RXRE engine

TL;DR — you can stop reading here. RXRE is an AI-assisted reverse-engineering research project. It has produced one strong artifact type: sealed, zero-tolerance, byte-exact reconstructions of specific GNU coreutils 9.10 binaries (e.g. filenvercmp inside ls, and wc — both now reproducible by anyone). Third-party independent verification has not happened. You can verify the entire claim for yourself with one command (evidence/r3/verify.sh, about 2 minutes, needs only gcc/sha256sum/pytest). MIT.

Status, stated precisely

• Public research engine, not a service. Its reconstructions are semantic contracts verified under specific corpora and locales — not drop-in replacements.
• Third-party verification: NOT_EXECUTED. All evidence is from one host, one UID, the same human author (with AI-assisted tooling). The blind-verification issue that decides the question lives in the sibling package rxre-cleanroom#1.
• No general claims. No PASS, no global correctness, no independent "RXRE is right" for any other binary, no coverage beyond the sealed artifacts.
• GPU access is restricted to shader compilation/execution via Vulkan/SPIR-V. No MMIO, no /dev/mem, no root, no --auto, no global rollback.
• PUERTA acceptance (its gate, condition 3): PENDING_EXTERNAL_REVIEWER by design.

What is actually verified

Only six targets exist, with sealed evidence inside evidence/r3/:

Target	Artifact	Corpus	Result
filenvercmp (sort = version) in /usr/bin/ls — sha256 54e936d3...	semantic C reconstruction	34,145 x 2 sealed pairs (seeds 20261112/13, 0 divergences); adversarial challenge 3,100+141 / 0 divergences; locale-invariance byte-equal C ≡ es_AR.utf8; i18n black-box 200 names identity vs native ls -v	ok byte-exact vs native
wc in /usr/bin/wc — sha256 81d61459...	semantic C reconstruction	C locale + es_AR.utf8; default / flags / -m multibyte / -L width / --files0-from / flags x multibyte / files0-from adversarial	ok 0 divergences vs native GNU
fold in /usr/bin/fold — sha256 95030461...	property reconstruction (line-wrap width)	2 sealed generations (2.554 each, -w 4, seeds 20261113/14) + gated w=1/2/3/8 corpora (1.054 each; 6 gated regenerations); valid UTF-8 only; byte-exact vs native /usr/bin/fold -w N	ok 0 divergences vs native GNU
tr in /usr/bin/tr — sha256 b1d1bfbe...	property reconstruction (squeeze of byte-x runs, tr -s x)	2 sealed generations (2.513 each, seeds 20261113/14); byte-exact vs native /usr/bin/tr -s x with harness over stdin	ok 0 divergences vs native GNU
dirname in /usr/bin/dirname — sha256 aac6a889...	property reconstruction (parent directory, bounded)	2 sealed generations (1.917/1.901, seeds 20261113/14); byte-exact vs native /usr/bin/dirname with harness over stdin	ok 0 divergences vs native GNU
basename in /usr/bin/basename — sha256 1ef99139...	property reconstruction (last component, bounded)	2 sealed generations (1.987/2.013, seeds 20261114/15); byte-exact vs native /usr/bin/basename with harness over stdin	ok 0 divergences vs native GNU

Details, logs and matrices live in evidence/r3/ (EXPEDIENTE_R3.md is the full case file — Spanish).

New here? Read QUICKSTART.md first (15 minutes, verification-first).

Verify it yourself — one command

sh evidence/r3/verify.sh

One command; it checks the SHA-256 of 65 sealed artifacts against their manifest, verifies the identity of the native reference binaries (/usr/bin/ls, /usr/bin/wc, /usr/bin/fold, /usr/bin/tr, /usr/bin/dirname, /usr/bin/basename; fail-closed), rebuilds the reconstruction, and re-runs the differential comparison on both generations of filenvercmp (34,145 × 2) plus the challenge corpus (3,100 + 141 deliberately identical a==b = 3,241 cases), the locale-invariance test, the i18n black-box, the fold property reconstruction (2 sealed generations, byte-exact), and all wc corpora (LC_ALL=C and es_AR.utf8; bytes/chars/-L/flags), the dirname and basename property reproductions (2 sealed generations each, byte-exact), then runs the engine's Python test suites. Expected tail: total=34145 mismatches=0 twice, total=3100 mismatches=0, total=3241 mismatches=0, locale invariance OK, i18n black-box OK, two total=2554 mismatches=0 lines for fold, twelve total=… mismatches=0 lines for wc, two total=1917/1901 lines for dirname, two total=1987/2013 lines for basename, 131 passed, 2 skipped (the 2 skips are test_replay_mision cases whose precondition is the author's private operations DB, deliberately not distributed; they run wherever that DB exists), then VERIFY: all steps passed.

For independent reviewers

reviewer-kit/ contains the step-by-step verification guide, the counterexample report template, and the scope/limits of this package. If you reproduce or refute the sealed chain, file it as an issue with your environment and exact commit.

Repository layout

tools/rxre/        engine source (auditor architecture, contracts, verifier, API)
tools/rxre/tests/  adversarial test suites (integrity, contracts, reproducibility)
evidence/r3/       sealed artifact set (corpora, matrices, reconstruction, seals, journals)
docs/              method identity documents (Spanish)
audit/evidence/    the two evidence records referenced by the integrity suite
GLOSSARY.md        terms in EN+ES (PUERTA, seal-before-reveal, semantic contract, R3…)

Not here on purpose

• No SQLite databases, no third-party source or binaries, no private captures.
• No marketing claims this repo can't show you the bytes of. The method is interesting precisely because it is falsifiable.

Corpus sensitivity (quantitative, not self-congratulation)

evidence/r3/records/coverage_report.txt (generated by coverage_audit.py) shows what the original seed 20261112/20261113 generations exercise — it was generated for those two sealed corpora, not for the challenge corpus. The challenge corpus deliberately targets the original weak spots instead: 1,875/3,100 cases with names ≥32 B, 95.1% with UTF-8 bytes, 225 distinct return values, plus 141 deliberately identical a==b pairs (= 3,241 total in hito 13b). Compare like with like:: UTF-8 content 3.89% of cases, tilde 13.45%, dot-leading 5.47%, prefix-related 5.09%, numeric-only 2.30%, long names (>32 B) 0%. Outcome balance: 49.6% negative / 0.66% equal / 49.8% positive, with ~640 distinct unclamped return values (not just -1/0/1). Gap honesty: UTF-8 filename coverage is thin in byte-level terms; a black-box es_AR.utf8 i18n seal (200 names, ls -Av vs reconstruction) lives in evidence/r3/filenvercmp_i18n_20261126/ — still not broad generalization. If your adversary generator beats this corpus, file it in the sibling issue — that's a positive result.

An optional Dockerfile exists for reviewers who want a pinned toolbox (gcc + python + pytest). Built and run with rootless podman 5.8.7 on this host: image boots verify.sh, and its suites pass (one host-artifact-bound test skips honestly because the sealed reference artifact lives on the author's private host; everything else passes). The image is pinned to fedora:44 precisely because its /usr/bin/ls matches the sealed reference hash of this package.

Documentation map

• English entry points: this README, GLOSSARY.md
• Spanish case files: docs/RXRE_ESENCIA.md, docs/RXRE_LIMITACIONES.md, docs/RXRE_PROCEDENCIA.md, evidence/r3/EXPEDIENTE_R3.md
• Sibling public package: rxre-cleanroom v1.9.0 — same brother-in-arms, aligned release; commended to the same blind-verification call.
• Own auditable snapshot: v1.9.0 release (rxre-engine-1.9.0.zip; SIX bounded targets sealed: filenvercmp, wc, fold, tr, dirname, and basename; 80 passed, 2 skipped at that tag; 16 stages; 65-artefact manifest). main keeps moving after that tag (eight commits since the tag; gate 131 passed, 2 skipped; see UPDATE-33 in evidence/r3/EXPEDIENTE_R3.md); a new tag/release that reseals that state is pending owner authorization and has not been published. Superseded snapshots (v1.8.0 and earlier) kept byte-exact: v1.5.3 (transparency/content fix), v1.5.1, v1.5.0 (FIVE targets + transparency), v1.4.0 (FOUR targets), v1.3.10, v1.3.9, v1.3.8.

License & authorship

MIT (see LICENSE). Original work by Lucas Cañete. GNU coreutils binaries, if used as runtime oracles, belong to their rights-holders; nothing here redistributes them. The corpus and matrices were synthesized by the author.

Co-developed with AI tooling under human ownership. Human ownership of claims and auditability rests with Lucas Cañete; AI assistance does not constitute an independent verification of any fact, and this repo does not treat it as one.