ADR-0064 — The verdict gets a report an auditor reads alone

Date: 2026-09-01 Status: proposed Deciders: the maintainer Method: measured against what runward emits today, and against what the category emits, sourced

Context

runward's artifacts target a regulated reader. Its ergonomics target an engineer. That gap is the subject here, and it is not a polish problem.

What runward emits today, and who it is for. check --sarif renders into a forge's security view. compliance <regime> writes three readiness drafts — ISO 42001, NIST AI RMF, EU AI Act — plus an OSCAL component definition. check --attest and bundle produce in-toto statements. Every one of those is a compliance surface. But runward does not only judge compliance: it judges DELIVERY — which phases were crossed, which decisions were traced to resolving evidence, what is deferred and why, what the gate deliberately does not cover. That verdict has no reader-facing form at all. It exists as a terminal rendering for whoever ran the command, and as --json for whoever writes code against it.

So the person who most needs the verdict — an assessor arriving months later, on the repository alone — is the one person the product currently asks to open a terminal.

What the category emits. Measured 2026-09-01 against a comparative analysis of the three spec-driven frameworks runward is positioned beside: "None of these frameworks are described as producing auditor-ready compliance reports, shipping verdicts, or governance dashboards." Spec Kit enforces human approval at phase gates and produces a constitution, specs, plans and tasks. OpenSpec produces a change archive and claims traceability by design through its proposal → spec delta → task → code chain — which is a HISTORY, not a verdict. BMAD produces role-specific documents for a product owner to read. None produces an object addressed to a reader who was not in the session.

Is that immaturity? Partly, and it would be lazy to stop there. The category is roughly a year old, and features arrive. But the deeper reason is an ORIENTATION: these tools optimise for the agent's next step, and their natural output is whatever the agent consumes next. A report for an assessor has a different addressee and a different lifetime — it must survive the session, the agent, and the tool. Maturity adds features to a design; it does not hand a tool an addressee it was never pointed at. That orientation is precisely what the survival thesis (ADR-0052) says a harness-native gate cannot hold, and it applies to the reporting surface as much as to the verdict.

What the refusals actually forbid. ADR-0054's five crossings constrain the VERDICT PATH: no network endpoint SERVING rules or verdicts, no long-lived process BETWEEN invocations, no key or state the operator does not hold, no reading beyond the tree judged, no network or LLM IN THE VERDICT PATH. None of them forbids the CLI from EMITTING a rendered artifact. A generated file is a build output, like the SARIF and the OSCAL already are — local, deterministic, owned by the operator, surviving the tool that made it. Reading the refusals as a ban on any interface is a misreading, and it has cost the product a reader.

Decision

Proposed: runward report emits ONE self-contained file that an assessor opens without a terminal, without an account, and without runward installed — addressed to the auditor, not to the CTO and not to the agent.

Self-contained is the load-bearing word: no external stylesheet, no script fetching anything, no link that must resolve for the document to be readable. It is committed, attached to a release, or emailed, and it says the same thing in five years as it does today. It is generated FROM the verdict and the artifacts the mission already holds — it computes nothing new, so it cannot disagree with check.

What it must answer, because that is what an assessment asks:

1. Where is this mission — which phases are crossed, which is current, what blocks the next.
2. What is proven, and by what — each gated claim with the evidence it resolves to, by path.
3. What is deferred, and why — the declared horizon, stated as a floor and not as a ceiling.
4. What the gate does not cover — GATE_NON_SCOPE, verbatim, where a reader cannot miss it.
5. How to re-derive this alone — the command, the attestation, the fact that it needs no account.

Alternatives discarded

• A hosted dashboard. Refused by construction, not by priority: it recreates the dependency the survival thesis exists to remove — a verdict that dies with the service that showed it. ADR-0054 crossings 1 and 2.
• More forge connectors instead of a report. Worth doing and not a substitute. GitLab does ingest SARIF (artifacts:reports:sarif, with direct upload being made first-class), so the forge surface is wider than GitHub already. But a forge view shows FINDINGS; it does not show a delivery verdict, and it lives in the forge rather than in the repository. A separate decision, not this one.
• Extending the readiness drafts. They are addressed to a regime and shaped by it. Folding a delivery verdict into a regime pack would make the verdict look like a compliance claim, which is exactly the conflation this ADR exists to undo.
• A cross-repository portfolio view. Genuinely useful and genuinely a different product: a portfolio cannot be a per-repository artifact. It is the ADR-0039 satellite, with a different buyer, and conflating it with this decision would sink both.

Consequences

• Positive: the reader runward's artifacts are FOR can finally read them; the delivery verdict stops being invisible to everyone but the operator who ran the command; and nothing about the verdict path changes, so no moat property is spent.
• Negative, accepted: a rendered document is a surface with taste in it, and taste ages. It also becomes a thing to keep byte-stable, which means a golden and a place in the whole net.
• On other boundaries: it is a fifth emission beside SARIF, OSCAL, in-toto and the readiness drafts, and it inherits their discipline — deterministic, schema-checked where a schema exists, and refusing to state anything the verdict does not.

What would settle it

An assessor — or someone briefed to read as one — opens a generated report for a mission they have never seen, without a terminal, and answers the five questions above. If any answer requires opening the repository, the report has not replaced the terminal for its addressee and the shape is wrong.

The second half is cheaper and equally binding: the report must be byte-identical across two runs on one tree, and must state nothing check --json does not already carry. A rendering that computes is a second verdict, and two verdicts that can disagree are worse than one that is hard to read.

Reevaluation trigger (mandatory, dated)

An assessor reads a generated report and still asks for the repository, or a regime pack and the report state different things about one mission. Either signal means the addressee or the derivation is wrong, and reopens this.

Trigger set on: 2026-09-01 · Watched via: the acceptance above, run on the first real external reader — the ADR-0052 pilot, or the first assessment runward is put in front of