SCIENTIFIC PAPER · O0-PAPER-007

The k̂ Flip Is Observation-Model-Specific, Not Intrinsic to Predictive Processing: A One-Line Ablation

STATUSPUBLISHED · single-study paper wrapping O0-CRP-014. Verdict: PRELIMINARY SUPPORT for CRP-013 §9.1 mechanism-localization prediction. Three-way head-to-head at Study B's seeds: k̂ d = +4.4 → −31.5 → +7.9 (Bayesian → PP → PP-A).
EVIDENCE TYPEFORMAL SCIENTIFIC PAPER · IMRAD-format wrapper around O0-CRP-014.
REPLICATIONDERIVED · study O0-CRP-014 is itself the ablation resolution of the CRP-013 k̂ flip.
PHYSICAL VALIDATIONNONE (computational simulation).
VERSION1.0.0
DATE

The k̂ Flip Is Observation-Model-Specific, Not Intrinsic to Predictive Processing: A One-Line Ablation

**Ozone Archive Research Program**

**Paper ID:** O0-PAPER-007

**Underlying study:** O0-CRP-014

**Date:** 2026-07-29

**Version:** 1.0.0

**Record class:** SCIENTIFIC PAPER · single-study IMRAD wrapper

---

Abstract

Preregistered mechanism-localization ablation. O0-CRP-013 reported that under a predictive-processing (PP) observer the k̂ (knowledge) contrast flips direction relative to the Bayesian baseline. That study's Section 9.1 predicted the flip was caused by the direct y_k = log|V(m)| observation channel, not by anything intrinsic to PP. This paper tests that prediction with a one-line ablation: PP-A is identical to R3's PP observer except μ_k receives no direct observation update — it is updated *only* through prediction-error credit assignment under opacity. Three-way head-to-head at Study B's own seeds (1000–1029): k̂ Cohen's d goes **+4.4 (Bayesian) → −31.5 (PP with direct y_k) → +7.9 (PP-A without direct y_k)**. The flip is fully restored. The R3 mechanism-disagreement is cleanly localized to a single observation-model term.

**Verdict:** PRELIMINARY SUPPORT for CRP-013's Section 9.1 mechanism-localization prediction.

---

1. Introduction

O0-CRP-013 reported a partial architecture-dependence finding: two of three central belief contrasts (ĥ, û) replicate from Bayesian to PP; the third (k̂) inverts. CRP-013 §9.1 predicted this flip was caused by the direct y_k observation channel present in PP but not directly parallel to Bayesian's evidence intake. This paper tests that prediction.

2. Methods

**PP-A observer.** Identical to CRP-013's PP observer with one change:

  • μ_k **no longer receives direct observation update**.
  • Instead, μ_k is updated only through prediction-error credit assignment under opacity.
  • Everything else (priors, μ_a, μ_h, μ_u update rules, credit assignment weights) preserved verbatim.

**Design.**

  • Confirmatory: 30 fresh seeds × 3 prior regimes × 2 conditions (R, D).
  • Three-way head-to-head at Study B's seeds 1000..1029: Bayesian vs PP (CRP-013) vs PP-A (this paper).

**Preregistered decision rules.**

  • k̂ d > 0.5 with p < 0.0125 in each of ANCHOR, FLAT, SKEPTICAL regimes to support the mechanism localization.
  • ĥ, û directions must remain preserved under PP-A.

3. Results

**At Study B's own seeds (1000–1029), three-way comparison:**

| Register | Bayesian d | PP (CRP-013) d | PP-A (this paper) d |

|---|---:|---:|---:|

| ĥ | +53.6 | +22.3 | +21.2 |

| û | +73.9 | +4.0 | +4.1 |

| k̂ | +4.4 | **−31.5** | **+7.9** |

**k̂ direction fully restored under PP-A.** Removing the direct y_k observation flips the flip.

**All three prior regimes** show PP-A k̂ d > 0.5 with p < 0.0125.

**ĥ and û unchanged** — this ablation targets only k̂'s update path.

4. Discussion

CRP-013's Section 9.1 mechanism-localization prediction receives PRELIMINARY SUPPORT. The k̂ flip in CRP-013 is caused by the direct y_k observation, not by anything intrinsic to predictive processing.

**Broader implication.** A single term in an observation model can invert a belief-update direction. Cross-architecture replication in Bayesian belief models is not purely a function of the "family" (Bayesian vs PP) — it depends on which observation channels are present. This is a general methodological finding, not just about CONTACT_AND_REVELATION.

**Adversarial interpretation.** A skeptic could argue that PP-A is "just Bayesian in disguise" — that removing the direct y_k observation makes PP-A converge toward Bayesian behavior. This is partially true: PP-A is designed to test exactly this — whether removing a single specific term restores the direction. The ablation's clarity is a feature, not a defect.

5. Limitations

  • Only tests one specific ablation. Other observation-model choices are not explored.
  • Effect sizes under PP-A are still smaller than under Bayesian, so the two architectures are not equivalent.
  • Only three prior regimes tested; broader prior space not swept.

6. Related studies

  • **O0-CRP-013** — the study this paper resolves the k̂ flip for.
  • **O0-CRP-011, O0-CRP-012** — Study B baseline (Bayesian).
  • **O0-SCOPE-002** — claim registry entry C012.

7. Data and code availability

  • Full record: `research/studies/O0-CRP-014/scientific_record.md`
  • Source: `research/studies/O0-CRP-014/src/`
  • Figures: `research/studies/O0-CRP-014/figures/`

8. References

  • Friston, K. (2010). "The free-energy principle." *Nature Reviews Neuroscience.*
  • Rao & Ballard (1999). "Predictive coding in the visual cortex." *Nature Neuroscience.*

Source proposition

“IMRAD paper wrapper for CRP-014, generated as part of the paper-format publication layer.”

Conceptual provenance is not empirical support.

Related records

O0-CRP-014O0-CRP-013O0-CRP-011O0-CRP-012O0-SCOPE-002