Repository navigation
feat(cli): install verified prebuilt releases - #11
Merged
Merged
Conversation
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars.
 |
Owner
Author
|
Validation update (2026-10-29): the new dependency constraint fix makes The Fedora Release matrix preflight on Fedora found the native Linux, macOS/Windows Zig, and iOS targets ready; Android NDK and wasm-pack are not installed yet. No release was published. A clean v0.6.2 release still requires the complete release matrix and a no-Rust download/install/E2E check. |
Owner
Author
|
Validation update (2026-10-29): latest PR head |
1 of 2 tasks
aliasfoxkde
added a commit
that referenced
this pull request
Nov 1, 2026
…11 integration (#13) * docs(planning): burst-compute queue — T5500 down, T3.4 fired T5500 powered off (idle, no training). Burst queue: T3.4 4B arm on Kaggle (trigger fired: 2B plateaued at 0.6840 < 0.75), JEV-27B probe on Modal or 30 min at next T5500 boot; consensus pass resolved as already-done on fedora 5060 Ti. Timeframes per item. Co-Authored-By: Claude <noreply@anthropic.com> * feat(bench): jev27 contamination-audit arm + Modal burst (B2 staged) autotrust/JEV-27B on JevBench public 231 through their own serve_decide.py POST /v1/decide (thinking off = pure System-1 audit). Runner gains --arm jev27 (native probabilities, never renormalized, noul false/true -> no/yes, failures score incorrect). Modal H100 burst app with weights volume + in-process replay is staged but blocked: the workspace spend limit refused the first launch (no cost incurred); raising the cap is operator action. Also: trainer v6 (in-loop HF resume-state upload after every ckpt + val-plateau early stop) uploaded to the t34 corpus repo; queue doc B1 session log records the session-1 12 h kill, the output-harvest recovery, and the v6 rollout plan. Co-Authored-By: Claude <noreply@anthropic.com> * feat(bench): d1-2B survey arm — JevBench-231 0.7489 via /v1/systemone LiquidAI/d1-2B (lfm1.0, extraction-teacher-only) through llama.cpp v0.8.0's native decision wire: 0.749 acc / 0.762 macro / ECE 0.065 / Brier 0.279, 231/231 valid, 0 failures, p50 5.9 s. Second-best 2B-class row (decider-2b 0.7576), best ECE of any 2B arm. Arm `d1` added to the runner (noul/choice/score native rendering, probabilities never renormalized); JEVBENCH.md arm 8 + results/anchor rows; B6 recorded in the burst queue doc. Co-Authored-By: Claude <noreply@anthropic.com> * feat(calib): per-kind T beats global 2x on E1-E board perkind_holdout_fit --probs on e1e-slot-231: board ECE 0.1305 (global T=3.362) -> 0.0678, auto@5% 0.039 -> 0.208. In-sample sizing probe; holdout-side exact fit is the ship step. Fixes ref path (datasets/public). * docs(oper): T3.4 session-3 record — cap kill at 4300, ckpt@4000 safe v6 upload path validated under a real 12 h kill. Next session blocked on Kaggle weekly quota reset; resume from step 4000 is automatic. * docs(calib): out-of-sample correction retracts per-kind claim Suite-fitted choice T (6.144) lands board ECE 0.1511, worse than the shipped global 3.362. Per-kind 0.0678 was mostly in-sample. Per-kind shape and fit-per-surface rules survive; cross-surface transfer does not. * feat(bench): trainer v6 partial manifest + T3.4 session-3 verify The manifest previously existed only after the training loop, so a cap-killed session lost its data-quality counts (loud over-max-len skips, bad slotprobs) and val history — fields recorded nowhere else. build_manifest() is now shared; save_resume_state writes a manifest_version *-partial copy at every checkpoint and uploads it to ckpt/train-manifest.json beside the resume state. Also: session-3 output verified (resume-state step=4000 byte-identical to the HF copy, val 0.0260->0.0119 still improving); the v6 trainer committed to the repo for the first time (was only in /var/tmp + HF); quota re-probe still blocked (30 h cap, resets after the Fri/Sat midnight-UTC boundary per community reports). Co-Authored-By: Claude <noreply@anthropic.com> * docs(research): §33 jevgrep re-read + Tev1-4B audit; d1 teacher runner Both production decision models ship uncalibrated (raw prob at fixed 0.5); Tev1 letter-only SFT loses to Jev on their own honest rerun — recorded as D15 evidence. Also commits d1_teacher_pass.py runner for the merged-v4 quarantine-pool teacher pass. Co-Authored-By: Claude <noreply@anthropic.com> * fix(bench): teacher pass survives transient server failures HTTP 5xx/timeouts were recorded as completed rows, so resume permanently skipped never-voted items (162 poisoned in the first run). Transient failures now stay pending for the next resume; permanent parse failures are still recorded so a malformed item cannot loop forever. Adds --retry-failed to strip failed rows from an existing output file. Co-Authored-By: Claude <noreply@anthropic.com> * feat(bench): 4-worker concurrent teacher pass Single-stream voting left 3 of 4 llama-server slots idle; workers fill them. Resume semantics unchanged (append-only, completed-id snapshot). Smoke: 16/16 ok on 4 workers. * docs(oper): record teacher-runner concurrency scope decision * docs(oper): d1 teacher-pass throughput finding — full stream host-blocked * docs(oper): GitForge CI trigger retry-storm diagnosis Orphaned local-only commits retried at checkout (retryable=false) left pending trigger rows that starved the pipeline queue; engine redrive via fresh push recovers. Permanent-fail on object-missing checkout remains an open ci-service defect. * fix(engine): rename betacf bindings for clippy similar-names qab/qap/qam trip clippy::similar_names under -D pedantic; renamed to sum_ab/a_plus_one/a_minus_one. No behavior change (251 lib tests pass). (cherry picked from commit ca8d64bae68240dad11bfdcdbb931edb7d8c231f) * fix(deps): pin internal config-crate path deps opencodifier-config was the only crate with bare path deps on core and engine; cargo-deny wildcard gate fails supply-chain. Pin to 0.6.0 to match every other crate. * chore(ci): redrive stalled pipeline triggers for e77105f/c6b3b1a * docs(oper): CI lane green on 03ed530 + lock-storm defects Co-Authored-By: Claude <noreply@anthropic.com> * chore(ci): redrive stalled pipeline triggers (e77105f/c6b3b1a/03ed530/e28f2b2) Co-Authored-By: Claude <noreply@anthropic.com> * chore(aegis): regenerate baseline for line shifts since 9cc08e0 The supply-chain gate reported 753 stale-fingerprint findings on 70f8d6f: files shifted after the last regen (which predated 9cc08e0), so every fingerprint went line-anchored stale. Fresh in-tree regen (default profile matching the gate), inspection_ledger trimmed before commit. Gate now reports 0 new findings. Co-Authored-By: Claude <noreply@anthropic.com> * test(engine): cover rotation combine and numeric edge branches The coverage floor fell to 96.28% on 70f8d6f because the rotation readout landed with the ArithmeticMean combine path, the incomplete- beta reflection branches, the ln_gamma x<0.5 branch, the binomial CDF saturation returns, and normalize's collapse refusal all unexercised. Four tests pin each branch to a known value or a cross-branch identity. Co-Authored-By: Claude <noreply@anthropic.com> * test(cli,engine): lift coverage floor over 97.5 in-image Init-doctor probes read hardcoded host paths inside the decision, so their missed branches were untestable; extract the cores to take their environment as data (probe_cpu_from, probe_ram_from, gpu_finding, server_finding, artifacts_finding, doctor_report) and cover all branches with deterministic tests. Add the over-ceiling image branch in decide, the local-2b lane proposal branch, boolean/score escalation partial answers, plus rotation/lane tests. Aegis baseline regenerated in the same commit (fingerprints are line-anchored), ledger trimmed. In-image floor check: 22032/22584 = 97.56% (floor 97.5). Co-Authored-By: Claude <noreply@anthropic.com> * docs(oper): CI triage runbook, in-image coverage rule, lane-fix debt Record the REST route that actually lists pipeline runs (GET /api/pipeline-runs?repo_id=) after the obvious guesses 404/405, and land scripts/ci-run-lookup.sh so the lookup is a single curl+jq command instead of rediscovered each session. Document the in-image coverage measurement recipe: the floor is measured in opencodifier-ci-rust:2, and a local rustc run reads ~8 points lower on the same tree, so test targets must be chosen from an in-image measurement. Cross-ref the GitForge per-ref trigger lane fix (14de7171): committed to gitforge-ci/main but not deployed -- the running release 69245e2b predates it. The retry-backoff and republish attempt-reset defects remain unfixed and are recorded here with the owner and what "fixed" means. Co-Authored-By: Claude <noreply@anthropic.com> * chore(aegis): regenerate baseline for docs line shifts in 6471ad7 The CI triage runbook and coverage-rule doc edits added fenced code blocks to docs/COVERAGE.md and OPERATOR_ACTIONS.md, shifting line-anchored fingerprints. Ledger units[] trimmed per the standing regen rule. Gate re-verified rc=0 locally. * chore(aegis): regenerate baseline for AEGIS_AUDIT doc section shifts The appended baseline-regeneration section in AEGIS_AUDIT.md shifted line-anchored fingerprints and its own prose matched 5 patterns (markdown-code-fence, eval-usage, trivy-config, missing-limit). Regenerated with the gate's flags; inspection_ledger units trimmed before commit (138 MB -> 30 MB). Gate re-verified rc=0. --------- Co-authored-by: Michael Kinney <aliasfoxkde@gmail.com> Co-authored-by: Claude <noreply@anthropic.com>
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
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.
Summary
just ci, GitForge CI, and the GitHub mirror.Validation so far
sh -n scripts/install-prebuilt.sh, ShellCheck, Python byte-compilation,cargo metadata --locked, andgit diff --checkpass.cargo check --workspacepassed on Fedora 44 after the version bump.just ci, the GitForge pipeline receipt, complete release matrix, and v0.6.2 download/install/E2E remain pending.Release boundary
This PR prepares a clean v0.6.2 release. It does not claim the new release is published. The full GitForge gate and release artifact validation are required before publishing.