Authorized: ODD_BRIEF.md and Thomas's execution request. Incremental Modal input compute cap: 14400 core-seconds. Profile teal-sea, one CPU per input, at most ten concurrent inputs, no detached app, no push. The present agent watches every owned app to a terminal state. Local work is limited to edits, reads, compilation and tiny sanity checks. Existing referee files are read-only.
Preflight: clean branch teal-sea/oob-odd-referee at 06fdc801. Fetched remote refs without integrating. The Orca worktree listing was used because the requested project list_sessions tool is not exposed in this session. Other visible oob lanes were done. Existing shared venv is linked locally and reports python-flint with both expected backends; the pre-push hook is installed. The full fast suite is not run on this host, per the brief's compute constraint.
Estimate before the pilot
The previous even referee measured 71.383 core-seconds for ten high-frequency 500-mode panels. This is a planning input, not a measurement of the odd code. First run one odd unit on panels 990 through 999 at the frozen main settings. Allow 180 seconds, one core, with a 165-second internal alarm, no retries. Measure it and replace the provisional extrapolation before any main batch. There are 100 main ten-panel units and 30 Zhu ten-panel units. The latter have 200 modes and lower precision. Reserve 1200 seconds for Zhu, 900 seconds for reduction and controls, 600 seconds for startup/verification, and a 30% reserve on the main extrapolation. Provisional total is 100*71.383*1.3+1200+900+600=11979.79 core-seconds, below the cap. Stop before the main launch if measured extrapolation exceeds the cap. Every unit checkpoints a manifest, budget, progress and exact-dyadic matrix. Quota refusal is a terminal blocker, not a mathematical failure.
Pilot and measured dispatch estimate
Source freeze commit 3ce6ab9. App ap-F5KLwbmFxNF1jLkY7PfYBU completed the main high-frequency pilot, panels 990 through 999. Manifest runtime: 72.40257549285889 core-seconds, including serialization and checkpoints; assembly progress at panel ten was 60.388236847 seconds. No numerical gate failed. No quota block was encountered.
Recorded before any full main dispatch: 100*72.40257549285889*1.3 plus 1200 for Zhu, 900 for reducers/controls and 600 for startup/verification is 12112.334814071656 core-seconds, or less than 3.365 core-hours. This includes the successful pilot. Proceed with Zhu's 30 units and reduction first, then the remaining 99 main units if observed runtime still fits the cap.
Local tiny hygiene run: 29 passed, five deselected, 3.84 seconds. The context index is current. Contribution structure check fails on existing hunt-level omissions: MISSION lacks a huntspec block, and hunt-level RUNS and RESULTS do not exist. None was changed by this review. The narrower referee documents and evidence remain within the authorized write scope.
Zhu validation completed; main dispatch
App ap-XXYjuxO0XEJyrQxtC3FKhQ: all 30 Zhu assembly units completed, 651.6811525821686 core-seconds. App ap-FRG5XG7ywfOtMwX6OlcjI9: reducer completed in 42.71830892562866 core-seconds, passing the positive endpoint, quadrature moments, pole normalization, matrix hashes and tail gates. Both local entrypoints exited zero. Total recorded work including pilot is 766.8020370006561 core-seconds. The original planning allowance still covers all remaining work. Before main launch, an updated conservative estimate is 766.8020370006561+99*72.40257549285889*1.3+900+600 =11585.013502931595 core-seconds, below 3.219 core-hours. Launch the 99 remaining main units with the byte-identical source and exact inputs.
Main assembly completed
App ap-oyXeqoiRrDLKIDCKlkzXlj completed all 99 remaining main units. Together with the pilot, all 100 main blocks are present. The client emitted one heartbeat deadline warning, but continued returning successful units and exited zero; no input failed or was retried. Total recorded work through all assemblies and Zhu reduction is 8197.112294197083 core-seconds. One main reducer has at most 900 input seconds available, leaving the total below 9098 core-seconds before conservative app-occupancy accounting.
Final result, counts, and terminal states
Main reducer app ap-MDXpk3KteZl2PtytdCXKFA completed in 255.8151834011078 core-seconds. The safe endpoint 2.99e-44, the actual pole-sign mutation and the negative-operator control all passed. Independent artifacts were frozen locally in cefd6363 before reading three worker comparison artifacts at 08ff0491. No worker Python was read. The measured reduced-form value agrees at the worker's displayed precision; the independent lower shift remains the one fixed before execution.
| app | task | terminal state | maximum allocated cores |
|---|---|---|---|
| ap-F5KLwbmFxNF1jLkY7PfYBU | main pilot | stopped, zero tasks | 1 |
| ap-XXYjuxO0XEJyrQxtC3FKhQ | Zhu assembly | stopped, zero tasks | 10 |
| ap-FRG5XG7ywfOtMwX6OlcjI9 | Zhu reduction | stopped, zero tasks | 1 |
| ap-oyXeqoiRrDLKIDCKlkzXlj | main assembly | stopped, zero tasks | 10 |
| ap-MDXpk3KteZl2PtytdCXKFA | main reduction and controls | stopped, zero tasks | 1 |
Counts: 132 completed inputs, zero failed inputs, zero inconclusive inputs, zero retries. These comprise 130 assemblies and two reducers, 1300 panels, 250900 CC node evaluations including repeated endpoints. Two positive factor proposals and one expected negative-control refusal. Two controls passed. No quota block. One early read requested a not-yet-written result and returned file-not-found; it was a read, not a numerical input or failure.
Recorded input work totals 8452.92747759819 core-seconds, 2.348035410443942 core-hours. Manifest elapsed time is recorded before the final manifest commit. For a separate conservative occupancy bound, the complete app lifetimes times maximum cores sum to 78+91*10+50+838*10+262=9680 core-seconds, less than 2.689 core-hours. Both are below the 14400-second authorization. All states, durations and package versions are saved in execution_summary.json and modal_terminal_apps.json. Container Python image series is 3.12, with python-flint 0.9.0, mpmath 1.3.0 and gmpy2 2.2.1. The container Python patch version was not logged. Modal client is 1.5.5; local Python is 3.13.14.
Seventeen deciding artifacts were downloaded. Their compressed and uncompressed hashes, plus every numerical source/input hash, match the unit manifests. The final head/factor and all source and exact inputs are local; intermediate panel matrices remain at the named volume paths and are covered by locally saved unit manifests. No heavy local numerical run, GitHub Actions, push, PR or publication occurred.
Final local checks
The affected fast hygiene selection passed again: 29 passed, five slow checks deselected, 4.44 seconds. Python compilation and git diff --check pass; CONTEXT.md is current. Nine reported outward error bounds were checked against their saved printed balls with directed decimal arithmetic. Exact rational comparison confirms the chosen independent endpoint lies strictly below the worker's exact endpoint, and both displayed Rayleigh comparisons agree. The composite arithmetic and ordinary-proof grades are unchanged.
The initial secret-scanner call omitted the required ref argument. A second call incorrectly supplied a directory instead of a ref and is not counted as a valid scan. The final check uses the real committed range 06fdc801..HEAD; its output is final_secret_check.log. These CLI invocation mistakes did not execute any numerical inputs. No out-of-scope repair to repo tooling was attempted.