Identity
Suite bytes, case trees, exact commits, per-case bundle sources, runtime adapters, the suite schema, and release locks.
Architecture ยท Evidence before claims
A pipeline where each authority stays in its lane.
Skivolve separates source identity, agent generation, objective verification, comparator judgment, provider capability, spend accounting, and holdout authority. A result advances only when the relevant layer has evidence.
Evaluation flow
From exact source to reviewable artifact.
Strict manifest
Duplicate-key rejection and schema validation establish the suite, cases, variants, modes, providers, limits, and artifacts.
Git-bound variants
Control, candidate, and no-skill arms resolve to exact, non-identical per-case sources before dispatch.
Provider adapter
A reviewed capability registry determines roles, concurrency, sandbox, billing evidence, provenance, and authority scope.
Isolated workspace
The harness works on a fixture copy while source, verifier resources, credentials, and unrelated host paths remain separated.
Objective oracle
A bounded case verifier evaluates declared behavior against the produced workspace diff, final text, or final JSON.
Eligible comparison
Where the suite permits judgment, a locked calibrated profile compares anonymized AB and BA presentations.
Evidence gates
Source drift, protocol locks, artifacts, spend, provider provenance, holdout consumption, and certification close independently.
Authority map
No output gets to authorize itself.
| Question | Authority | What cannot answer it |
|---|---|---|
| What source ran? | Git object and canonical source fingerprint | Prompt text or provider response |
| Did behavior pass? | Case-specific objective verifier | Compilation alone or model self-report |
| May ambiguity be judged? | Suite mode and comparator profile | Provider name or result field |
| May this provider act? | Immutable reviewed capability registry | Manifest-supplied arbitrary imports |
| Was spend bounded? | Dispatch journal and configured ceilings | Successful calls alone |
| Can this support a release? | Private frozen suite, independent review, consumed plan, and required certification | Public validation scores |
Reproducibility
The run records more than a score.
Execution
Provider and model identity, requests, dispatch state, time and spend ceilings, output artifact, and verifier outcome.
Decision
Objective evidence, eligible comparator evidence, AB/BA ordering, calibration support, aggregation, and gate status.
Current contract
Suite schema v1 makes every evaluation boundary explicit.
Skivolve accepts suite schema v1. It requires the evaluation mode, provider adapters, shared verifier boundary, release comparisons, per-case bundle source, and per-case artifact contract instead of deriving them from repository layout or names.
Each case declares a canonical workspace diff, final text, or final JSON artifact. The parser and JSON schema enforce the same closed contract, and unsupported input fails before provider dispatch.