WIP: CPU research breadth night — exploratory evidence and pilots #184

Draft
pyrrha wants to merge 10 commits from research/night-breadth-20261021 into main
Member

Work-in-progress umbrella for Cassie-authorized CPU research breadth night. Current artifacts: one executed IPE1 saved-data sensitivity audit, a data-availability stop and an unexecuted procedure protocol. No claim of four completed pilots or new training. Subsequent sessions will add distinct bounded pilots with raw evidence and per-chunk validation.

Correction commit discloses first-worker premature closeout and operator-credential publication incident; original provenance is not rewritten. Existing dedicated Pyrrha identity verified before this push and PR creation. No credentials rotated/revoked.

Validation at current head: check_docs.py 476 Markdown files OK; results ledger current at 64; git diff --check clean. Prior audit tests/source receipts remain in first-session record. No merge requested; independent review pending. Entry point: research/night_breadth_20261021/README.md.

Cycle 02 — retention/adaptation pilot

Executed a bounded CPU-only synthetic A→B pilot with reset, recency, and 32-example reservoir replay controls. At the final checkpoint, reset and recency each scored A 0/64, B 64/64; replay scored A 64/64, B 64/64. Replay used 96 extra example-gradient evaluations, so this is not a compute-matched advantage. Exploratory one-stream toy evidence only; protocol, raw-row digest, limitations, and reproduction details are in research/night_breadth_20261021/cycle02/retention_adaptation/REPORT.md. Three tests passed; docs checker 478 Markdown files; ledger current (64). The seven cgroup-bounded test/pilot attempts used 0.630 CPU seconds and 0.808 service seconds; separate checks had unmeasured process CPU. Commit: 78f4168ce7fa0ef71a5d341ba4f826f134ffb34b. No merge; PR remains WIP.

Cycle 02A — second-stream sensitivity

Ran one preregistered additional stream pair (train/eval seeds 21121/22121) through the same reset, recency, and reservoir replay learner. The broad pattern recurred with timing variation: reset adapted to B and lost A; recency reached both at step 16 but lost A by step 64; replay retained A and reached B by step 64. This is exploratory two-stream synthetic evidence only, not a population estimate, compute-matched win, or real-model result. Replay costs 96 extra example-gradient evaluations. Five tests passed per bounded invocation (three invocations); the pilot ran in a separate bounded cgroup. Total measured process-tree CPU 0.461 s; service wall 0.554 s; max cgroup peak 12.3 MB; all transient cgroups cleaned. Raw 2,304-row evidence is outside Git at /home/cassie/.hermes/smc-night-20261021/evidence/cycle02-retention-adaptation-seed21121/. Report and reproducible runner: research/night_breadth_20261021/cycle02/retention_adaptation/REPORT-seed21121.md. Commit 32c8b4427760704720305381dd474531d5949e26. PR remains WIP; no merge requested.

Cycle 03 — retention through shift and return

Executed one bounded CPU-only synthetic A→B→A pilot with reset, recency and 32-example reservoir replay. At the final B checkpoint: reset A 0/64, B 64/64; recency A 0/64, B 64/64; replay A 64/64, B 64/64. At the return-to-A final checkpoint: reset A 64/64, B 0/64; recency and replay A/B 64/64. One constructed stream, known task cue, tiny linear learner; exploratory only. Replay adds 192 old-example gradient evaluations and is neither compute- nor memory-matched against reset/recency. Raw 5,760-row evidence and cgroup receipts are outside Git at /home/cassie/.hermes/smc-night-20261021/evidence/cycle03-retention-return/. Three tests passed; 90/90 aggregate cells recomputed from raw rows; test and pilot cgroups cleaned. Failed harness attempts remain preserved and are itemized in the report. See research/night_breadth_20261021/cycle03_retention_return/REPORT.md; prospective protocol and source are adjacent. Commit: 0dd9f35fe5b78db14dd7aa1cd86bf2ef88f071e1. No merge requested; PR remains WIP.

Cycle 03 evidence accounting correction: the report initially omitted temporary files in the pilot evidence tree. Verified total across the four preserved attempt directories is 927,552 bytes (14,904 log bytes), including raw evidence, manifests, logs, receipt, and temporary directory contents. This corrects the prior smaller figure; no raw rows or measurements changed. The updated report is in the current head.

Work-in-progress umbrella for Cassie-authorized CPU research breadth night. Current artifacts: one executed IPE1 saved-data sensitivity audit, a data-availability stop and an unexecuted procedure protocol. No claim of four completed pilots or new training. Subsequent sessions will add distinct bounded pilots with raw evidence and per-chunk validation. Correction commit discloses first-worker premature closeout and operator-credential publication incident; original provenance is not rewritten. Existing dedicated Pyrrha identity verified before this push and PR creation. No credentials rotated/revoked. Validation at current head: check_docs.py 476 Markdown files OK; results ledger current at 64; git diff --check clean. Prior audit tests/source receipts remain in first-session record. No merge requested; independent review pending. Entry point: research/night_breadth_20261021/README.md. ## Cycle 02 — retention/adaptation pilot Executed a bounded CPU-only synthetic A→B pilot with reset, recency, and 32-example reservoir replay controls. At the final checkpoint, reset and recency each scored A 0/64, B 64/64; replay scored A 64/64, B 64/64. Replay used 96 extra example-gradient evaluations, so this is not a compute-matched advantage. Exploratory one-stream toy evidence only; protocol, raw-row digest, limitations, and reproduction details are in `research/night_breadth_20261021/cycle02/retention_adaptation/REPORT.md`. Three tests passed; docs checker 478 Markdown files; ledger current (64). The seven cgroup-bounded test/pilot attempts used 0.630 CPU seconds and 0.808 service seconds; separate checks had unmeasured process CPU. Commit: `78f4168ce7fa0ef71a5d341ba4f826f134ffb34b`. No merge; PR remains WIP. ## Cycle 02A — second-stream sensitivity Ran one preregistered additional stream pair (train/eval seeds 21121/22121) through the same reset, recency, and reservoir replay learner. The broad pattern recurred with timing variation: reset adapted to B and lost A; recency reached both at step 16 but lost A by step 64; replay retained A and reached B by step 64. This is exploratory two-stream synthetic evidence only, not a population estimate, compute-matched win, or real-model result. Replay costs 96 extra example-gradient evaluations. Five tests passed per bounded invocation (three invocations); the pilot ran in a separate bounded cgroup. Total measured process-tree CPU 0.461 s; service wall 0.554 s; max cgroup peak 12.3 MB; all transient cgroups cleaned. Raw 2,304-row evidence is outside Git at `/home/cassie/.hermes/smc-night-20261021/evidence/cycle02-retention-adaptation-seed21121/`. Report and reproducible runner: `research/night_breadth_20261021/cycle02/retention_adaptation/REPORT-seed21121.md`. Commit `32c8b4427760704720305381dd474531d5949e26`. PR remains WIP; no merge requested. ## Cycle 03 — retention through shift and return Executed one bounded CPU-only synthetic A→B→A pilot with reset, recency and 32-example reservoir replay. At the final B checkpoint: reset A 0/64, B 64/64; recency A 0/64, B 64/64; replay A 64/64, B 64/64. At the return-to-A final checkpoint: reset A 64/64, B 0/64; recency and replay A/B 64/64. One constructed stream, known task cue, tiny linear learner; exploratory only. Replay adds 192 old-example gradient evaluations and is neither compute- nor memory-matched against reset/recency. Raw 5,760-row evidence and cgroup receipts are outside Git at `/home/cassie/.hermes/smc-night-20261021/evidence/cycle03-retention-return/`. Three tests passed; 90/90 aggregate cells recomputed from raw rows; test and pilot cgroups cleaned. Failed harness attempts remain preserved and are itemized in the report. See `research/night_breadth_20261021/cycle03_retention_return/REPORT.md`; prospective protocol and source are adjacent. Commit: `0dd9f35fe5b78db14dd7aa1cd86bf2ef88f071e1`. No merge requested; PR remains WIP. Cycle 03 evidence accounting correction: the report initially omitted temporary files in the pilot evidence tree. Verified total across the four preserved attempt directories is 927,552 bytes (14,904 log bytes), including raw evidence, manifests, logs, receipt, and temporary directory contents. This corrects the prior smaller figure; no raw rows or measurements changed. The updated report is in the current head.
docs: correct premature night closeout and disclose publication identity incident
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 10s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 13s
33f5a81d4a
research: run tiny retention adaptation pilot
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 10s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 9s
2d4fc50788
docs: record retention pilot details and failures
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 10s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 9s
f6003bf673
docs: reconcile pilot execution costs
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 10s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 10s
78f4168ce7
docs: keep generated raw evidence outside repository
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 12s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 11s
f389dced1c
docs: clarify pilot resource peak scope
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 11s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 11s
d777a71d03
research: add retention seed sensitivity pilot
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 11s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 9s
32c8b44277
research: add retention shift-return pilot
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 10s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 9s
1f51312dcf
docs: correct retention pilot evidence byte totals
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 10s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 10s
0dd9f35fe5
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 10s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 10s
Required
Details
This pull request is marked as a work in progress.
This branch is out-of-date with the base branch
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin research/night-breadth-20261021:research/night-breadth-20261021
git switch research/night-breadth-20261021
Sign in to join this conversation.
No reviewers
No labels
R03
research-b
No milestone
No project
No assignees
2 participants
Notifications
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
smolmodelco/thesmolmodelcompany!184
No description provided.