Version. v0.2 Date. 2026-06-25 (v0.1) · amended 2026-06-25 (v0.2) Surfaced by. CR-2026-122 Step 0 (S0-4), grounding the conversational-fidelity fix.
> Amendment — v0.2 (2026-06-25). The v0.1 finding below is unchanged and preserved. v0.2 adds the result of the follow-on investigation the v0.1 note called for: the structural fact (second voice, outside the persona) is confirmed, but the CR-2026-122 state-fidelity defect is not reachable on this voice. The "could occur independently" concern is, for that specific defect, not realized. A separate, low-priority, unevidenced hypothetical is recorded. See "Investigation result (v0.2)" at the end.
Finding. The Companion has at least two reply-composing conversational surfaces governed by different prompt assets:
operator_converse general conversation → assemble_prompt → companion_persona.md. This is the surface CR-2026-122 fixes.POST /engagements/{eid}/seed/converse (seed_conversation.py) → composes replies from CREATION_CONVERSATION_PROMPT, NOT companion_persona.md.Consequence. The CR-2026-122 persona state-fidelity mapping (held draft is "waiting," never "saved"/"in the record") does not reach the seed-conversation voice. If that voice describes held/committed state, it may carry the same state-fidelity gap independently, unfixed by CR-2026-122.
Status (at v0.1). Out of scope for CR-2026-122 (whose observed defect was on operator_converse general conversation). Recorded as a follow-on to check: does CREATION_CONVERSATION_PROMPT make any held-vs-committed / saved-vs-waiting state claim, and if so, does it need the same state-to-word mapping? Not yet investigated.
Related. Same failure shape as CR-2026-121 item #8 — a fidelity guarantee applied to the surfaces in view while a parallel surface ran its own vocabulary unchecked. Worth a general sweep someday for all conversational reply-composing surfaces and whether each is under a state-fidelity rule.
The follow-on the v0.1 note called for was run as a read-only investigation. Result: the structural fact holds, but the CR-2026-122 defect is not reachable on this voice. No fix needed for that defect.
Structural fact — confirmed. The seed-conversation voice is genuinely separate from the persona:
seed_conversation.py:639 sets the system prompt to CREATION_CONVERSATION_PROMPT (the asset at creation_conversation.py:26–197); it does not load companion_persona.md and does not call assemble_prompt._call_llm (seed_conversation.py:197), not the operator_converse responder's generate_response / _call_responder_llm.The CR-2026-122 defect cannot occur here — two independent reasons.
in_progress | ready_for_review only — there is no "saved"/"committed"/"created" status in the flow. There are no held assertions, no Memory writes, and no commit transition on this surface. The held-vs-committed distinction CR-2026-122 fixed concerns assertions in an existing engagement's Memory — a different object that does not exist in the creation conversation.CREATION_CONVERSATION_PROMPT offers no saved/committed vocabulary at all and frames the review moment as a brief "under review" ("take a look at what we have so far — does this capture what you're trying to do?"). A scan found zero occurrences of held / committed / retracted / saved / in-the-record / on-the-record / waiting / stored / persisted in the prompt. Its own "never do" list bans methodology nouns (seed, assertion, engagement, induction, manifestation, shaping, rendering) and does not even include held/committed — because the voice never goes near them.
The actual "project created/committed" acknowledgment is elsewhere and already covered. It lives on the operator_converse intents _route_create_project / _route_finalize_project / _route_commit_project_draft (router.py:1290/1398/1615), which are persona-covered and server-composed (delegated_response). Not on this voice.
Net. The second voice is real and outside the persona fix (v0.1 correct on the structure). On inspection it does not carry the CR-2026-122 state-fidelity gap: its surface has no held/committed object to misreport, and unlike the persona it contains no state-blind word-to-meaning mis-mapping. The "could occur independently" concern in v0.1 is, for the CR-2026-122 defect specifically, not realized. This follow-on is closed.
Distinct from the CR-2026-122 defect, the investigation noted one generic-LLM-honesty risk worth recording so it is not rediscovered as if new: at ready_for_review, an LLM could in principle volunteer "I've created / saved your project" before anything is instantiated. This is a premature-"created"-claim about a brief, not a held-vs-committed-assertion misreport — a different concern. The prompt neither directs nor forbids it, gives it no "saved" anchor, and frames the moment as under-review rather than done. No evidence the voice actually does this was found; it is a possibility, not an observed behavior.
Disposition: recorded, not actioned. Priority low; evidence none. If a future observation shows the creation voice claiming a project is created/saved before instantiation, this note is the prior context. Chasing it absent evidence would be premature build against an unrealized risk.
The v0.1 "Related" note's wider suggestion still stands and is now partly discharged: a sweep of all conversational reply-composing surfaces for state-fidelity coverage. Two surfaces are now characterized — operator_converse (under the rule, post CR-2026-122) and seed-conversation (no state object, defect not reachable). Any other reply-composing surface that loads neither companion_persona.md nor this prompt remains unswept. Not known to exist; flagged for completeness if a third voice is ever found.
DUNIN7 — Done In Seven LLC — Miami, Florida Standing note — seed-conversation second voice — v0.2 — 2026-06-25