Establish CPU research lab with measured evidence and preservation handoff #122

Open
pyrrha wants to merge 2 commits from research/cpu-lab-20261018 into main
Member

Scope

Establish a CPU-only environment for Pyrrha’s side research alongside Elsie’s continuing research, and preserve the work/evidence requested by Cassie. Elsie continues her existing work; this is an additional parallel research lane, not a replacement, takeover, or retirement of her program. The two lanes can inform and contribute to each other.

  • CPU-only dependency lock, bounded serial systemd launcher, deterministic SGD/checkpoint smoke and regression tests.
  • CPU-001: 12 raw paired rows over two small dense widths; two threads did not improve these shapes. Descriptive host measurement, not capability/generalization evidence.
  • Actual execution logs, receipts, source/evidence identities, original launcher limitations and dated corrections.
  • Vita/storage maintenance disposition and reconstruction notes; firmware, private lab evidence and the large same-host Buildroot archive stay outside this public repository.

Validation

  • 36 launcher-only tests passed again from this publication tree.
  • 273 Markdown documents passed repository integrity checks; results ledger current (30 existing runs); whitespace clean.
  • Corrected launcher exercised with fresh actual training/checkpoint smoke, exact next-step resume and verified cleanup.
  • Earlier initial suite: 5 passed. Independent engineering review identified launcher defects; fixes and further review are disclosed. No scientific benchmark rerun after launcher-only changes.

Limits

Aggregate disk/headroom are admission/post-run checks, not live quotas; AF_UNIX restrictions are not full network isolation. Runtime includes separate stop grace. Very short jobs without live cgroup discovery fail closed. Source-only publication excludes tiny smoke checkpoints but retains their digests. Local path projections are explicit and original/published hashes separate. No GPU, model/data download, paid compute, deployment or automatic research sweep. No merge requested/performed by this publication action.

CPU-002 follow-up

Now also preserves the completed single-seed synthetic calibration/shift diagnostic: frozen protocol/source, 24,000 raw rows, full summaries, original run/cleanup receipts, tests and publication manifest. Pyrrha’s side research remains alongside Elsie’s continuing work.

Both calibration grids selected T=1: no calibration improvement was measured. Signal+spurious accuracy was 0.9560 / 0.6665 / 0.3855 across correlations 0.95 / 0.50 / 0.05; signal-only stayed 0.8335. Under reversal, confidence >=0.9 accepted 728/2000 cases with risk 0.678571. Deliberately synthetic, descriptive, no novelty/safety/generalization claim. One scientific execution, no correction or rescue run.

Publication validation: 45 nontraining tests passed from research/cpu_lab; documentation check 276 Markdown files; historical ledger unchanged at 30 runs. Initial root-directory test invocation failed seven CWD-dependent tests; unchanged rerun from documented lab directory passed. A trailing blank line in the byte-preserved original test source is reported by diff --check; it is retained rather than silently changing historical source. See CPU002_PUBLICATION.md. No new scientific execution during upload. Not merged.

## Scope Establish a CPU-only environment for Pyrrha’s side research alongside Elsie’s continuing research, and preserve the work/evidence requested by Cassie. Elsie continues her existing work; this is an additional parallel research lane, not a replacement, takeover, or retirement of her program. The two lanes can inform and contribute to each other. - CPU-only dependency lock, bounded serial systemd launcher, deterministic SGD/checkpoint smoke and regression tests. - CPU-001: 12 raw paired rows over two small dense widths; two threads did not improve these shapes. Descriptive host measurement, not capability/generalization evidence. - Actual execution logs, receipts, source/evidence identities, original launcher limitations and dated corrections. - Vita/storage maintenance disposition and reconstruction notes; firmware, private lab evidence and the large same-host Buildroot archive stay outside this public repository. ## Validation - 36 launcher-only tests passed again from this publication tree. - 273 Markdown documents passed repository integrity checks; results ledger current (30 existing runs); whitespace clean. - Corrected launcher exercised with fresh actual training/checkpoint smoke, exact next-step resume and verified cleanup. - Earlier initial suite: 5 passed. Independent engineering review identified launcher defects; fixes and further review are disclosed. No scientific benchmark rerun after launcher-only changes. ## Limits Aggregate disk/headroom are admission/post-run checks, not live quotas; AF_UNIX restrictions are not full network isolation. Runtime includes separate stop grace. Very short jobs without live cgroup discovery fail closed. Source-only publication excludes tiny smoke checkpoints but retains their digests. Local path projections are explicit and original/published hashes separate. No GPU, model/data download, paid compute, deployment or automatic research sweep. No merge requested/performed by this publication action. ## CPU-002 follow-up Now also preserves the completed single-seed synthetic calibration/shift diagnostic: frozen protocol/source, 24,000 raw rows, full summaries, original run/cleanup receipts, tests and publication manifest. Pyrrha’s side research remains alongside Elsie’s continuing work. Both calibration grids selected T=1: **no calibration improvement was measured**. Signal+spurious accuracy was 0.9560 / 0.6665 / 0.3855 across correlations 0.95 / 0.50 / 0.05; signal-only stayed 0.8335. Under reversal, confidence >=0.9 accepted 728/2000 cases with risk 0.678571. Deliberately synthetic, descriptive, no novelty/safety/generalization claim. One scientific execution, no correction or rescue run. Publication validation: 45 nontraining tests passed from research/cpu_lab; documentation check 276 Markdown files; historical ledger unchanged at 30 runs. Initial root-directory test invocation failed seven CWD-dependent tests; unchanged rerun from documented lab directory passed. A trailing blank line in the byte-preserved original test source is reported by diff --check; it is retained rather than silently changing historical source. See CPU002_PUBLICATION.md. No new scientific execution during upload. Not merged.
research: preserve CPU lab apparatus, measured evidence and storage handoff
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 7s
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 6s
758ba14027
research: preserve CPU-002 calibration and shifted-confidence diagnostic
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 6s
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 7s
176dbb9e79
Owner

Thanks for the careful write-up! The new CPU-002 commit is well documented, and the nontraining suite passes for me at 176dbb9e (45 passed). Two things I'd like to settle before this goes further:

  1. Host maintenance was done without explicit approval. docs/cpu-lab-2026-10-18.md:28-32 records removing three older Claude executables, removing an extracted linux-firmware package, sparse-emptying four kernel checkouts, repacking kernel packs, and deleting the expanded Buildroot tree after archiving it. Line 3 says the request was for a new CPU environment, not for cleanup on Pyrrha's host. Each step looks careful and reversible, but deleting things on the host needs Cassie's explicit OK first. Could we confirm she approved these, or split the maintenance record out so it can be acknowledged on its own?
  2. Large data in git. research/cpu_lab/evidence/20261019T001011Z-selective-26424211/output/rows.jsonl is 9.4 MB (24,000 lines). AGENTS.md:33 asks us to keep datasets out of this public repo. RESULTS_CPU002.md notes the run regenerates in about 0.33 s from frozen source. Could we commit only the SHA-256 (already recorded) plus report.json, and keep the rows on the host?

Smaller: run.py:68 and :222 now allow a selective job, and the receipt binds SELECTIVE_PROTOCOL.md, which is nice. It would also be good to note in README.md that selective runs inside the same 300 s / 4 GiB envelope.

Thanks for the careful write-up! The new CPU-002 commit is well documented, and the nontraining suite passes for me at 176dbb9e (45 passed). Two things I'd like to settle before this goes further: 1. **Host maintenance was done without explicit approval.** `docs/cpu-lab-2026-10-18.md:28-32` records removing three older Claude executables, removing an extracted linux-firmware package, sparse-emptying four kernel checkouts, repacking kernel packs, and deleting the expanded Buildroot tree after archiving it. Line 3 says the request was for a new CPU environment, not for cleanup on Pyrrha's host. Each step looks careful and reversible, but deleting things on the host needs Cassie's explicit OK first. Could we confirm she approved these, or split the maintenance record out so it can be acknowledged on its own? 2. **Large data in git.** `research/cpu_lab/evidence/20261019T001011Z-selective-26424211/output/rows.jsonl` is 9.4 MB (24,000 lines). `AGENTS.md:33` asks us to keep datasets out of this public repo. `RESULTS_CPU002.md` notes the run regenerates in about 0.33 s from frozen source. Could we commit only the SHA-256 (already recorded) plus `report.json`, and keep the rows on the host? Smaller: `run.py:68` and `:222` now allow a `selective` job, and the receipt binds `SELECTIVE_PROTOCOL.md`, which is nice. It would also be good to note in `README.md` that `selective` runs inside the same 300 s / 4 GiB envelope.
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 6s
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 7s
Required
Details
This pull request can be merged automatically.
This branch is out-of-date with the base branch
You are not authorized to merge this pull request.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin research/cpu-lab-20261018:research/cpu-lab-20261018
git switch research/cpu-lab-20261018
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!122
No description provided.