feat(lilith): run 4 plan and run id (lilith-program-r4-20261006) #552

Merged

toasterson merged 1 commit from feat/lilith-r4-run-config into main 2026-10-06 10:45:02 +00:00

Conversation 0 Commits 1 Files changed 3 +104 -2

toasterson commented 2026-10-06 08:54:32 +00:00

Owner

Copy link

Run 4 of the Lilith program: docs/ai/plans/2026-10-06-lilith-program-r4-siblings-in-context.md.

Run 3 (PR #550, NOT CERTIFIED) showed that the shadow mechanism was mostly absent from the data she trained on: 25 of 31 teacher siblings unfit at 16k tokens, every sibling-contrast pair refused by the trainer (pair_prefix), one hand-off race. Run 4 keeps belts, families, quotas, teachers and the init adapter identical to run 3 (the control) and changes only how sibling data reaches the trainer: 32k training context after a memory probe, episode tool-result budgets for the teachers (PR #551), first-divergence preference pairs (PR feat/lilith-first-divergence-pairs).

This PR: the plan, the default run id lilith-program-r4-20261006 in program/schema.py, the test pin. Anima run WI 01a11069-448a under plan 01a0e26f-d032.

Tests: uv run --no-project --with pytest python -m pytest program/tests -q → 137 passed, 4 skipped.

Run 4 of the Lilith program: `docs/ai/plans/2026-10-06-lilith-program-r4-siblings-in-context.md`. Run 3 (PR #550, NOT CERTIFIED) showed that the shadow mechanism was mostly absent from the data she trained on: 25 of 31 teacher siblings unfit at 16k tokens, every sibling-contrast pair refused by the trainer (`pair_prefix`), one hand-off race. Run 4 keeps belts, families, quotas, teachers and the init adapter identical to run 3 (the control) and changes only how sibling data reaches the trainer: 32k training context after a memory probe, episode tool-result budgets for the teachers (PR #551), first-divergence preference pairs (PR feat/lilith-first-divergence-pairs). This PR: the plan, the default run id `lilith-program-r4-20261006` in `program/schema.py`, the test pin. Anima run WI `01a11069-448a` under plan `01a0e26f-d032`. Tests: `uv run --no-project --with pytest python -m pytest program/tests -q` → 137 passed, 4 skipped.

toasterson added 1 commit 2026-10-06 08:54:32 +00:00

feat(lilith): run 4 plan and run id (lilith-program-r4-20261006) ... e2d18c54b5

Run 3 (PR #550) showed the shadow mechanism was mostly absent from the trained data: teacher siblings unfit at 16k, sibling-contrast pairs refused (pair_prefix), one hand-off race. Run 4 keeps everything else fixed and tests the three data fixes with run 3 as the control.

toasterson referenced this pull request 2026-10-06 08:55:33 +00:00

feat(lilith): first-divergence preference pairs at export #553

toasterson referenced this pull request 2026-10-06 09:01:34 +00:00

fix(lilith): panel orders state the grounding rule (a cited receipt is not visible evidence) #554

toasterson merged commit 98de3c2faf into main 2026-10-06 10:45:02 +00:00

toasterson referenced this pull request from a commit 2026-10-06 10:45:04 +00:00

Merge pull request 'feat(lilith): run 4 plan and run id (lilith-program-r4-20261006)' (#552) from feat/lilith-r4-run-config into main

toasterson referenced this pull request 2026-10-06 22:20:23 +00:00

docs(lilith): RESULT for program r4 (siblings in context) — NOT CERTIFIED #568

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!552

No description provided.