# SAIE-N Prime Core Release Note (v2) Release date: 2026-11-19 (revision v2a) Canonical page: https://www.shirania-branches.com/research/saie-n-prime-core Archive: `saie-n-prime-core-submission-v2.tar.gz` sha256 `a29d4edf54ff4fe66c36d0d352d2bf9283cd5a5905ace3f7466f4a0c6326a226`, 102,283 bytes, 112 files (111 MANIFEST entries plus `MANIFEST.tsv` itself). ## No new theorem validation is implied v2 is a repackaging. The mathematics, the claims, their scope labels (`proved`, `tested`, `fitted`) and the April 2026 evidence are unchanged. Rebuilding the outputs byte for byte shows that the package is complete and the scripts are deterministic. It does not strengthen any claim beyond what the April claims matrix says. ## What changed from v1 - **Self-contained.** v1 shipped the runner and 3 evidence files, but none of the scripts the runner calls, so it could not be run. v2 adds: - the 13 called scripts; - the 2 cited-only probes; - every cited result file, and every document the manuscript links to. The closure check (`tools/closure_check.py`, output in `CLOSURE_CHECK.txt`) passes. - **Relative paths only.** Script path defaults, the runner and doc links no longer point at the author's machine. The extracted tree contains 0 occurrences of the old local root. Every script change is a path change, and the diffs are in `CHANGES/`. The original local root is redacted there as ``. - **Evidence kept intact.** The v2 runner writes into a separate `OUT_DIR` and never overwrites the bundled results. `tools/compare_reproduced.py` compares the two. - **The comparison fails closed (v2a).** The first v2 candidate (sha256 `058380856b36…`) had a checker that returned exit 0 in two cases where it should have failed. One was an empty run log, or an empty figure expansion, with 0 files compared. The other was a logged file with no bundled copy (Inariael's review, 2026-11-19). That candidate also said "109 files" where the tree held 110, counting the manifest. `tools/compare_reproduced.py` now exits 1, with a message, when the run log is absent, empty or malformed; when a figure expansion is empty; when a bundled figure CSV is not reproduced; when a logged file is missing from the output or has no bundled copy; when any file differs; or when 0 files were compared. The runner also deletes each step's seeded copy before the step runs, so a step that exits 0 without writing can no longer pass as a match. The controls are in `tools/checker_controls.py`, and their build-time output is in `CHECKER_CONTROLS.txt`. - **Integrity.** `MANIFEST.tsv` lists the path, byte count and sha256 of every file, and `tools/verify_manifest.py` checks it. - **Omitted on purpose:** - the April outbound-messages / announcement draft, which is not evidence; - the v1 machine-local bundle builder. ## Verification run 1: full maths (2026-11-19, first v2 candidate, Python 3.12.3) This run was made on the first candidate. The v2a changes touch only the checker, the controls and the runner's pre-step deletion, not any script that computes. The archive was extracted into a fresh, unrelated folder. Every Python step ran under `python3 -I` with an open-audit hook. The rebuild command was `HIGH_RANGE=1 FORMULA_CHECK=1 INDEPENDENT_CHECK=1`. - verify_manifest: PASS (109 entries in that candidate), and PASS again after the run, so the bundled evidence was untouched. - closure_check: PASS. - The runner finished with rc 0 and a wall time of 328 s. The slowest step was the k=4 coupling map at n = 5,000,000, at 104 s. - **Reproduced and byte-identical: 46 of 46 files** that the run wrote. That is 26 probe/fit/simulation JSONs, 18 figure CSVs, the anchor-trend table and the table snippets. There were 0 differences. - Audit of opened files: 99 distinct paths. All are inside the extraction or the Python standard library, apart from the audit hook itself. None is in the original experiments folder, and there are 0 hits for the old local root, both in the audit and in the bash trace. ## Verification run 2: checker fix (2026-11-19, this archive, fresh clean extraction) - verify_manifest: PASS (111 entries), before and after the runs. closure_check: PASS. There were 0 occurrences of the old local root. - Base runner (no flags) under `python3 -I` with the open-audit hook: exit 0 in 6 s. `compare_reproduced.py reproduced`: **28 of 28 byte-identical, exit 0**. - `checker_controls.py reproduced` ran 11 controls, and every one behaved as expected (overall exit 0): - Must pass, and exited 0: the real base run, and a synthetic normal case. - Must fail, and exited 1: an empty run log; an absent run log; an empty figure folder; an absent figure folder; one figure CSV not reproduced; a logged file with no bundled copy; a logged file missing from the output; a differing file; a malformed log line. - Runner-level control: the runner was run with `PYTHON="python3 -I -c pass"`, an interpreter that exits 0 and writes nothing. The runner itself exited 0. Comparing its output then exited 1, with MISSING-REPRODUCED=10 and an empty figure expansion. - Audit of the base run: 88 distinct opened paths, none outside the extraction, the standard library or the hook. There were 0 hits for the old local root in the audit and 0 in the bash trace. - Not re-run on this archive: the full `HIGH_RANGE=1 FORMULA_CHECK=1 INDEPENDENT_CHECK=1` rebuild (run 1 above covers its maths). The pre-step deletion was exercised on the base path and in the no-op control only. **Existing evidence, bundled but not regenerated by the runner** (shipped as the April record): - `prime_channel_shadow_probe_results.json` and `prime_channel_slice_dimension_probe_results.json`. These are cited by the docs but not called by the runner. - `saie_n_prime_protocol_sim_results.json`. This is an earlier simulation run with no runner step. - `table_basis_costs.csv`, `table_context_threshold_scaling.csv` and `table_top_coupling_subsets.csv`. The table exporter reads these as static inputs. **Unrun:** nothing in the runner. Commands that the docs list historically but the runner does not call were not executed. The 2 cited-only probes were not re-run. **Limits of the audit:** strace was not available. The hook records Python `open`/`listdir`/`mkdir`-class events. It does not record `stat`/`exists` calls, and it does not record files the bash runner itself touches. The bash trace was grepped separately and showed 0 hits. ## Rebuild command See `saie-n-prime-core-rebuild-command-v2.txt`. From the extracted folder: ``` OUT_DIR="$PWD/reproduced_full" HIGH_RANGE=1 FORMULA_CHECK=1 INDEPENDENT_CHECK=1 bash prime-engine/run_all_prime_core.sh python3 -I tools/compare_reproduced.py reproduced_full ``` ## v1 [v1 (April 2026, the submission record, not runnable)](archive/saie-n-prime-core-submission-v1.tar.gz), with its [release note](archive/saie-n-prime-core-release-note-v1.md) and [rebuild command](archive/saie-n-prime-core-rebuild-command.txt). It is kept unchanged as the historical record; its sha256 (`28756022984adbb5c28828a30f5fdf6c621dfcee2787cb8bb672618f55a4b50e`) matches the copy served on the live site.