceConstraints answeredAdmissibility rule
declared_mirrorapp/data/catalog.json paired via platform/catalog/sku-map.mjsrequiresCapabilities, minContextTokensNarrows candidate eligibility only. Never crosses a quality or economic threshold and never satisfies requiresVerifiedCapabilities.
measured_third_partyapp/data/speed.json paired via sku-map.mjs + baseIdminThroughputToksSec, maxTtftMsEvaluates throughput and time-to-first-token bounds while naming the measurer, request sample count, and retrievedAt. Refuses when unmeasured.
measured_customerWorkload evaluation evidence (verifiedCapabilities, n > 0, sourceRef)requiresVerifiedCapabilitiesRequires a customer evaluation measurement with positive sample count n and non-empty sourceRef. Only provenance class allowed to satisfy requiresVerifiedCapabilities.

Canonical capability tokens

Both requiresCapabilities (declared_mirror) and requiresVerifiedCapabilities / verifiedCapabilities (measured_customer) take comma-separated values on the CLI or string arrays in JSON (vitna.stackfile.v1 and POST /api/v1/decide) from the canonical token set:

  • tools
  • structured_outputs
  • reasoning
  • vision
  • audio_input
  • video_input
  • file_input
  • image_output

Registering capability constraints and verified evidence

bash
# Narrow eligibility using declared catalog capabilities, context window, and third-party speed
npm run aperture:workload:edit -- convexity/invoice-extraction \
  --constraint requiresCapabilities=tools,structured_outputs,vision \
  --constraint minContextTokens=128000 \
  --constraint minThroughputToksSec=40 \
  --constraint maxTtftMs=1200

# Require customer-verified structured_outputs and record the evaluation that proved it
npm run aperture:workload:edit -- convexity/invoice-extraction \
  --constraint requiresVerifiedCapabilities=structured_outputs

npm run aperture:evidence:add -- \
  --workload convexity/invoice-extraction \
  --sku anthropic/claude-sonnet-4-6 \
  --kind eval \
  --label invoice-eval-2026-09 \
  --measured-at 2026-09-14 \
  --n 500 \
  --source-ref https://github.com/org/evals/commit/4a9c210 \
  --metric successRatePct=95.4 \
  --metric verifiedCapabilities=tools,structured_outputs

The same constraints travel cleanly in a declarative Stackfile under workloads[].constraints and in stateless checks via POST /api/v1/decide and MCP vitna_decide_stack.

Refusal and stored-brief reproducibility

Every capability check follows the engine's general eligibility rule: a requirement the record cannot check is not satisfied.

  • If a workload sets requiresCapabilities or minContextTokens and a candidate has no paired catalog mirror entry (for example, an unmapped SKU or private infrastructure), the candidate is refused with the missing source named.
  • If a workload sets requiresVerifiedCapabilities and a candidate has declared_mirror support in the catalog mirror with no customer verifiedCapabilities evidence row, the engine refuses eligibility and states explicitly that a vendor declaration cannot satisfy a customer-verified requirement.
  • Whenever a decision consults declared_mirror or measured_third_party records, the exact readings (listingId, retrievedAt, snapshot sha256, measurer, and requests) are frozen into decision.inputs.capabilityReadings and printed under ## What it read in the decision brief so stored briefs regenerate byte for byte even after daily mirror updates.