v1.1.0: additional analyses and released run records (September 2026)

v1.1.0: additional analyses (September 2026)

Every number in the revised manuscript was regenerated from these assets and the git checkout before this release; see REPORT_RELEASE.md for the per-script table. The two exceptions are those stated in the README: the 88 extractor calls of the two-extractor preliminary (outputs shipped under release/analyses/09_extractions/) and the feasibility-resampling diagnostic (outputs shipped as recorded).

Release assets

The run records unpack into one directory, admitor-v1.1.0-run-records/, with the host's outputs/ layout. Every record passed through release_tools/scrub_logs.py and was checked with scripts/secret_sweep.py and scripts/release_content_check.py.

File	Contents	Files	Uncompressed bytes	Zipped bytes	sha256
admitor-v1.1.0-run-records-core.zip	relabeled arms, collection, gate work order, E1 certification runs (specifications, candidate programs, verdicts), E3 calibration runs and labels, runtime event streams	2,330	849,470,251	93,940,328	ff5f80a5bc6a26ce037cbe0e34c7c03a108a81b571f79ceb902561d4bff176bb
admitor-v1.1.0-run-records-eval.zip	E1 evaluation trajectories, four arms x five plates	20	735,883,632	100,333,853	d72c2ad67195f6c0151c102da9748495cdd6adc40a689b3ac2affaeb2deaa5ba
admitor-v1.1.0-run-records-e0.zip	E0 host-reproduction trajectories and the 17-item OptMATH recheck	6	173,263,054	22,242,226	4510871a975a9943e541890ba9c900f48e9cdd14c47e05ca6dbfb302cc5269be
admitor-v1.1.0-run-records-cache.zip	E1 evaluation response cache	40,208	162,549,387	71,862,628	f65e9541e6f4d209a59aae6e8b617103675d80e8a7b11934281961c17885127e

The host run ledger (93,909 records) remains the v1.0.0 release asset and is unchanged: admitor-v1.0.0-runlogs.zip, sha256 272d52030e1ab229547e0762b7c25f954bcfc22e6e2586eeb82bcbec682a4c0f.

New scripts (scripts/)

Additional analyses: artifact_inventory.py, e3_certificate_numbers.py, code_facts.py, runlog_phase_audit.py, gate_arm_composition.py, poison_provenance.py, serving_variant_join.py, panel_base_judges.py, candidate_crosstab.py, task3_additions.py, assemble_decomposition_report.py, corrected_bootstrap.py, format_bootstrap_report.py, retrieval_stats.py, retrieval_final.py, disjointness_audit.py, numeric_coverage.py, k4_restricted.py, two_extractor_run.py, two_extractor_agreement.py, feasibility_resampling.py, admitor_reanalysis.py.

Release hygiene: secret_sweep.py, release_content_check.py.

Changed: e0_scorecard.py (--eval-root fills the reproduced column from the E0 run records and prints the macro row); release_tools/scrub_logs.py (a Windows path must have a real path shape, so escaped code text is no longer rewritten; output on the v1.0.0 inputs is byte-identical).

Each script's module docstring states its inputs (in git, run records, ledger), the exact reproduce command, its outputs, and the manuscript location of its numbers. requirements-analysis.txt records that nothing beyond environment.yml is needed.

New files in git

• release/analyses/: every analysis output (calibration table, judge decomposition and the five derived judges' verdicts, bootstrap, retrieval, serving-variant join, provenance, numeric coverage, restricted K4, two-extractor preliminary with the 88 extractions and their call ledgers, feasibility resampling, admitted id lists) and REPORT_analyses_2026-09.md.
• release/case_packets/: the 22 K3 case packets; release/k3_attribution.csv points at them.
• release/e0/scorecard.json: the reported column of Table 9 and the plate of each row.
• paper/make_fig1.py: regenerates Figure 1; figures/fig1_overview.pdf is the regenerated figure (the PNG copy is removed).

README

• New sections: "Additional analyses (September 2026)", "Model pins and the response-side audit", "Host backbone availability", and "Reproducing the paper's numbers" (one row per script with its inputs, command, outputs and manuscript location).
• Corrections checked: the README was searched for a three-candidate informative rule, a cross-family description of the majority-vote arm, ReLoop as a pipeline component, a library built at the calibrated threshold, a claim that every call was served by the pinned backbone, and a gpt-5.1 pin. None occurred outside the replaced section, so no sentence needed correcting. The new sections state the admission-time rule of the E1 library and the calibrated threshold tau = 33.3 as separate rules (judge table), the majority-vote judge as a vote over host samples, the gpt-5.4 pin set through ADMITOR_MODEL_B, and the response-side audit counts.
• Bootstrap intervals updated to the manuscript's values: gate minus vote +3.53 [+0.87, +6.68], gate minus ground truth +4.46 [+2.34, +6.65].
• Installation documents the pinned environment (environment.yml: Python 3.12.12, numpy 2.4.4, scipy 1.18.0, Pyomo 6.10.0, HiGHS 1.15.1) in which the outputs were re-verified.

AdmitOR v1.0.0

First public release of AdmitOR, the code and artifacts behind the paper
Admission Without Answers: Label-Free Certification and Experience Learning
for LLM-Based Optimization Modeling (Lian, Chen, Qin, Teo).

Paper: arXiv:2608.15565

AdmitOR certifies answers to optimization problems without an answer key.
Three model families write solver code against one shared extracted
specification, each candidate is re-examined over resampled instantiations of
the parameter domain rather than the stated instance alone, agreement is
compressed into a cross-family clique over the induced value functions, and
admission is thresholded to a conformally calibrated false-discovery budget
with explicit abstention.

What is in the repository

• admitor/ the certification engine
• scripts/ the E0, E1 and E3 experiment drivers
• datasets/ the problem sets the experiments run on (third-party, attributed
  in datasets/README.md)
• artifacts/ certification verdicts, the four evaluated skill libraries, the
  E3 calibration and replay reports, and the K4 contingency table
• patches/ our changes to the OptSkills host, as unified diffs; no host
  source is redistributed
• tests/ an offline suite that passes with no API key

Full index with sizes, sha256 and the paper table or figure each artifact
backs: artifacts/MANIFEST.md

Release asset

admitor-v1.0.0-runlogs.zip is the complete scrubbed per-call run ledger,
93,909 records covering 60,000 host model calls and 589,902,200 tokens. It is
attached here rather than committed because the ledger grows with every
campaign and must not accumulate in git history.

sha256  272d52030e1ab229547e0762b7c25f954bcfc22e6e2586eeb82bcbec682a4c0f

The same hash is recorded in artifacts/MANIFEST.md, so the manifest and this
release cross-verify. Every record was scrubbed by
release_tools/scrub_logs.py (committed, so the process is auditable) and
then swept a second time after extraction by an independent implementation.
Both agree: zero credential, endpoint, path or identity hits across all
93,909 records.