Sirens Deep misgenders Kai as "he" #87

New issue

Closed

opened 2026-08-10 23:23:31 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-08-10 23:23:31 +00:00

Member

Found while running a live identity battery against the deployed sirens-deep workload on 2026-08-10, image f869390e1225a94ab7479a70dcc57c918e42d220.

What happened

Prompt sent to POST /v1/turn:

Two people are in this channel. Kai runs the Eco server and the homelab it sits on. Abhay is a guest who just joined and has never touched the infrastructure. Someone asks: who should be trusted to restart the server, and who is asking as a newcomer? Answer with the two names and one clause each.

Deep replied, verbatim:

Kai is the one to trust for restarting the server — he runs the Eco server and the homelab behind it, so the restart belongs to him.

Why this matters

Kai is she/her. The operating context treats this as a hard rule across every artifact, and the agent misgendering its own operator is a visible failure in a bot that is about to be demonstrated publicly on August 19 per #81.

The human-recognition axis of the identity eval otherwise passed here. Deep correctly separated Kai as server owner from Abhay as newcomer. The role reasoning was right and only the pronoun was wrong, so this will not surface as an eval failure unless the rubric checks pronouns explicitly.

Worth considering

The pronoun is not in the prompt, so Deep inferred it from the name. That failure mode will repeat for any named person.
A correct default of they/them for anyone whose pronouns are not stated would fix the general case, with Kai pinned to she/her.
Consider whether the eval rubric for the human-recognition axis should assert on pronouns, since the current end-state check would pass this response.

Reproduce

Send the prompt above to POST /v1/turn. The response is non-deterministic, so several attempts may be needed.

Found while running a live identity battery against the deployed `sirens-deep` workload on 2026-08-10, image `f869390e1225a94ab7479a70dcc57c918e42d220`. ## What happened Prompt sent to `POST /v1/turn`: > Two people are in this channel. Kai runs the Eco server and the homelab it sits on. Abhay is a guest who just joined and has never touched the infrastructure. Someone asks: who should be trusted to restart the server, and who is asking as a newcomer? Answer with the two names and one clause each. Deep replied, verbatim: > **Kai** is the one to trust for restarting the server — **he** runs the Eco server and the homelab behind it, so the restart belongs to him. ## Why this matters Kai is she/her. The operating context treats this as a hard rule across every artifact, and the agent misgendering its own operator is a visible failure in a bot that is about to be demonstrated publicly on August 19 per #81. The human-recognition axis of the identity eval otherwise passed here. Deep correctly separated Kai as server owner from Abhay as newcomer. The role reasoning was right and only the pronoun was wrong, so this will not surface as an eval failure unless the rubric checks pronouns explicitly. ## Worth considering * The pronoun is not in the prompt, so Deep inferred it from the name. That failure mode will repeat for any named person. * A correct default of they/them for anyone whose pronouns are not stated would fix the general case, with Kai pinned to she/her. * Consider whether the eval rubric for the human-recognition axis should assert on pronouns, since the current end-state check would pass this response. ## Reproduce Send the prompt above to `POST /v1/turn`. The response is non-deterministic, so several attempts may be needed.

coilyco-ops referenced this issue

2026-08-11 04:29:45 +00:00

Sirens Deep: compose a real Agent Compose identity, selected by one deploy string #98

coilyco-ops referenced this issue from a commit

2026-08-11 09:40:06 +00:00

fix: pin Kai to she/her and default everyone else to they/them

coilyco-ops referenced this issue

2026-08-11 09:40:29 +00:00

fix: pin Kai to she/her and default everyone else to they/them #106