# RN003 support1.1 — current checks and historical pilot

Supersedes public support1.0 on4 October2026. Current article edition1.4 is a selected literature synthesis and methods proposal. Canonical edited prose replaces earlier public reports and packets with successor notices; original identifiers/hashes remain in the review/publication records.

## Current deterministic checks

Run `node content/reproducibility/rn003/evaluate.mjs .` from the generated ARC repository. The script uses the local Worker and public files, without provider calls, remote requests or analytics writes. The result records eight fixed search probes and four current representation checks, hashes, UTF8 bytes, exact query behavior and evidence targets. `current-deterministic-results.json` is the separately dated executed result. Presence/absence checks are not semantic comprehension scores. Bytes are not tokenizer counts. Exact source hashes identify these checked local artifacts; later generated metadata or service shells may differ.

## Historical pilot

`historical-pilot-summary.json` describes the September29 four-reader Luna exercise and retains original input hashes. Those inputs were local pre-release artifacts, unequal in completeness. One separately initialized reader per condition and parent editorial scoring cannot identify causal format effects or independent evidence lineages. The original answers are not rewritten as new outcomes, and no new reader experiment is claimed. Earlier full packets and raw answer records are withdrawn with notices. Full historical public replay is consequently unavailable.

The new review found a missed inference: the TXT reader called RN002 adjacency interaction. The original rubric scored identity and causal-preference qualifications but did not separately score adjacency versus a reply. That limit qualifies its original perfect Stage2 score. Two unsupported attribution additions were already recorded. Local artifact counts omit chunk rereads/checksum operations and are not HTTP requests.

## Proposed benchmark

The hidden graph, extended precision/recall measures, disagreement handling, planning ranges, counterfactual controls and retrieval comparisons remain proposals to validate. Synthetic generation does not alone prove novelty; annotation labels are fallible; distinct model families and packets can share priors, sources or benchmarks. Source pointers preserve provenance, not complete claim verification. No writes or third-party probing are part of these checks.