# Measured Claims This is the **canonical claim boundary** for Persona Harness (PH). The README and [CONTRIBUTING](../CONTRIBUTING.md) summarize and link here; they do not restate it. If a claim is not on this page with evidence, PH does not make it. Persona Harness only claims what measurement supports. When a measurement is negative, PH records it and keeps the feature default-off or parked — it does not quietly drop the result. ## Claim levels (summary) A claim may only assert what its evidence supports. Climb in order: 1. Surface exists → 2. PH invokes it in a fixture → 3. PH generates the evidence (not hand-written) → 4. Adversarial inputs fail honestly → 5. External smoke reproduces from a fresh tarball/npm → 6. Repeated A/B improves a named scenario → 7. Only now may a scoped claim be written. Full detail and the reject rule live in [CONTRIBUTING → The Claim Ladder](../CONTRIBUTING.md#the-claim-ladder). ## Supported claims Each supported claim is scoped to bounded local fixtures and is a **completion-integrity** claim, not an app-quality claim. P3 held Stable/GA and npm `latest` movement from the 2026-07-12 local production audit, which found stronger completion-integrity bypasses outside these bounded fixtures. The condition that hold named — trusted external signed CI/verifier attestation — was demonstrated on 2026-08-09 and is recorded below. ### 1. PH blocks missing completion evidence > PH can block completion when required reports, evidence, or test results are > missing for explicitly defined workflow gates. Boundary: this does not prove generated app quality or strong anti-forgery integrity. ### 2. Forged TDD evidence is rejected > PH ignores a simple hand-written/forged TDD evidence file in the measured > completion-integrity fixture; `workflow finish` still exits non-zero. Boundary: this does not prove full TDD sufficiency or rejection of every internally consistent forged evidence chain. P3 is the accepted remediation program for that gap. ### 3. Green-only completion is blocked when the TDD rail is enabled > In the measured completion-integrity fixture, with the TDD rail on, PH blocks > green-only completion that has no red-first evidence (measured 5/5 vs 5/5 > allowed with it off). Boundary: PH does not write tests for you and does not prove test quality. ### 4. Compile errors are not accepted as red evidence > PH does not treat a compile failure as valid red TDD evidence in the measured > fixture. Boundary: this is an evidence-integrity claim, not a testing-framework claim. ## Negative / parked ### P3 completion-integrity hold A local production audit dated 2026-07-12 against `persona-harness@0.7.0-rc.2` recorded bypasses and safety gaps that block Stable/GA and npm `latest` movement until P3 closes. Local evidence paths: - `/Users/yongtae/Downloads/persona-harness-production-audit.md` - `/Users/yongtae/Downloads/persona-harness-audit-execution-log.txt` Accepted boundary: - Unsigned project-local JSON can remain diagnostics/audit material, but must not be finish authority after P3. - Local fresh fixed-command verification is cooperative-project integrity only. - Strong integrity/GA claims require trusted external signed CI/verifier attestation. **Demonstrated 2026-08-09** — see below. - Current RC evidence is not generated-app certification, product efficacy, broad reliability, security certification, token-saving evidence, or a strong anti-forgery guarantee. See [`current/p3-integrity-roadmap.md`](current/p3-integrity-roadmap.md). #### Level 2 external attestation, demonstrated 2026-08-09 The boundary the hold names as a precondition for a strong integrity or GA claim was exercised end to end, on `persona-harness@0.8.0-beta.34` installed from the registry, against a separate public consumer repository producing its own GitHub-signed attestations. > An enrolled external repository signed a receipt over its own source in GitHub > Actions; the consumer fetched it, consumed it once, and the replay was > refused. Refusals were exercised individually and each reported a distinct diagnostic: wrong repository, wrong workflow, wrong ref, excluded pull-request path, three malformed-archive shapes, expiry, source drift, replay, all-skipped tests, and a severed network. A real signature from an unenrolled repository was refused on identity rather than on cryptography. Bounded plainly: - **Level 2 is now evidenced. The remaining P3 items are not.** This records one boundary, not the hold's closure. - The excluded **fork event** was not exercised — it needs a push from an account outside the project — and is reached only indirectly. - One operator, one day, one project shape. Evidence and the lines that were **not** covered are on issue #116. ### Runtime injection Measured in accepted 10-pair local-current OpenCode fixtures. Both PH ON and PH OFF succeeded 10/10, but PH ON increased provider-token total, read chars, tool calls, and elapsed time in **all 10 pairs**. See [`current/injection-value-status.json`](current/injection-value-status.json). - Allowed: runtime guidance exists as an explicit **default-off, opt-in preview**. - Forbidden: that runtime injection improves generated code quality, saves tokens, or improves broad product efficacy. ## Preview / not proven ### Role Checklist Relay - Allowed: PH includes a Role Checklist Relay preview for role artifacts and gate-oriented workflow guidance; role boundaries are report-only/heuristic. - Forbidden: reliable automatic multi-agent orchestration, production-ready delegation, or deterministic role enforcement. ### Workflow loop (Ralph-loop-style continuation) - Allowed: `ph workflow loop` is an explicit capped fresh-session blocker loop command with fixture-scoped evidence. It is **not** a default hook. - Forbidden: autonomous completion, closure guarantee, broad reliability, or token saving. ### CodeGraph / LSP - Allowed: CodeGraph and LSP wrappers are optional preview integrations that report an unavailable status when their external tools are missing. - Forbidden: default effectiveness, broad navigation benefit, or being a required product path. ## Forbidden claims | Claim | Status | Why | | :--- | :--- | :--- | | token saving | forbidden | runtime injection data is negative / not proven | | generated app quality | forbidden | PH checks rail/evidence, not app quality | | full TDD framework | forbidden | PH gates evidence, not test sufficiency | | broad AST/linter enforcement | forbidden | checks are scoped/report/closure surfaces | | production-ready multi-agent orchestration | forbidden | relay is a checklist preview | | closure guarantee | forbidden | workflow loop has fixture-scoped evidence only | | runtime injection benefit | forbidden | measured negative | ## How to add a new claim 1. Add the measurement first. 2. Link the evidence (archive, fixture, or status file). 3. State the scope (which fixture/model/conditions). 4. State the limitation. 5. If it is not supported, keep the claim out of the README.