# Perceptual Twins v2.2.0 — correction and validation summary

**Date:** 12 August 2026  
**Author:** Jandislei Antonio Genova  
**Status:** independent protocol preprint; not peer reviewed

## Outcome

Version 2.2.0 is ready for use as a protocol/method manuscript with internal
software audits. It is not an empirical validation of epistemic autonomy and
cannot confirm or exclude a “great leap” of AI.

## Manuscript corrections

- reframed the title and contribution from a tested benchmark to a proposed protocol;
- defined the causal unit as a pre-assignment exogenous anchor package;
- separated anchor-level recovery from the population macro-F1 functional;
- replaced the unconditional `tau_couple` interpretation with conditional
  `tau_decouple`, while retaining the historical software key;
- removed the “two of six” shortcut and made claims endpoint-specific under Holm control;
- replaced arbitrary sample floors with simulation of the full decision rule;
- froze held-out evaluation before final model selection;
- recomposed the principal equations as native Word mathematics;
- reported corrected Replication Kit v1.0.2 values and Stage 0.5 v0.2.0 results;
- converted 47 references to ABNT NBR 6023:2025 style and aligned citations with
  NBR 10520:2023;
- kept the 189-word abstract and 220-word resumo within the 100–250-word range
  used for article abstracts under NBR 6028:2021;
- synchronized PDF metadata and removed the duplicate keyword token.

## Replication Kit v1.0.2

- 8/8 automated tests passed;
- restricted randomization fails closed;
- named PCG64 streams remove evaluation-order dependence from stochastic ties;
- the bootstrap quantile method is explicitly `linear`;
- regenerated JSON matches to absolute tolerance `1e-12` and CSV matches byte for byte;
- all six encoded efficacy references are covered and both E/C_E diagnostics equal zero;
- the 52-file SHA-256 manifest is complete.

## Stage 0.5 Validation v0.2.0

- 16/16 automated tests passed;
- baseline SHA-256 is verified before execution;
- sample and bootstrap streams are independent;
- every repetition, seed, estimate, and interval is exported;
- decisions use simultaneous exact-binomial `PASS`, `FAIL`, `INCONCLUSIVE`, and
  `NOT_APPLICABLE` states;
- the broad diagnostic has 18 informative `INCONCLUSIVE` cells and six
  degenerate `NOT_APPLICABLE` checks;
- the critical nondegenerate-null gated-revision `tau_tag` cell passed locally:
  coverage 0.954 [0.9391, 0.9661] and FPR 0.046 [0.0339, 0.0609];
- 3,400 released repetition records reconcile with the summaries;
- the 35-file SHA-256 manifest is complete.

## Final hostile review

The second adversarial review assigns **82.4/100** and recommends **major
revision, publishable only as a protocol/method paper**. It rejects use as an
empirical benchmark paper or proof of autonomy. The two immediate blockers are:

1. deposit version-specific, citable releases for Replication Kit v1.0.2 and
   Stage 0.5 v0.2.0;
2. add the author's real ORCID and correspondence email if required by the
   selected venue.

## Legitimate next use

The software is suitable for estimator regression, interval auditing, and
Stage 1 engineering. Stage 1 must add a correct rule absent from the initial
hypothesis space, structurally distinct fault generators, multiple agent
architectures, frozen held-out evaluation, and independent replication.