Skip to content

REDIRECTION_EXTERNAL_ACTIONS.md

Latest commit

 

History

History
158 lines (119 loc) · 6.35 KB

REDIRECTION_EXTERNAL_ACTIONS.md

File metadata and controls

158 lines (119 loc) · 6.35 KB
Raw

Redirection external-action ledger

This ledger records the external actions for PLAN_REDIRECTION.md, including completed mutations and the optional inactive human-only branch. External steps mutate GitHub, the remote GPU host, or human evidence and therefore must not be inferred from an ordinary build/test request.

Invariants

  • Do not install, download, or run Ollama/model weights on the developer Mac.
  • windows-gpu is the only model-execution host for P5, P6, and P6-M.
  • P5, P6, and P6-M are sequential; never run their model services concurrently.
  • Every remote starter is default-deny and checks a clean exact commit, idle GPU, immutable model digest, frozen inputs, and dedicated output root.
  • P5/P6/P6-M use development evidence only. They never authorize held-out/test use.
  • P6 remains retrospective and cannot confirm H4 even after human completion.
  • artifacts-v1 is immutable. Future accepted P6 human results must be preserved in a new versioned archive and independently read back before the packaged manifest is switched.

1. Publish the reviewed source commit (complete)

The reviewed source and subsequent P10-B/CI repairs are on origin/main. Future source pushes still require explicit GitHub push authorization:

git status --short
git push origin main
test "$(git rev-parse HEAD)" = "$(git rev-parse origin/main)"

Do not prepare the remote host if local HEAD and origin/main differ.

2. P5 registered remote ablations (complete)

The three registered 350-row runs completed, were returned, and passed the independent verifier. The commands below remain the historical runbook; re-running them would require new authorization:

scripts/prepare_windows_p5_ablations.sh
FAR_P5_PREP_ALLOWED=1 scripts/prepare_windows_p5_ablations.sh --execute

scripts/start_windows_p5_ablations.sh
FAR_P5_TRAINING_ALLOWED=1 scripts/start_windows_p5_ablations.sh --execute
scripts/check_windows_p5_ablations.sh

After all three 350-row manifests complete, stop the dedicated services and return the bundle:

scripts/stop_windows_p5_ablations.sh --execute --stop-ollama
FAR_P5_FETCH_ALLOWED=1 scripts/fetch_windows_p5_ablations.sh --execute

The fetch command ends with the independent zero-model verifier. H3/H5 remain unset until it returns valid:true from the complete remote bundle.

3. P6 machine prelabels (complete)

The authorized run completed 217 rows in 221 attempts, returned the native provenance bundle, and stopped its dedicated model service. The commands below are retained as the historical runbook and are not authorization to rerun it:

scripts/prepare_windows_p6_prelabels.sh
FAR_P6_PREP_ALLOWED=1 scripts/prepare_windows_p6_prelabels.sh --execute

scripts/start_windows_p6_prelabels.sh
FAR_P6_PRELABEL_ALLOWED=1 scripts/start_windows_p6_prelabels.sh --execute
scripts/check_windows_p6_prelabels.sh

After 217 rows complete:

scripts/stop_windows_p6_prelabels.sh --execute --stop-ollama
FAR_P6_FETCH_ALLOWED=1 scripts/fetch_windows_p6_prelabels.sh --execute

The installer preserves and checks every native prompt/raw-response hash. The accepted result is present in the published artifacts-v1 diagnostic release.

4. P6-M machine-only stability audit (complete)

The separately authorized non-human substitute audit completed three distinct remote families, two views per item, and 1,302 accepted rows with zero failed attempts. Returned raw outputs and the deterministic report verify. Consensus covered only 15/217 samples and 202 remained contested, so P6-M did not provide a broad human substitute and did not modify any P6 reviewer/adjudicator slot. Both dedicated remote services are inactive. Re-running P6-M would require new authorization and a new explicit protocol decision. The minimal 1.7 MiB juror evidence tree is tracked with the report so a fresh clone can rerun prompt/raw-response and identity verification after installing the published diagnostic packet; no model weights or caches are included.

5. Optional future P6 human branch (inactive)

No real reviewers or adjudicator are available. This branch is outside the accepted no-human profile, has no active owner or deadline, and is not a current project blocker. It reopens only if real people become available and the project explicitly chooses to pursue strict human-mappability, human-IAA, or adjudicated-gold claims. It cannot be automated or substituted with more models.

If it is ever reopened:

  1. Give two distinct reviewers only their respective blank 217-row templates and visible contexts; hide machine labels, scores, analysis strata, and the other reviewer.
  2. Freeze and install both complete reviewer files.
  3. Give a third distinct adjudicator the two frozen reviews and machine labels.
  4. Install all 217 adjudications, analyze, and run the independent verifier as described in P6_EXECUTION.md.

Low agreement is a result, not permission to remove rows or change the ontology.

6. P10-B completion record

P10-B received separate authorization and completed on 2026-07-12. The exact archive was published as artifacts-v1:

  • bytes: 5,639,635;
  • SHA-256: 5e3f28dcd81d2af3170f740611b9f59b8bbe1ee6e869379d5794730db4ecf96e;
  • files: 336;
  • tree SHA-256: 8f620af737f3b04f5b3813b06c7183743a3ae14f7c8fd869f43a83b9a821dbff.

The asset was downloaded from GitHub through the installer into an empty directory and compared against the complete source tree before tracked diagnostics/ removal. Fresh checkouts and CI now install the ignored local tree from the immutable manifest:

uv run falsirag ops diagnostic-data install

This authorization did not include a Git-history rewrite or LFS migration; historical blobs remain reachable from older commits and the immutable tag.

Authorization boundaries

The following are deliberately separate approvals:

  • push the reviewed source commit;
  • modify the remote worktree/install units;
  • start P5 model execution;
  • fetch P5 artifacts;
  • start P6 model execution;
  • fetch/install P6 artifacts;
  • start/fetch P6-M machine-only audit artifacts;
  • publish the diagnostic GitHub release and remove tracked diagnostics;
  • supply real reviewer/adjudicator work.

The final item is an optional future authorization, not an active dependency.

An approval for one item does not imply any later item.