Fix Verification Profile v1 (FVP-1)

Status: draft · Version: 1.0.0-draft.1 · License: MIT, same as the project — copy it, fork it, implement it, no permission needed.

0. What a fix record claims, and what it does not

Read this section before any other. Everything below exists to keep this boundary intact.

A Verified Fix Record asserts exactly this:

These named findings were present in a scan of this code. After a change, a re-scan by the verifier no longer reported them, and these named commands were executed by the verifier, which observed these exit codes.

It does not assert, and MUST NOT be presented as asserting:

• that the code is free of vulnerabilities;
• that the change is correct, or that it preserves behavior;
• that the scanners were capable of finding the class of defect in question;
• that anything outside the recorded coverage was examined at all.

A record is evidence about a procedure that was carried out. It is not a safety certificate, and an implementation that markets it as one is not conformant with this profile regardless of how many assertions below it passes.

Scope

When an agent writes a repair for a security finding, someone has to decide whether the repair worked. In most tools that decision is made by asking the model that produced the repair — directly, or by parsing its own summary of what it did. That is the one question a model is structurally least able to answer against its own interest, and no amount of prompt wording changes it.

FVP-1 specifies an arrangement where the decision does not depend on the repairer at all, and where the resulting record can be re-derived by a third party who trusts neither the repairer nor the verifier's own report of itself.

This document specifies:

• the roles and the separation between them (§1–§2);
• what a conformant verifier MUST do before issuing a record (§3);
• the record format and its integrity properties (§4);
• how a record is re-derived offline (§5);
• coverage and the honesty rules that bound a claim (§6);
• what may be claimed at each conformance level (§7).

It does not specify a scanner, a rule format, a policy language, an audit format, or a programming interface. It constrains observable behavior only. An implementation with a completely different architecture can be conformant.

Audience

Authors of security-scanning and auto-remediation tools; agent runtime authors; security reviewers deciding how much weight to put on an automated fix; anyone building a release gate that agent-written code has to pass.

Conventions

The key words MUST, MUST NOT, SHOULD, SHOULD NOT, and MAY are to be interpreted as described in RFC 2119. Assertion IDs are stable across patch and minor versions of this profile; an assertion is never renumbered, only deprecated.

1. Terms

Term	Meaning
Executor	Whoever or whatever edited the code: an agent, a person, another tool.
Verifier	The component that issues the record. It re-scans, runs the checks, and decides the verdict.
Target	A finding the repair is meant to eliminate, identified by a fingerprint stable under small edits.
Check	A command belonging to the project — its tests, type checker, build — executed by the verifier.
Observed exit code	An exit code the verifier read from a process it started itself.
Coverage	What the scan actually examined, and what it did not.
Record	The issued artifact described in §4.

2. Separation of executor and verifier

FV-1. The verifier MUST determine the verdict without consulting the executor. Specifically, the executor's natural-language output, self-report, summary, or any structured claim it emits MUST NOT be an input to the verdict.

FV-2. The executor's identity MUST be recorded and MUST NOT affect the verdict. A record produced for a repair written by a person and one written by an agent are evaluated identically.

Rationale. If the identity of the repairer changed the standard of proof, the record would be measuring reputation rather than the repair. Recording the executor is still useful — a reviewer may want to know — so the requirement is that it be present and inert, not that it be absent.

FV-3. The verifier MUST start every check process itself and read the exit code from that process. Accepting an exit code, test result, or pass/fail verdict reported by another party does not satisfy this profile.

FV-4. Where the verifier and the executor are parts of one product, the implementation MUST document the boundary between them and MUST NOT allow the executor to configure, select, or suppress the checks the verifier runs.

3. What a verifier MUST do before issuing a record

FV-5. The verifier MUST re-scan the code after the change, using at least the engines that produced the targets.

FV-6. The verifier MUST determine, per target, whether it is still present, using a fingerprint that is not defeated by the finding moving within a file or by its message being reworded.

FV-7. The verifier MUST attempt the project's own checks and MUST record each one it ran with the command and the observed exit code.

FV-8. A project for which the verifier could execute no check MUST NOT receive a verified record. An unverifiable repair is not a verified one, and "there was nothing to run" is not a pass.

FV-9. Every check command MUST pass whatever authorization the implementation applies to other command execution. A verifier MUST NOT hold a privileged path that ordinary tool execution does not. A denied command MUST be recorded as a check that did not pass, never silently omitted.

FV-10. The verdict MUST be verified: false if any of the following hold: any target is still present; any executed check did not pass; no check was executed.

FV-10a. (v2) The verdict MUST also be verified: false if the repair introduced a finding at or above the gate's threshold that the originating scan did not report, or if the verifier did not determine what the repair introduced.

Rationale. A repair is a change, and a change does two things: it removes what it was aimed at, and it may add something that was not there. FV-10 only asks the first. A verifier that removes an injection and adds a different one satisfies FV-10 completely, which makes verified mean less than a reader would reasonably assume.

The second half — not determining it fails — matters as much as the first. Without it, a verifier that never looks scores better than one that looks and finds something, and a rule that rewards not looking is not a rule. It is the same principle §6 already applies to coverage: absent because unlooked-for is not absent.

FV-10b. (v2) The record MUST carry the complete set of introduced findings, not a set pre-filtered by severity, together with the threshold used to decide which of them blocked. A reader applying a stricter threshold MUST be able to reach its own conclusion from the record alone, without re-scanning.

Rationale. Store the observation, derive the judgement, and record the parameters of the derivation. This is the same rule the profile already applies to exit codes in §FV-6: what was observed is evidence, what it means is interpretation, and a record that keeps only the interpretation cannot be re-examined.

4. Record format and integrity

FV-11. A record MUST carry, at minimum:

Field	Meaning
schema	Identifies the format and version.
generatedAt	When the record was issued.
tool	The verifier's name and version.
executor	Who edited the code (§FV-2).
before	The originating scan: identifier, completion time, coverage, and the targets.
after	The verifying scan: identifier, completion time, coverage, and the targets still present. (v2) Also what the repair introduced, or an explicit "not determined".
gate	(v2) The threshold and mode the verdict was reached under.
checks	Each executed check: what was run, the observed exit code, whether it passed.
assurance	Whether checks were executed at all.
verdict	The boolean plus the reasons and caveats behind it.
outcome	(v2, optional) Which way the repair fell short: verified, target-remains, regressed, or unverifiable.
recordHash	Integrity over everything above.

A record MAY additionally carry the changed files and a hash over them, an anchor into a tamper-evident log, and the hash of the governing policy.

FV-12. recordHash MUST be computed over a canonical serialization of the record with recordHash itself excluded, such that two implementations serializing the same record in a different key order compute the same hash.

FV-12a. A verifier MUST re-derive a record under the rules of the version the record declares, not the newest version it knows. A record that stops verifying because the rules changed after it was issued would make offline re-derivation conditional on the reader's build date, which §5 exists to prevent. New rules go in a new version; old versions stay readable, and a renderer SHOULD say which questions the older version did not ask.

FV-13. The verdict MUST be derivable from the rest of the record by the rules in §FV-10. It is stored for convenience, not as an independent input.

Rationale. Storing a verdict that cannot be recomputed would make the record only as trustworthy as its issuer's honesty at the moment of issue. Because it is derived, a record whose verdict was edited and whose hash was then recomputed to match still fails re-derivation. Both defenses are required; neither is sufficient alone.

FV-14. A verifier MUST issue a record whether the verdict is positive or negative. Issuing records only on success turns the artifact into an award and makes its absence unreadable.

FV-15. A record MUST NOT contain credentials, tokens, or the contents of scanned files. It names locations and rules; it is not a copy of the code.

5. Offline re-derivation

FV-16. An implementation MUST provide a way to re-derive a record that requires no access to the workspace, no network, and no state held by the verifier.

FV-17. Re-derivation MUST recompute recordHash and MUST re-derive the verdict per §FV-13, and MUST fail if either disagrees with the record.

FV-18. Re-derivation answers whether the record is sound and unmodified. It does not answer whether the repair is still good; an implementation MUST NOT present a successful re-derivation as a current statement about the code.

FV-19. A re-derivation failure MUST be distinguishable, by exit status or equivalent, from a malformed input and from an internal error. A pipeline has to tell "this record does not hold up" from "you pointed me at the wrong file".

6. Coverage and honesty rules

FV-20. Both scans in a record MUST carry a coverage status of complete, partial, or unknown.

FV-21. Coverage MUST be partial when any selected engine did not complete, when results were discarded, or when the scan was narrowed to a subset of the code. It MUST be unknown when the implementation cannot establish what was examined — including for records written before it tracked coverage.

FV-22. A finding absent from a scan whose producing engine did not complete MUST NOT be reported as resolved. It MUST be reported as unknown.

Rationale. This is the most commonly violated rule in existing tools, and the most damaging, because it silently converts "we did not look" into "it is fixed" — in the direction that makes the tool look effective.

FV-23. Coverage below complete MUST be surfaced in the record's reasons. It does not by itself invalidate a passing verdict — a check that passed did pass — but a record MUST NOT be presented without it.

FV-24. An implementation MUST NOT present a fix record as a statement that the code is free of vulnerabilities (§0).

7. Conformance levels

Level 1 — Independent verification. FV-1 through FV-15 and FV-20 through FV-24. Permits: "repairs are verified independently of whoever wrote them."

Level 2 — Re-derivable evidence. Level 1 plus FV-16 through FV-19. Permits: "the verification can be re-derived by a third party offline."

An implementation claiming a level MUST publish which assertions it satisfies and MUST NOT claim a level on the strength of a subset.

8. Reference implementation

DvalinCode implements this profile. The verifier is src/remediation/verify.ts (observed exit codes, policy-gated commands) driven through src/security/verifyRun.ts; the record format and its derivation rules are in src/security/fixRecord.ts; coverage and the resolved/unknown distinction are in src/security/contracts.ts.

Offline re-derivation is dvalin verify-fix <record.json>, and the same check is exposed to other agents as the dvalin_verify_fix MCP tool. This is published as a profile rather than as a feature so that it can be held against DvalinCode too: an assertion this implementation fails is a bug in this implementation, not an amendment to the profile.