Repository navigation
Conversation
…a harness error (L2) llama-cli writes to /dev/tty when the cell inherits a controlling tty (tmux): 0 B captured, rc 0, a blank cell. Test fails without setsid. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…A-007 (Refs #5639) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Six tests: missing config, refused connect, immediate Shutdown, one full shard/gradient/averaged round trip with shard clamping, Shutdown during AllReduce, and a non-shard message. Mutating the shard_end clamp turns the round-trip test red. Adds FALSIFY-DISTRIBUTED_TRAINING_V1_003. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… truncation, origin) (Refs #5601) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…set on content rows (Refs #5601) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… HEAD (Refs #5762) A stale-pv regen on #5776 rewrote shapes.ttl in an old form and cost ~25 min. pv_bin.sh proves the tree; the resolver now also requires pv and pv-sat to be newer than every tracked verifier source, ignores an inherited PV_SAT, and checks before the conflicted files are dropped. Not run against a real pv build. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…5762) The guard's loop ended on a false test, so $(...) returned 1 and set -e killed the resolver with no message. Found by running it against a real pv build; the stub case table had no set -e. Now fresh passes (rc 0, no tree change) and a touched verifier source is refused with the rebuild command. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…egenerates (Refs #5762) pv_bin.sh builds --bin pv only, so pv-sat stayed whatever an earlier build left (the #5776 stale-verifier regen, ~25 min). Build both with cargo (its fingerprint is the freshness authority) before pv_bin.sh snapshots them and before the conflicted files are dropped; ignore an inherited PV_SAT. Supersedes the mtime guard, which a quorum lane showed refuses on merged non-compiled files and a lane found a corrupted header in. Run against a real build: a touched pv-sat source is rebuilt and snapshotted, a fresh tree regenerates with no diff. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…drops the conflicted files (Refs #5762) Drops the extra cargo build (a quorum lane: pv_bin.sh already builds pv and pv-sat with --features update-check,build-sha, so a bare build thrashes the fingerprint and the stale-pv-sat premise is unproven). Keeps the reorder: a refusal or failed build now leaves the merge as it was; PV_SAT is the one beside the resolved PV, never inherited. Real run: tree regenerates with no diff. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
§13.11 rung 1 — quorum shadow verdict Shadow mode: this records a verdict and merges nothing. A refusal |
noahgift
left a comment
noahgift
left a comment
There was a problem hiding this comment.
seat 1, aprender-pace @1a26ef25fda8f65708beaa4ae8186c388c0ceeba
Non-author seat (0 aprender-pace trailers in origin/main..head).
Read: the net diff against main (excluding generated contracts/witness, contracts.nt, README) is 328 added lines: tests + two contract rows. The five Revert commits (crux setsid, contracts_resolve x4) cancel their originals, so those scripts are unchanged vs main. New tests: aprender-explain tui/source_pane_tests.rs (#5601), aprender-qa-cli certification_processing.rs profiling_assertion_tests, aprender-train classify_trainer_tests.rs run_worker_tests. Contract rows FALSIFY-BIN-APR-QA-007 and FALSIFY-DISTRIBUTED_TRAINING_V1_003 name tests that exist and run.
Ran at this head (detached worktree): source_pane 6/6 pass, profiling_assertion_tests 5/5 pass, run_worker_tests 6/6 pass.
Not verified: no mutation of my own on any of the three test sets (the qa-cli row's unwrap_or(5.0)->6.0 mutation claim is the author's, not re-run); generated files not regenerated or diffed; CI state not read; the revert pairs were judged by net diff, not commit by commit.
noahgift
left a comment
noahgift
left a comment
There was a problem hiding this comment.
Seat 2 (aprender-w5018, non-author; 0 w5018 commits in the PR): APPROVE @1a26ef25fda8f65708beaa4ae8186c388c0ceeba
Read the full net diff (6 files, +328/-0, tests and 2 contract rows only; the 4 contract-script fixes and the crux setsid commit are reverted in-branch, so none of that code lands). Checks: profiling_assertion_tests pins the 5/50 defaults, the strict-below boundary, legacy and per-backend override; run_worker_tests drives a real loopback coordinator for the clamp, the three fail-closed paths and the one-step round trip; source_pane_tests assert cells, not glyph art. Contract rows FALSIFY-BIN-APR-QA-007 and FALSIFY-DISTRIBUTED_TRAINING_V1_003 name tests that exist. Not verified by me: the stated mutation RED runs; CI 8/8 green covers execution.
…5789) * fix(deps): wasmtime 48.0.6 -> 48.0.7 for RUSTSEC-2026-0378 and RUSTSEC-2026-0379 Advisories published 2026-10-19 turned sov.security red on main and every merge-group run (MG 39413691864, #5781, job 118597427403). Lockfile-only bump inside the declared 48 range; no ignore added, no gate edited. Two-way door. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * fix(deps): keep main's lock edges; only the wasmtime family moves (#5789) The first bump let cargo 1.93 flip 17 unrelated dependency edges to older versions (windows-sys, windows-link, unicode-width, itertools). Rebuilt from main's lock with only the wasmtime/cranelift/pulley package blocks taken from the 48.0.7 resolve. 56 changed lines vs main, all in that family. cargo metadata --locked passes; cargo audit exit 0. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
|
Closed per Noah's 15:34Z ruling (verbatim via arbiter-cop). Tree diff: 5 of 6 files have identical blobs on main or in #5840; the 6th, contracts/bin-aprender-qa-cli--apr-qa-v1.yaml, differs from #5840 by ONE line, the same mutation note with a newer run stamp (2026-10-21T10:15Z in #5840 vs 2026-10-20 here), so #5840 carries the newer version of this content. |
Batch 7: 9 constituents (one commit each), full local T1 gate green on this exact tree.
setsid -w, rc 0 + 0 B stdout is a harness errorcheck_profiling_assertionstests + FALSIFY-BIN-APR-QA-007run_workertests (fake TCP coordinator)render_source_panetestscontracts_resolve.shfixes (Refs #5762)Gate receipt:
/mnt/nvme-raid0/cop-inbox/handoff/b7f-gate-receipt-e0824a23e0.md— tree e0824a2, quick tier 8339/8339, tree-reader set 16338/16338,make gateGATE_RC=0. Earlier b7c/b7e runs were RED (TMPDIR; a load-sensitive test) and are not receipts.No gate, workflow, or CLAUDE.md change in the diff. Not armed: needs the missing receipts and two non-author seats at the head.
keep-open: none
ont-delta: none
🤖 Generated with Claude Code