docs(lilith): run 5 plan and ADR 126 — CLI teachers, pre-collapse signals, transcript probes #569

Merged

toasterson merged 1 commit from docs/lilith-program-r5-plan into main 2026-10-07 06:52:29 +00:00

Conversation 0 Commits 1 Files changed 2 +148

toasterson commented 2026-10-07 05:59:18 +00:00

Owner

Copy link

Run 4 (PR #568) trained a candidate that loops on tool reads and nobody saw it before the exam. Operator directives (2026-10-06): the teachers are Claude Fable 5.1, Claude Opus 5.5 and GPT-6 Astra through their CLIs (the sealed Ollama teachers are capped until the monthly reset); pre-collapse signals from the literature decide when a run is adjusted; her transcripts are observed during the run and loops are intervened on immediately.

ADR 126 records the signals (likelihood displacement — Razin et al. 2024; entropy collapse — Cui et al. 2025; length exploitation — Park et al. 2024 / SimPO; gradient spikes; the transcript loop probe with run-3/run-4 reference numbers) and the protocol (probe points every 40 steps and at stage ends, guard trips acknowledged by hash, rollback actions, serving faults rerun, no idle pod time). The plan lists the five implementation slices (CLI teacher adapter, trainer guard + per-token preference normalization, loop probe + supervisor hand-off, trainer probe protocol with rollback, answer-turn pairs) and the run procedure.

Run 4 (PR #568) trained a candidate that loops on tool reads and nobody saw it before the exam. Operator directives (2026-10-06): the teachers are Claude Fable 5.1, Claude Opus 5.5 and GPT-6 Astra through their CLIs (the sealed Ollama teachers are capped until the monthly reset); pre-collapse signals from the literature decide when a run is adjusted; her transcripts are observed during the run and loops are intervened on immediately. ADR 126 records the signals (likelihood displacement — Razin et al. 2024; entropy collapse — Cui et al. 2025; length exploitation — Park et al. 2024 / SimPO; gradient spikes; the transcript loop probe with run-3/run-4 reference numbers) and the protocol (probe points every 40 steps and at stage ends, guard trips acknowledged by hash, rollback actions, serving faults rerun, no idle pod time). The plan lists the five implementation slices (CLI teacher adapter, trainer guard + per-token preference normalization, loop probe + supervisor hand-off, trainer probe protocol with rollback, answer-turn pairs) and the run procedure.

toasterson added 1 commit 2026-10-07 05:59:20 +00:00

docs(lilith): run 5 plan and ADR 126 — CLI teachers, pre-collapse signals, transcript probes with intervention ... 3d1766eee3

Operator directives after run 4's looping candidate: Fable, Opus 5.5 and Astra act as teachers through their CLIs (the Ollama teachers are capped until the monthly reset); likelihood displacement, entropy collapse, length exploitation and gradient spikes are measured every step with a guard; her transcripts are probed every 40 steps and at stage ends with immediate rollback.

toasterson referenced this pull request 2026-10-07 06:04:43 +00:00

feat(lilith): CLI-driven teacher adapter and sibling teacher-kind (claude -p, codex exec) #570

toasterson merged commit ff8032e6f9 into main 2026-10-07 06:52:29 +00:00

toasterson referenced this pull request from a commit 2026-10-07 06:52:32 +00:00

Merge pull request 'docs(lilith): run 5 plan and ADR 126 — CLI teachers, pre-collapse signals, transcript probes' (#569) from docs/lilith-program-r5-plan into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels   

follow-up

Documented operational follow-up

No labels

follow-up

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

toasterson/akh-medu!569

No description provided.