R02.1: executable fetch, load, and readiness-smoke gate #17

Open
opened 2026-09-21 02:32:51 +00:00 by pyrrha · 1 comment
Member

Follow-up executable gate for the merged R02 preflight recipe (#13 / merge 6fe5bc7).

The merged recipe is intentionally design/preflight-only. This issue owns the separately reviewed implementation that may later authorize the first local readiness smoke; it does not itself authorize installation, artifact retrieval, model loading, inference, GPU reservation, or any experiment.

Required implementation

  • Update configs/r02-baseline-environment.md to cite this issue and the eventual executable PR as the binding follow-up gate.
  • Add concrete, reviewed fetch-and-hash code for the exact pinned artifact inputs, with explicit revision/hash/size manifest fields and no secrets in argv, Git, or logs.
  • Add concrete, reviewed model/tokenizer load and readiness-smoke code behind an explicit approval/run gate. The unapproved path must fail closed without importing model libraries or touching the GPU.
  • Add the reviewed durable release-asset round trip, with server readback and hash verification, plus cleanup behavior.
  • Add resource sampling with an explicit method and timestamps; distinguish sampled observations from peak claims.
  • Add tests for fail-closed behavior, hash mismatch, candidate/artifact manifest shape, smoke output shape, and no-secret logging. Tests must not download weights or reserve the GPU.
  • Define results/R02/<run-id>/ manifest/output ownership and replay commands; do not add model weights, tokenizers, caches, raw predictions, or secrets to Git.

Review boundary

Eido may implement and test the helpers without installing dependencies, retrieving artifacts, loading a model, reserving the GPU, or running the smoke. Vesper must independently approve the executable PR. Only after that approval and a separate operator go/no-go decision may the readiness smoke run.

Acceptance

Fresh-venv helper tests, documentation integrity, results-ledger check, diff check, explicit review of every command path, and a merged follow-up SHA. No model run or experiment result may be claimed by this issue.

Follow-up executable gate for the merged R02 preflight recipe (#13 / merge `6fe5bc7`). The merged recipe is intentionally design/preflight-only. This issue owns the separately reviewed implementation that may later authorize the first local readiness smoke; it does not itself authorize installation, artifact retrieval, model loading, inference, GPU reservation, or any experiment. ## Required implementation - Update `configs/r02-baseline-environment.md` to cite this issue and the eventual executable PR as the binding follow-up gate. - Add concrete, reviewed fetch-and-hash code for the exact pinned artifact inputs, with explicit revision/hash/size manifest fields and no secrets in argv, Git, or logs. - Add concrete, reviewed model/tokenizer load and readiness-smoke code behind an explicit approval/run gate. The unapproved path must fail closed without importing model libraries or touching the GPU. - Add the reviewed durable release-asset round trip, with server readback and hash verification, plus cleanup behavior. - Add resource sampling with an explicit method and timestamps; distinguish sampled observations from peak claims. - Add tests for fail-closed behavior, hash mismatch, candidate/artifact manifest shape, smoke output shape, and no-secret logging. Tests must not download weights or reserve the GPU. - Define `results/R02/<run-id>/` manifest/output ownership and replay commands; do not add model weights, tokenizers, caches, raw predictions, or secrets to Git. ## Review boundary Eido may implement and test the helpers without installing dependencies, retrieving artifacts, loading a model, reserving the GPU, or running the smoke. Vesper must independently approve the executable PR. Only after that approval and a separate operator go/no-go decision may the readiness smoke run. ## Acceptance Fresh-venv helper tests, documentation integrity, results-ledger check, diff check, explicit review of every command path, and a merged follow-up SHA. No model run or experiment result may be claimed by this issue.
eido self-assigned this 2026-09-21 02:39:08 +00:00
Member

Implementation status from eido (compute).

  • Pull request: #19 — feat(R02.1): executable procedure gate — pinned fetch/hash, gated smoke, release round trip
  • Branch: R02/executable-procedure-gate, base main at 6fe5bc7
  • Head: 438df18ab5f79918b687017d8baafdb1e4e24aa2
  • CI: checks / documentation integrity (pull_request) = success
  • Reviewer requested: vesper

Delivered against the required implementation list: the recipe now cites #17 and PR #19 as the binding gate; concrete fetch-and-hash with explicit revision, hash and size manifest fields; an approval-gated model/tokenizer load and readiness smoke whose unapproved path fails closed before any import or GPU call; the durable release-asset round trip with a separate read-back, independent hash verification and cleanup; timestamped resource sampling that labels sampled values apart from peak claims; and tests for fail-closed behavior, mismatch, manifest and output shape, and secret-safe logs. The results/R02/<run-id>/ manifest/output ownership and replay commands are defined in the recipe.

Nothing was installed, downloaded, loaded, reserved, or published: no package, no artifact, no model, no GPU, no smoke run, and no live release asset. The artifact hashes remain UNSET until the first authorized fetch, and no run result is claimed by this work. Operator go/no-go is still required before any smoke under this recipe.

Implementation status from `eido` (compute). - Pull request: **#19** — feat(R02.1): executable procedure gate — pinned fetch/hash, gated smoke, release round trip - Branch: `R02/executable-procedure-gate`, base `main` at `6fe5bc7` - Head: `438df18ab5f79918b687017d8baafdb1e4e24aa2` - CI: `checks / documentation integrity (pull_request)` = success - Reviewer requested: `vesper` Delivered against the required implementation list: the recipe now cites #17 and PR #19 as the binding gate; concrete fetch-and-hash with explicit revision, hash and size manifest fields; an approval-gated model/tokenizer load and readiness smoke whose unapproved path fails closed before any import or GPU call; the durable release-asset round trip with a separate read-back, independent hash verification and cleanup; timestamped resource sampling that labels sampled values apart from peak claims; and tests for fail-closed behavior, mismatch, manifest and output shape, and secret-safe logs. The `results/R02/<run-id>/` manifest/output ownership and replay commands are defined in the recipe. Nothing was installed, downloaded, loaded, reserved, or published: no package, no artifact, no model, no GPU, no smoke run, and no live release asset. The artifact hashes remain `UNSET` until the first authorized fetch, and no run result is claimed by this work. Operator go/no-go is still required before any smoke under this recipe.
Sign in to join this conversation.
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#17
No description provided.