40points
Implementation evidence
A matching implementation declaration exists. A schema alone is insufficient.
Evidence has a score. Approval has a reviewer.
SofiCore parses source, retrieves candidate matches and applies explicit review rules. The score describes supporting evidence; it is never a probability that the product is correct.
Loading the saved review record…
Reading the saved review status and any recorded response. This does not start a model request.
This recorded review is separate from the live workspace’s optional model setting. No completed result is assumed from configuration.
THE ANALYSIS PATH
From a fixed revision to a traceable record.
- 01
Capture
Resolve the GitHub revision and retrieve bounded source files. Oversized or incomplete scans are rejected.
- 02
Parse
Extract supported declarations, imports and active named tests with the TypeScript compiler. Repository code is never executed.
- 03
Match
Compare requirement phrases with parsed identifiers and test names. Comments, arbitrary prose and imports alone cannot supply implementation evidence.
- 04
Explain
Retrieve analysis records with their file paths and citations. Keep the acceptance decision separate from the explanation.
THE EVIDENCE INVENTORYFour inputs.
Four inputs.
Up to 100 points.
Each requirement receives a coverage score. Project coverage is the rounded average across requirements.
The Nano ID study illustrates the distinction: 100% evidence coverage, zero verified requirements. Every acceptance mapping remains unreviewed.
Read the study findings20points
Acceptance criteria mapped
Points are proportional to criteria with candidate matches. Mapping does not mean reviewer confirmation.
25points
Named tests present
A matching active named test declaration exists in the captured source.
15points
Exact-commit CI passing
The GitHub adapter records passing Actions for the analyzed commit.
A SEPARATE REVIEW GATE
When does “verified” apply?
All four conditions must hold for the analyzed revision. “Verified” means reviewed evidence plus commit-level CI; it does not independently certify runtime behavior.
- Implementation evidence is present.
- Every defined criterion has matching evidence and explicit owner confirmation.
- A matching active named test is present.
- GitHub Actions passed for the exact analyzed commit.
A critical requirement without implementation or test evidence, or failed CI, blocks release. Unknown CI or pending acceptance review keeps it under review. Review flags are cleared when the repository revision changes or requirement definitions are edited in the workspace.
EXPLANATION MODES
The same records. An explicit choice.
DEFAULT
Deterministic evidence retrieval
Copilot answers from saved analysis records and links to their evidence. It does not call a language model. Interactive public Copilot uses this mode; the Azure review above displays a separate saved record.
CONFIGURATION + OPT-IN
Optional model explanation
In a signed-in workspace, a configured provider can receive the question and a bounded set of source evidence after explicit opt-in. Invalid citations or provider failures fall back to evidence mode. Viewing the recorded review does not enable this setting or send a new request. Model prose still needs review.
Keep the proof in its proper scope.
On 28 Oct 2026, all 42 application tests passed: 20 engine and 22 server checks. Azure adapter tests use mocked provider responses; they do not establish a live model request or hosted integration. Inspect the saved review above for the model and its recorded outcome.
Nano ID source hashes and recorded exact-commit CI are inspectable. The 55 indexed tests are named declarations; SofiCore did not execute upstream code or collect individual test results.
The saved Azure record reports its own request outcomes, usage and citation checks. It does not establish hosted sign-in, live repository import, GitHub App authorization or live workspace model configuration. Those require separate deployment checks. The documented AWS target is not provisioned.