EMCA / agent_arch — publication package, campaigns v1–v20

Corrected preprint and reproducibility archive — publication version 2: Zenodo — DOI: 10.5281/zenodo.22772224. This version incorporates errata 10 and 11, updates the PDF and checking scripts, and corrects both language versions of the announcement. Frozen experimental code, reports, preregistrations and raw results are unchanged. The Zenodo record links to the exact Git commit contained in its ZIP archive.

The original publication remains available at DOI: 10.5281/zenodo.22762848, with the original archive at commit 678dee6. The all-versions DOI resolves to the latest published version.

This folder is a packaging of work already done and already frozen. It adds no new measurement, moves no threshold, and softens no result. Every number in the section documents is quoted from a report that is itself in reports/, and every report is backed by raw cells in evidence/, a preregistration in preregistrations/, and a verification log in evidence/results/.

The preprint itself is paper/emca_preprint.pdf (29 pages, LaTeX source paper/emca_preprint.tex, build + independent verification paper/build_paper.sh). It is the publication form of this package: abstract, introduction, related work, method, results (v1--v9 abbreviated as foundation, v10--v20 as the main body), external tests, formal verification, independent convergence, discussion, limitations, references. The title page carries the two authors (Oxunjon Ubaydullayev, Aiodam) and an author-contributions statement — see erratum 9 in ERRATA.md, which reverses the earlier "no mention of the tooling" instruction; the body of the paper still describes the experiment and its results, not the tool.

If you are reading this for the first time, read in this order:

1. 00_OVERVIEW.md — what the whole thing is, in one page, with the honest headline of each campaign.
2. sections/ — six sections, each self-contained:
   • 01_v1_v9_causality_vs_economics.md — the founding line, kept short (it is the ground everything else grew on).
   • 02_external_tests.md — the honest re-assessment of scientific novelty against published literature (Sachs 2005, causal bandits, the union mechanism), with exact citations.
   • 03_formal_verification.md — the Lean 4 proofs: what is proved, what remains a conjecture, and the honest boundary — including the turn-154 close of the modelling-identification item (§5.1: the identification is a conditional, hgood is an independent premise, and at the agent's own parameters the bound is vacuous).
   • 04_v10_v16_safety_line.md — the central, strongest part: wayfinding-shaped harm, the purchasability of built-in values, the ledger and its hole, the truth-does-not-select-action collapse, the blast-radius boundary, the bought auditor (v17), the world-side boundary (v18), the 30-seed replication, the learning attacker (v19), and the bought enforcer (v20).
   • 05_independent_convergence.md — the NVIDIA/Rutgers (Che & Wu, 2026) convergence, with the exact status of that citation.
   • 06_limitations_and_honesty.md — every declared limit, every refuted prediction, every defect found in the author's own work.
3. REPRODUCE.md — how to re-run everything from scratch, one command per campaign, with the exact environment and the layout the scripts expect.
4. MANIFEST.md — the file inventory with sizes and SHA-256 for the frozen artefacts.
5. PACKAGING_NOTES.md — what was copied verbatim, and the two path-only edits made so the package runs from its own folder.
6. ERRATA.md — read this before quoting anything. Ten over-statements in this package's own section documents: four found by the owner's re-reading in turn 151 and corrected here, two found in turn 152 — v16's withdrawn absolute headline (caught by the author's own fresh-seed replication) and the author's own preprint verifier reporting a catch it did not make — one in turn 154: a preregistered prediction (v19's HQ4d) refuted by the author's own code defect, with eight further defects of the checking layer recorded alongside it, and one in turn 157: v20's preregistered invisibility claim (HV4's "refused is False in both") refuted by the author's own measurement — the flag does change; the invisibility is a property of the agent not reading o["scope"]; and two in turn 160, both found by recounting from the raw data rather than reading it: the replication count "26 of 27" was inflated by an analyzer line counting its own meta-row as a verdict — recounted from the raw cells, the honest count is 25 of 26 — and the abstract's "learns its world to within 10 % of an oracle ceiling" was 11.3 % above the ceiling, not within 10 % (0.0364 against 0.0327). The frozen reports themselves are unchanged.

Directory map

folder	what is in it
paper/	the preprint — emca_preprint.pdf, its LaTeX source, the build script and the independent verifier
sections/	the six section documents the preprint was written from
reports/	every results/report document, verbatim, as frozen (RESULTS_*.md, CAMPAIGN_RESULT.md, EPISTEMIC_VERDICT.md, TURN*_STATUS.md, …)
preregistrations/	every PREREG_*.md, written before the first cell of its campaign
code/	every producer: environments, agents, runners, drivers, analyzers, oracles, independent verifiers, factchecks, run_all_*.sh
evidence/results/	the raw matrices (matrix_*), oracle outputs, independent-pass outputs, factcheck outputs, diagnostics
external/bench_cb/	the causal-bandit external test: ported published code (ext/), Lean proofs (lean/), matrices, verifiers
external/bench_ext/	the Sachs-2005 external test: data (ds/), measurement scripts, head-to-head against published methods
sources/	the registered external sources cited by the reports (see sources/README.md)

Plus: REPRODUCE.md (how to re-run everything), MANIFEST.md (size + SHA-256 for every file), PACKAGING_NOTES.md (what was copied verbatim and the two path-only edits), verify_package.py (the package's own checker, with two live negative controls), and make_manifest.py.

What this package is NOT

• It is not a new result. Nothing here was measured for the first time in the packaging turn.
• It is not a claim about any frozen line it builds on: v10–v20 are verdicts on their own instruments, never on the frozen v1–v9 line, and the reports say so in their first paragraphs.
• It is not a novelty claim. Section 02 and section 05 exist precisely to put the published prior art next to the campaign's claims.
• It does not hide failures. Refuted predictions, checks that could not fail, a verifier that over-claimed, a first-pass matrix deleted rather than kept — all of these are in the reports and collected in section 06.