runward

RW™ · V0.43.0

Docs · Décisions · ADR-0082

ADR-0082.

ADR-0082-an-agent-may-ratify-under-its-own-name-never-under-a-humans.md

ADR-0082 — An agent may ratify, under its own name, never under a human's

Date: 2026-09-27 Status: accepted 2026-09-28 — the recommended direction, chosen by the maintainer, who opened the question on 2026-09-27 (« ça peut très bien être un autre agent IA qui va ratifier aussi, pas forcément un humain » [it may very well be another AI agent that ratifies too, not necessarily a human]) Deciders: the maintainer Method: the current doctrine read against the code; the ratification of runward's own mission on 2026-09-27 as the one measured session; the forges' own answer to the same question

Context

What the product says today. A ratification is the operator answering displayed evidence (ADR-0066 decision 4). ratify refuses a terminal-less run; the only way through without one is --attest-blind, recorded mode: BLIND. The charter tells agents to propose and never to run ratify (ADR-0066 decision 3, AGENTS.md). Under the regulated tier a BLIND ratification does not count (ADR-0080), so an agent working a regulated mission has no path at all: the gate's Next line sends it to ratify --decided, which refuses it, and --attest-blind leaves the rows counted.

What the product cannot tell. A ### Ratification line typed by hand, or ratify driven by a script through a pseudo-terminal, produces the same line-by-line trace as a person at the keyboard (RWD-2026-0119). So an agent can already ratify: it just has to do it under a human's name, which is the one outcome every other rule here exists to prevent. Forbidding the gesture does not stop it; it only removes the honest way to make it.

What the maintainer said. The reader of a refusal, and the one who ratifies, may be another agent, not necessarily a human. The CLI audit of 2026-09-27 found every surface written for a human at a terminal, and the agent path ending in a refusal or in BLIND.

What the forges did with the same question. GitHub's answer to "may an AI approval count" moved on 2026-09-01: « When enabled, Copilot can submit an approval that counts toward the repository's required-approvals rule. » It is off by default (« By default, Copilot will not approve pull requests »), administrators enable it at enterprise, organisation or repository level, and the review is Copilot's own, never a person's. Its separate rule still holds: the person who asked the cloud agent for the change cannot approve it. The pattern is: allowed, labelled, off by default, and never by the party that did the work.

Decision

An agent may ratify, as itself. ratify --agent <name> runs without a terminal, resolves and records the same evidence a person would be shown, and writes the trace with mode: agent and by: <name> (declared, agent). It never writes a human's name, and ratify without --agent keeps refusing a terminal-less run.

  1. Labelled everywhere. The ledger counts agent ratifications apart (ratification.agent), and check, the JSON, the SARIF and the attestation say how many rows an agent ratified. A reader never has to infer it.
  2. Never the party that proposed. An agent ratification of a row whose proposer segment names the same agent is refused. Both names are declared, so this stops the honest mistake, not the liar (RWD-2026-0119 applies unchanged).
  3. Counts by default, not under the regulated tier. In a default mission an agent ratification is a ratification, disclosed. Under the regulated tier it counts only if the mission's lock also declares "agentRatification": true: the organisation's explicit choice, off by default, the forge's shape. Without it, an agent-ratified row stays a strict gap there, named as such.
  4. The charter follows. AGENTS.md and ADR-0066 decision 3 change from "never run ratify" to "never ratify under a human's name; ratify only with --agent, only rows you did not propose".

Amended 2026-09-28: the trace names the person accountable for the agent. Chosen by the maintainer after a sourced check of how agent identity works where it already exists. Every forge and directory that gives an agent its own identity also records a person answerable for it: the GitHub cloud agent authors its commits and marks the requesting developer as co-author, GitLab's composite identity puts a service account and « The person who initiated the request » in one token, for Microsoft Entra « At least one sponsor is required for each agent identity », and OAuth token exchange (RFC 8693) keeps the agent as actor beside the person as subject. None has an agent with no one behind it, and none lets an agent act under a person's credentials. So:

  • The trace carries two names. ratify --agent <name> --for <person> writes by: <name> (declared, agent) and for: <person> (declared, accountable); --for is mandatory. Both are declared; the forge alone proves accounts (ADR-0080).
  • Decision 2, extended. Refused when the agent, or its accountable person, is the row's proposer.
  • Decision 3, extended. Under the regulated tier, agentRatification counts an agent ratification only when its accountable person differs from the proposer. The independence DORA's RTS 2024/1774 art. 17(1)(b) asks of the « functions » that approve and implement is, read here, that of the person answerable for the agent, not of the agent (no text says so; it is this ADR's reading). It never relaxes ADR-0080 part 2: the forge approval still requires a human account, not a bot or an app.
  • Consequences, added. ANSSI's PA-102 recommendation R30 asks that a human regularly review AI-generated code; for an organisation that follows it, the human-sample alternative discarded below stays the documented choice. The AI Act's human oversight (art. 14) is exercised « by natural persons » on a high-risk system in use; it does not address the tools that built it, and this ADR does not claim otherwise.

Alternatives discarded

  • Keep "agents never ratify". It holds only on paper: the gesture is already possible under a human's name, and an agent on a regulated mission is left with a Next line it cannot follow.
  • Let an agent ratify as the operator, by delegation. A delegation the trace does not show is the false record this ADR exists to prevent.
  • Agent ratification with a mandatory human sample. Heavier, and it needs a second, human gesture per session; it is the natural next step if agent ratifications turn out to be caught wrong.

Consequences

  • Positive. An agent-built mission gets an honest path to green, and the attestation says which rows a person answered and which an agent did, the question an assessor asks first.
  • Negative, accepted. A mission may cross the default gate with no human answer at all; the label is the only safeguard there, and the regulated tier stays the place where that is refused.
  • On other boundaries. No model call enters runward: the agent runs ratify, runward does not run an agent. The verdict path is unchanged except for one more disclosed counter.

What would settle it

Chosen on 2026-09-28. What would reverse it, on use: whether agent ratifications are later found wrong more often than human ones on the same missions (the ADR-0052 pilot is the first place to count it), and whether a regulated organisation accepts agentRatification or asks for the human sample instead.

Reevaluation trigger (mandatory, dated)

Reopen if a forge or a supervisor states whether an AI approval may stand for a human one in change management, or if agent-ratified rows are caught wrong where human-ratified ones are not.

Trigger set on: 2026-09-27 · Watched via: the ratification.agent counter on this repository's mission and on the pilot; forge changelogs.

References

← Tous les ADR