Rules and Evidence Model
Ten rules
Each rule is a policy. The table separates what code enforces, what is guidance, and what is not implemented, so no rule is overstated.
| Rule | Policy | Enforced by code | Guidance only | Tested | Not implemented |
|---|---|---|---|---|---|
| INT-001 | Human approval | approve/attest need reviewer, note, exact revision, --confirm-human | who counts as a legitimate approver | yes | verified identity |
| INT-002 | Revision-bound evidence | link rejects stale revision; old links kept as history; AUD-05 | migration of evidence across revisions | yes | automated migration |
| INT-003 | No completion by proxy | no completed state; attest needs four evidence types; AUD-04, AUD-11 | wording in downstream tools | yes | independent execution proof |
| INT-004 | Untrusted content boundary | sources labelled untrusted; text never executed; size caps | agent behavior in IDEs | partly | host-level sandboxing |
| INT-005 | Data minimization | local only; heuristic redaction | what teams choose to import | partly | DLP-grade detection |
| INT-006 | Read-only default | API is GET-only; CLI writes only .intentsutra/ and chosen report dirs | connector write policy | yes | connectors |
| INT-007 | Provenance | SHA-256 of sources and evidence; provenance labels; AUD-06, AUD-09 | reviewer discipline | yes | signatures |
| INT-008 | Unknown stays unknown | questions; [TO CONFIRM]; unknown outcomes | stakeholder follow-up | yes | semantic gap detection |
| INT-009 | Independent verification | audit re-derives state from raw data and cross-checks the log | an auditor outside the team | yes | verified test execution (M1) |
| INT-010 | Truthful publishing | docs state limits; labels in reports | release review | partly | automated claim linting |
Evidence kinds
| Kind | Recorded | Hashed | What it can support |
|---|---|---|---|
code | workspace file | yes | "this file existed and had these bytes" |
test | workspace file | yes | "a test file exists" (not that it passed) |
test-run | workspace file + outcome | yes | "a human imported this result" (not independently verified) |
release | workspace file + outcome | yes | "a human imported this release claim" |
ticket | reference text | no | a pointer |
pr | reference text | no | a pointer |
Trust labels
Future schemas separate observed-file, imported-claim, self-attested, measured-test, verified-external-signature, unknown, not-run and stale. This release implements the first three plus stale; measured-test and verified-external-signature are planned and are never shown because the code cannot substantiate them.
Stale evidence
When a revision changes canonical content (title, actor, trigger, behavior, constraints, acceptance), older approvals, links and attestations remain in the record for that older revision and no longer count. Reports show the stale count. Nothing is migrated silently.
The attestation gate
attest succeeds only when, for the current revision, there is: an approval, an observed code link with matching hash, an imported passed test-run, and an imported released release claim. The reviewer then confirms by hand. The state name human-attested-evidence-linked deliberately avoids "complete".
Independent audit checks
| ID | Check | Severity |
|---|---|---|
| AUD-01 | audit hash chain intact | critical |
| AUD-02 | content hashes match stored fields | high |
| AUD-03 | evidence files still match recorded hashes | high |
| AUD-04 | no false completion (attested state re-derived) | critical |
| AUD-05 | approvals bound to existing revisions | high |
| AUD-06 | imported claims keep "not independently verified" label | high |
| AUD-07 | approvals never claim verified identity | high |
| AUD-08 | registry state is explained by audit events | critical |
| AUD-09 | sources untrusted-labelled and hash-pinned | medium |
| AUD-10 | stale evidence visible (informational) | info |
| AUD-11 | proof-by-proxy gaps (test file without pass, PR without code, approval with open questions) | medium |