EPS
← All docs·Activity

Daily — 2026-07-30 (8 follow-up rounds landed; 18 fixes filed; 0 promotions)

updated: 2026-09-03

What happened

  • #1739 (behavior-prediction map horse race) ran end-to-end: all 3 GCP lanes terminal (1,906 cells), upload-verification PASS, interpretation + clean-result critics PASS, merged (7ac4098f45), parked awaiting_promotion. Thomas drove a heavy interactive day on it (963k-context map reuse directive, WildChat rung re-scope after a teammate caught a false row-identity premise, natpv round, nonlinear-map fan-out overnight).
  • 8 same-issue follow-up rounds completed + merged: #1768 (r2 matched-text-capture-noise-floor + r3 on-target-prefix-corpus), #1776 (operator_swap_success, p3p4, slot_patch_sufficiency), #1689 (derived-vs-free-answer-map re-sweep + wellposed-shared-readout; cheap-band r3 dispatched), #1738 (fu2 sae-arm + fu3 crossed-multiturn — all 3 user-approved follow-ups done), #1482 (matryoshka-tier), #1769 (fu1 alpha-subgrid), #1775 (dedup-refit) — all re-parked at awaiting_promotion.
  • The 07-29 daily's fix wave largely landed: ~15 workflow-fix/infra tasks merged to main today (#1813–#1828 band, #1835–#1852 band incl. the SLURM rsync-lane stranding fix #1835 and the thread-caps main-red fix via #1884's lineage).
  • New experiments in flight: #1887 (ridge λ-selection hardening, running), #1900 (leakage-predictor horse race at per-prompt grain, fellows lane), #1901 (mapping-quality metric characterization), #1902 (post-training stages reshape, planned + spawned).
  • Codex quota outage all day (sentinel live until 2026-08-06) — every doubled review site ran single-Claude per the no-show fallback; cross-family review diversity is off fleet-wide for ~a week. No clean-results were promoted today.

Applied workflow improvements

  1. Target: .claude/agents/implementer.md — what: clarified --map-files <diff-list-file> takes ONE newline-delimited path-list FILE (positional paths exit rc=2) — commit: e9a6c6068b Why: parked candidate on #1896 (implementer round 1 burned a retry rediscovering the flag shape via --help: "the first invocation exits rc=2 (argparse usage error)"). Verified: --map-files FILE per select_step9c_tests.py:212-214; 1 placeholder occurrence in the file. Applied edit: 
    - ... --map-files <diff-list> --repo-root "$WT"` (a SECOND selector call — ...)
+ ... --map-files <diff-list-file> --repo-root "$WT"` (a SECOND selector call — ...). `--map-files` takes ONE argument: a FILE containing the newline-delimited changed-path list (compose it first, e.g. `printf '%s\n' <paths> > /tmp/diff-list.txt`) — NOT positional paths, which exit rc=2 (argparse usage error).
  2. Target: .claude/rules/gotchas.md — what: billing-reads bullet: gcloud billing is PERMISSION_DENIED under the eps-gcp SA; use [email protected] — commit: b9b43a3eea Why: miner-4 P15 (session 5fb88cc0, 18:29Z): Thomas asked "How many credits are left on GCP"; the eps-gcp configuration read failed PERMISSION_DENIED and only the personal-account retry worked. Applied edit: 
    + - **`gcloud billing` reads fail `PERMISSION_DENIED` under the `eps-gcp` configuration** — ... run the read under Thomas's personal account instead: `gcloud billing accounts list [email protected]` (verified working 2026-07-30, session 5fb88cc0 ...)
  3. filed for review #1916: judge max_tokens floor → ≥600 for reason-then-score JSON rubrics (rule 23 measurably failed twice: #1739 at 400, #1769 at 300 with 12.5% censored draws, arm-asymmetric to 42.5%, full mt600 re-judge).
  4. filed for review #1928: git guards cwd/compound awareness — ssh-payload + cd-to-worktree false positives on guard_repo_root_branch.sh, compound pathspec-limited commits blocked by guard_root_code_commit.sh despite its own remedy text (9+ blocked turns today; #1710's 07-26 mask extension didn't cover these shapes).
  5. filed for review #1930: Step 9a-ter inline compute routing — ≥50 GB staged fit inputs route to cpu-bigmem AT DISPATCH (the #1345 fits died silently 4× on the VM over ~2.5h), and iterative-fit legs >15 min/cell route to GPU at dispatch (the #1768 MLP battery ran CPU-bound until Thomas ordered "just run on GPU").
  6. filed for review #1929: detached-launch discipline — relaunch re-runs the launcher FILE (a re-typed chain dropped #1768's completion sentinel), watches key on the WORKER pid (a launcher-pid watch false-fired EXITED on #1769's live re-judge), push-verify names expected paths (#1482's vacuous verify left 29 eval files uncommitted).
  7. filed for review #1919: plan-compute-sizing measurement basis — pilots bind to their own lane's production budget/shape (4 of 6 #1739 lanes halted on proxied walls), and GPU-util claims need a ≥10-sample window (a one-shot nvidia-smi "90%" was actually 12.6%; ~7h H100 at ~87% idle).
  8. filed for review #1921: four CLAUDE.md discipline clauses — absence claims need a relocation sweep ("fourth time today" self-admission, one false claim in a durable marker), "you can do X" after a cost quote = authorization to run (the declined 8.5k call), glossary retired-terms check for mapping writeups (#1768 Motivation misattribution), file-scoped teammate git-log probes (a 10-commit-window probe stood down a teammate whose commit had landed).
  9. filed for review #1924: Monitor-until-loop quiet-wait branch for the Step 6d.2 poll (parked on #1818 — the clamp leaves §7's ~1-wake-per-1800s savings structurally unrealized).
  10. filed for review #1914: widen the body-audit condition-label regex to H1c/H4b forms (parked on #1826 — sibling of the check-28 class it fixed for sidecars only).
  11. filed for review #1915: rsync_covered() checks RSYNC_EXCLUDE_PATTERNS (parked on #1835 — nested-excluded-dir false-PASS residual, a deliberate extension its plan fenced out).
  12. filed for review #1925: pre-commit hook for the agent-memory index-size lint (parked on #1891 — the direct-to-main regrowth channel is ungated; parity with #1661's agent-spec hook).
  13. filed for review #1926: SLURM intent-table completeness — lora-7b-h100 (+ class) absent like capture-7b was; add a completeness pin test tying gcp intents to SLURM rows (parked on #1896).
  14. filed for review #1927: canonical gist-update recipe (gh api -X PATCH + content verify; the EDITOR-trick gh gist edit silently no-oped rc=0 on #1769's mirror).
  15. filed for review #1917: bounded index.lock retry in front of set-status's deliberate fail-loud raise (a transient collision left #1815's status move half-applied; body carries the #898 deliberate-design context — planner may deflect to a documented reconcile recipe).
  16. filed for review #1918: a /daily route-1 fix that credits an open filed task archives that task in the same pass (#1823's spawned session booted only to discover mootness).
  17. filed for review #1920: living_docs.py bounded index.lock retry (raw traceback on a routine lock collision, #1901).
  18. filed for review #1931: pod bootstrap retry + fail-loud provision verdict (pod-1773-regsteer's bootstrap died exit 100 leaving /workspace empty, surfaced only at pilot launch).

Other problems & notes

  • Held (route 3, needs you): #1922 — triage the root-sync KEPT stash 319c2bf16e7c (rescue patch at ~/.task-workflow/root-sync-rescue/stash-319c2bf16e7c.patch, 9.3 KB; re-flagged by ≥6 sessions today and again by tonight's own stub push). Carve-out: destructive/irreversible (dropping a stash). Yesterday's #1870 covers the surfacing mechanics only.
  • Held (route 3, needs you): #1923 — re-add the VM's public SSH key to the RunPod TEAM account key list (provision preflight WARN; pre-injection pod resumes may refuse SSH). Carve-out: external side-effect on shared account settings.
  • Main-red ledger: the 12-node tests/test_issue1689_os_exit_shutdown_bypass.py red is STILL red on main (verified tonight: 12 failed / 81 passed) — fix task #1903 is filed at proposed; every intervening Step 9c gate keeps paying a pristine-oracle classification leg, so it deserves priority dispatch. The thread-caps red (test_no_new_torch_before_dotenv_vm_entrypoints) is GREEN tonight (fix lineage #1884/b9920c4a06); a NEW main-red (test_backend_poll.py::test_every_lazy_scripts_local_import_is_bootstrap_guarded, offender scripts/issue1092_build_corpus.py:2029, verified red tonight) was urgent-parked on #1852 and routed by the watcher's urgent-park router as #1912 minutes into this run. test_issue1481_analysis and test_step9c_baseline are green again.
  • Recurring injection vector flagged on #1864 (open since 07-29): analyzer direct-to-main script commits bypass the branch-side gate — the thread-caps red recurred today with a rotating cast of offenders (two #1739 intrusion-scan scripts landed AFTER the first fix). Posted a recurrence note on #1864 instead of double-filing; consider prioritizing it.
  • Step C sweep hygiene: 10 parked candidates enumerated, all dispositioned — 5 filed (#1924, #1914, #1915, #1925, #1926 with routed-records on #1818/#1826/#1835/#1891/#1896), 1 route-1 self-applied (implementer.md, entry 1 above, record on #1896), 1 already watcher-routed (#1852's urgent park → #1912, record posted by the watcher itself), 2 deduped against #1884 (records on #1831/#1835; the shared failing test is green tonight), 1 deferred (the #1771 architectural-greenlight mechanization stays parked — #1771 itself is still plan_pending and is rewriting that exact gate; re-evaluate after it merges). 1 benign skipped row (json-decode on #1333, relevant_kind: false — known shape); unmatched_record_fps advisory ~300 refs (the #1703 standing condition, grows with each day's routed-records).
  • Retraction re-check: 6 marker-mined candidates scanned fresh at filing time (source tasks #1818/#1826/#1835/#1852/#1891/#1896); 0 retracted; the scan is what caught the watcher's just-posted #1912 routing and prevented a double-file. Miner-sourced items were tail-checked by the miners against session self-corrections.
  • Already filed by their own sessions (no /daily action): #1892 (GCP PENDING provisioning-model misread), #1893 (code-sync vs live-run race), #1897 ("already merged" false success — PR-state probe), #1908 (pytest test-order leak), #1911 (pgrep -c || echo 0 double-print — the ~11-idle-H100-hours near-miss on #1773's release watcher), #1845/#1877 (HF Xet wedge gotchas), #1882 (TG file-grain gate false positives), #1886 (SLURM rsync gotchas).
  • Skipped filings (family already covered): branch-vintage gate staleness cost ~4 extra full gate runs (~30-45 min each) across #1775/#1896/#1769 plus #1776's 4-run and #1768's 5-run merge sagas (~7h on #1768) — the freshness family is already dense (landed #1807 today + open #1882/#1883/#1897); re-evaluate tomorrow if the tax persists post-merge. gcp.fetch_results erroring on TERMINATED instances is documented fail-soft design (log-loudly-and-continue with the HF partial prefix as recovery) — noted, not filed. Fellows lane-staging gaps (3 job crashes) are covered by today's landed #1835 + filed #1886/#1915 — verify tomorrow that both the #1689 and #1768 shapes are closed.
  • 529 Overloaded storm ~08:53–09:21Z killed turns across ≥6 sessions (two autonomous sessions ended mid-tick, one implementer's final report, a clean-result critic died silently ~101 min); all recovered via tick crons / watcher respawns / durable-marker reconstruction. One usage-policy refusal each on #1776 and the #1773 session (single-kill, recovered next tick; one forced a Fable→Opus switch while paging raw LMSYS text — the documented digest-only trap).
  • Near-misses worth remembering: the #1739 wildchat-rung brief asserted 2,400 rows were "real WildChat crossing rows" (they were #594 synthetic battery rows) — the teammate's pre-spend STOP-AND-CHECK prevented a wrong results column; the #1689 within-model battery's checkpoint keys omitted the MODEL and double-counted to a fake "504/504" coverage pass (fixed in-session, d63ec44ac7); the #1768 p8 fit ran its plan-named 8-GPU axis serial on 1 of 8 H100s (~$375 idle width; #1873 daily-held from yesterday covers the review).
  • Guard/idiom recurrence counters (firings, echoes excluded): ~20+ hook blocks today across piped git commit/push (~10), repo-root reset/rebase/merge/restore (4), repo-root pull (2), foreground sleep-chains (4), root-code-commit cert ordering (5) — every block recovered next call; 8 of one day's 11 in the #1739 interactive session alone. TaskOutput misuse ×5 (unloaded schema ×3, name-keyed ×2). Shared-root index.lock collisions ×7 (all absorbed by bounded retry/defer except the #1815 set-status case → #1917).
  • #1310 lost activation store (run-2 script-format store died with its instance — a pre-existing persist-by-default violation surfaced by the nd-audit): ~1 GPU-h recapture was dispatched inline by Thomas's directive; tomorrow's daily should verify its upload-verify ran BEFORE teardown.
  • /daily self: tonight's stub push was initially rejected and landed via sync_repo_root (state=synced); the same KEPT-stash residue as #1922 was reported. Miner infrastructure ran clean (7 parallel miners, 84 transcripts, 203 MB).
  • title-sync drift unchanged (8 rows: #432 #458 #654 #658 #664 #665 #763 #1226 — see logs/daily/2026-07-25.md). Read-only sweep exited 0.

Living-docs drift

  • living_docs.py check exited 1 with 199 findings (hash 078285c14444): ~190 completed-with-clean-result tasks missing from any question's evidence (ids #376–#549 band and onward) + stale spec-role-header State trailers vs results promoted 06-09..06-16.
  • drift proposal already open (same hash logged 2026-07-07/07-22/07-23/07-24/07-27/07-28); not re-proposing. No event-stream row appended (dedup rule 1a).

My thoughts

Highlighted results

  • no results promoted today