Memory Room — Conformance Bar
Version. 0.2 (was v0.1, four checks; v0.2 adds check 5 + held-question) · Date. 2026-06-29
Status. Active standard for the Memory room conformance walk. Supersedes the implicit four-check bar.
Origin. Agent/boundary discussion, 2026-06-29 — the bar was drawn before the agent-boundary lens existed; this revision derives the missing check from the seed's Operator-authority commitment rather than from bugs-found.
The bar (a room is conformant only when all five hold)
1. Does-its-job — accumulates knowledge with provenance; never forgets; append-only.
2. Add — content can enter the room and reach durable commit.
3. Invoke — the Operator can drive the room's motions through the Companion (its actual interface), not only via direct API.
4. Correct-while-preserving — a correction supersedes (new version, wasRevisionOf), the prior remains visible; trajectory walkable.
5. Actor-boundary-completeness (added v0.2) — wrong-actor boundaries are structural throughout the room, not merely at the one commit/discard gate CR-2026-120 fixed. The seed's Operator-authority principle ("the machine surfaces and signals; the Operator approves") must be enforced in code at every point in the room where a non-human actor could take a human-authority action — not held by convention anywhere.
Held-question against Room 1 (recorded, not yet actioned)
Monitor-observability. Are Memory's consequential actions (add / commit / discard / revise) emitted such that a future boundary-monitor could watch them — not merely recorded for audit, but observable to a monitor evaluating them against a limit?
Dependency: this only matters if the boundaries thesis holds (the keystone test — "decisions are asset-transfers FORAY can monitor" — is parked pending Room 1 closure). Record the dependency; do not act until the keystone is tested. If the keystone holds, the boundaries arc revisits Memory against this question. If Memory's actions turn out already-observable, the question closes cheaply.
Why check 5 was added (trajectory preserved)
The original four-check bar was descriptive of failures found by walking, not derived from what the seed requires. CR-2026-120 made one actor-boundary structural (companion cannot commit/discard) — but that was found via a Step-0 that happened to surface the gate hole, not via a systematic audit. The agent/boundary discussion surfaced that the seed's Operator-authority principle commits Memory to actor-boundary enforcement everywhere, and we had enforced it once. Check 5 finishes what CR-2026-120 started: a focused audit for other unenforced actor-boundaries in the room.
Discipline applied: check 5 earns its place because it finishes an existing seed commitment (not future-proofing). The monitor-observability question did NOT earn a place in the bar — it depends on an untested thesis — so it is carried as a held-question, not a check.
Consequence for "Finding 1 complete"
Closing Room 1 Finding 1 now requires, beyond CR-2026-120 landing and verifying:
- A focused actor-boundary sweep of the Memory room: where else can a non-human actor take a human-authority action, held only by convention? Any found are same-bug-class as CR-2026-120 and belong in its family (fix-now, not discover-later).
- The monitor-observability held-question recorded against the room for the boundaries arc.