Repository navigation
docs: register C-161 — the pod runner archives half the posterior and reports OK - #559
Merged
Merged
Conversation
… reports OK Found 2026-10-13 retrieving the amber_kestrel calibration run, on a re-check one step before the pod was to be destroyed. `pod_run_model.sh` section 5 selects draw files with `-path '*/lr_*'`, so it archives three of the six target directories HydraNet writes. Measured on this run: 6.2 GB of 13 GB archived, 6.2 GB of gate draws left to die with the pod. The header comment says "the lr_* posterior" and is honest; MANIFEST prints `raw draws: 13G` beside `compressed: 34M` as though one were the compression of the other, and the run guide calls draws/ "the full posterior". The guard is the part worth recording. N_DRAWS/N_ARCHIVED exists so an empty archive cannot pass as success -- its own comment says that is "the failure this block exists to make impossible". It counts lr_*, archives lr_*, and compares the two, so it certifies the subset it was written for and cannot see the other half. A guard that defines its own scope cannot detect that the scope is wrong. What it cost: amber_kestrel is the truncated-body / soft_gate arm, where emitted P(y>0) equals the gate exactly, and its config pre-registers gate AP as a one-sided kill. No metricframe holds AP -- the gate's only metric is Brier_cls_sample, checked on both the arm and the control. `find models -name "*_by.tar.zst"` returned nothing: no model in this repository had its gate posterior, and violet_visitor's 2026-10-12 control is past recovery. The arm's gate h=1 slice survives only because it was built by hand on the pod before it was cleared. Tier 2, Open. Filed as #558 with the options; the script is production and is not touched here. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GG2cY2HaqmUMpwUyR5V82K
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Register entry only. No code, no config, no
run.sh.Records what #558 describes:
tools/podrun/pod_run_model.shsection 5 selects draws with-path '*/lr_*', so it archives three of the six target directories HydraNet writes — 6.2 GB of 13 GB on this run — and reportsSTATUS: OK.The part that makes it a register entry rather than a bug report is the guard.
N_DRAWS/N_ARCHIVEDwas added so an empty archive could not pass as success; its own comment calls that "the failure this block exists to make impossible". It countslr_*, archiveslr_*, and compares the two — so it certifies the subset it was written for and is structurally incapable of noticing the other half. Same family as C-159 and C-160: a check that reports a clean answer about input it cannot see.Measured consequence.
amber_kestrelis the truncated-body /soft_gatearm, where emitted P(y>0) equals the gate exactly, and its config pre-registers gate AP as a one-sided kill. No metricframe holds AP — the gate's only metric isBrier_cls_sample, verified on both the arm and the control.find models -name "*_by.tar.zst"returned nothing: no model in this repository had its gate posterior, andviolet_visitor's 2026-10-12 control is past recovery. The arm's gate h=1 slice (842,634 rows per target = 13 × 64,818, matching the config's own figure) exists only because it was built by hand on the pod before it was cleared.Tier 2, Open. Header counts updated 168→169 / 159→160 concerns / Open 77→78 / T2 54→55.
Noted and deliberately not changed: the tier counts have been 4 short of the concern total since before this entry (4 untiered merge stubs). That belongs to a
review-rr triage, not to this commit.Closes nothing — #558 stays open as the fix.
🤖 Generated with Claude Code
https://claude.ai/code/session_01GG2cY2HaqmUMpwUyR5V82K