Feedback & tuning
Example data. No family has rated a reply, flagged a fact or scored a voice, so every item here is an opinion nobody held.
Safety tuning signal
Paused a conversation that did not need pausing — reported by the member
A person in crisis not escalated — reported by nobody
This is not zero. It is the absence of a signal: nobody files a ticket about the help that did not arrive, and the person it happened to is the person least placed to report it. Better tooling does not fix it, because the error consists of an absence. RFD-0026.
Every item below argues for loosening the threshold. Not because the items are wrong, but because the error that would argue the other way is the one nobody can file. Four weeks of this, reviewed each Monday with a client asking why the product keeps interrupting people, is the pressure RFD-0026 exists to answer — and its default is to freeze the threshold for the duration of the pilot and ship no safety tuning at all.
What runs with no counsellor
1 of 4 feedback paths runs without a counsellor — wrong_fact.
H-10 has never landed, so the other three are waiting on a person who does not exist. This one is not waiting, and what it edits is the record of what a dead man said. S41 narrowed it: a correction auto-applies only when every telling of the fact is the corrector’s own. One reaching somebody else’s telling becomes a dispute, routes to a counsellor and puts the persona in the priority lane — because the owner created the persona, which is not the same as being the person whose memory of the dead man is authoritative.
Feedback
| Item | Type | Persona | Waiting | Approval path |
|---|---|---|---|---|
fb-118 | out_of_character | persona-a31f | 5 days | counsellor regression probes · |
fb-117 | wrong_fact | persona-77c2 | 4 days | dispute Reaches a telling that is not the corrector’s. Counsellor, and the persona moves to the priority lane no further gate |
fb-116 | wrong_fact | persona-77c2 | 2 days | auto-apply no further gate |
fb-115 | low_voice_rating | persona-51ae | 1 days | counsellor owner preview · |
fb-114 | consistent_positive | persona-0b94 | 0 days | counsellor no further gate |
Waiting time is in calendar days. WF-09 says business days and nothing here computes a business calendar, so this figure runs slightly ahead of the SLA rather than behind it — which is the direction to err.
What tuning may touch
prompt_weightingmemory_updateper_persona_embedding_refinement
A closed list, and membership is exact. shared_embedding_refinement is one word from a permitted action and is precisely what SoW §4.8 forbids; it is refused, as is anything else not named above. Customer data never trains or fine-tunes a shared foundation model — npm run guard fails the build on any path that would.
What this screen cannot do
It applies nothing. There is no approve control and no threshold slider, for the reason the other screens give — no actor — and for one more: the decision about whether safety tuning ships during the pilot is RFD-0026, which is open, and a control here would be that decision taken by whoever happened to click.
Nor does it show a sentiment, a trend or a net score. A single number summarising a one-sided stream is the most quotable thing this screen could produce and the least true. RFD-0023 is the same problem for persona fidelity: promoting well-rated replies tunes toward comfort, because a bereaved family rates the comforting reply above the accurate one. Its default was taken on 2027-08-13 and logged 14 days late.