test(unified): add deriver_no_memory_signal naturalistic suite
L1 reproducer for the deriver_no_memory_signal failure mode: the
deriver writes observations that don't add signal about the user or
their work -- bare acknowledgments paraphrased as facts ("user said
yes"), shell-history fragments ("user executed cd ~/foo"), instruction-
event-records that look like durable user attributes but aren't.
Inputs: 8 sessions sampled from SALT-NLP/SWE-chat (ODC-BY,
https://arxiv.org/abs/2604.20779), persona-balanced across Expert
Nitpicker / Vague Requester / Mind Changer. Two sessions from the
underlying extract are excluded for data quality (one Japanese with
a heavily skewed user/assistant role split; one near-monologue);
documented in the fixture description.
Per session: ingest messages, wait for queue empty, run an llm_judge
asking whether the user's representation contains any Tier 1 bare-
speech-act observation, then save_artifact-dump the representation
for offline analysis. The suite uses continue_on_failure to aggregate
per-session pass/fail. Judge prompt is calibrated with annotated
example observations across three tiers (Tier 1 / Tier 2 / Signal) to
distinguish bare-event records from specifics-anchored observations.
Inputs file at tests/unified/test_cases/data/swe_chat_sample_persona_balanced.json
-- the committed dataset extract. Metadata embedded in the file
documents its provenance (SALT-NLP/SWE-chat, ODC-BY, selection
criteria, extraction timestamp).
This commit is contained in:
parent
d7c7568ceb
commit
a59d69a9da
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
Loading…
Reference in New Issue