Synthetic primary study

Equations in these preserved Markdown records retain their TeX notation where present; this is not a newly typeset proof edition.

Frozen synthetic pair-token study and historical P14 replay

The signed/permuted synthetic primary cell was fixed under protocol v4, then measured on 64 held-out formulas. It failed the unchanged v3 utility rule: median direct-packed/CM production time ratio 0.2479, with a 95% hierarchical bootstrap interval [0.2462, 0.2495]. Direct packed evaluation was about four times faster in this cell. All four assignments agreed on every formula; all CM outcomes were pure structural. The median process peak-working-set difference was -1.56% for CM under the declared metric, which is dominated by shared interpreter and import memory.

The run is a post-import, cache-reset token-output comparison on unshared two-variable formulas with N=U=1023; it does not test natural circuits, ABC/AIG, sharing, prepared reuse, dense output, or process-start costs. The original protocol v3 remains preserved. V4 was disclosed before held-out timing because no ABC executable was available for the full circuit arm. The eight-formula pilot used disjoint seeds and did not enter the primary analysis. Post-run clarifications record the exact without-replacement sign sampling and an environment-capture limit without changing the frozen analysis.

Result files

• Run freeze: hashes of the fixed protocol, scripts, 64-case corpus, configuration, environment, and initially empty result files, recorded before timing.
• Primary analysis: the frozen median-ratio rule, 10,000-draw hierarchical interval, memory statistic, and per-formula results.
• run_v4_001/TIMING.jsonl: 3,840 append-only arm/repetition rows (64 formulas × 30 repetitions × 2 arms).
• run_v4_001/MEMORY.jsonl: 128 separate fresh-process arm/formula observations.
• run_v4_001/FAILURES.jsonl: empty; no recorded failure.
• Independent recount: raw-row time and memory point estimates and run counts checked separately from the frozen analyzer.

The source implementation remains unchanged at commit 0ab8ffd0c23ffa71ee951d375b1c170ffdcc084b; its 63-file snapshot and transitive dependency lock are in ../02_CONTRACT_REVISION_20260926/reproducibility/.

Reproduce on Windows

Keep this directory beside 02_CONTRACT_REVISION_20260926 in the CM Computation project. Run these commands in a copy of the project so the delivered files and hashes remain an immutable record. Python 3.10.11 and the pinned NumPy 2.2.6 were used for the measured run.

cd 'CM Computation\03_STUDY_DESIGN_20260926'
python -m venv .venv
.\.venv\Scripts\python -m pip --isolated install -r ..\02_CONTRACT_REVISION_20260926\reproducibility\requirements-checks-lock.txt
.\.venv\Scripts\python verify_v4.py
.\.venv\Scripts\python verify_p14_runtime.py
.\.venv\Scripts\python reproduce_v4.py --output .\replication_01

The last command is a new sequential local timing run. It creates a fresh child directory, copies the frozen inputs, checks their hashes, executes the 64 cases and memory observations, and writes a separate analysis. It does not overwrite run_v4_001. Allow roughly ten minutes on hardware comparable to the recorded Windows host; timing values can change on another machine, while correctness and the analysis procedure should reproduce. No paid compute, network service, or production code change is required. A separately frozen ABC executable and admission corpus are now in 04, with bounded natural secondary timing in 05; these do not complete the full v3 circuit study.

Historical P14 reproduction

The independent P14 bootstrap replay reads the preserved 51,102 timing rows and reproduces all 17 archived interval sets using the recorded 10,000 draws and seed. To rerun that analytical check in a project copy:

.\.venv\Scripts\python replay_p14_intervals.py

The original P14 runtime tree was located at [local research path omitted] The [recovery manifest](P14_RUNTIME_RECOVERY.json) identifies 2,035 packaged files, including all 14 natural BLIFs with holdout-matching SHA-256 hashes. [The portable candidate ZIP](P14_PORTABLE_RUNTIME_CANDIDATE.zip) is independently hash-checked by verify_p14_runtime.py`. The located code imports successfully. A full invocation of the archived analyzer was attempted and timed out after 600 seconds; the standalone numerical bootstrap replay above is the completed analysis check. The historical freeze hashed its six top-level scripts and two holdouts, but not the complete imported runtime. A compatible local tree therefore does not establish exactly which runtime bytes produced the September 18 timing observations. The identity-evidence checklist states what contemporaneous record would resolve the gap and what to report if none exists.

No public archive identifier or license/provenance clearance is claimed. The September 22 audit score is a historical assessment of that older draft and has not been reassigned to this study.