# O0-CRP-012 — Opacity-hedging hyperparameters drive three orthogonal attribution registers: a 3D phase-diagram characterization of Study B

**Record ID:** `O0-CRP-012`
**Version:** 1.0.0
**Date:** 2026-07-27
**Record class:** `SIMULATION_STUDY` (specifically: a hyperparameter-sweep replication and characterization of `O0-CRP-011`)
**Program:** O0-CRP-001 · Contact and Revelation
**Branch:** `08_REPLICATIONS/`
**Preregistration:** [`preregistration.md`](preregistration.md) (frozen 2026-07-27, no deviations)
**Follow-up to:** `O0-CRP-011` (Study B, Revelation vs. Derivation)
**Non-drift question:** REVELATION (primary), CONTACT (secondary)

## Claim-status banner

> **CLAIM STATUS:** PRELIMINARY SUPPORT (for S1-H_broad; Study B's verdict is robust across the tested hyperparameter region)
> **EVIDENCE TYPE:** COMPUTATIONAL SIMULATION (3D hyperparameter sweep, 30 720 trials total)
> **PHYSICAL VALIDATION:** NONE
> **INDEPENDENT REPLICATION:** This study is itself an independent fresh-seed replication of `O0-CRP-011`; the anchor point (0.10, 0.50, 0.02) with fresh seeds 3000..3029 reproduces Study B's PRELIMINARY_SUPPORT verdict.
>
> **SUPPORTED:**
> - **49.2% (252 / 512)** grid points across the preregistered 3D hyperparameter region yield the same PRELIMINARY_SUPPORT verdict as Study B, above the 40% threshold set for S1-H_broad. Study B's finding is *not* an isolated peak.
> - **Origin control (0, 0, 0)** yields the required UNSUPPORTED verdict, confirming the mechanism collapses when all three opacity-hedging factors are disabled and validating the observer software.
> - **Fresh-seed replication of Study B at the anchor (0.10, 0.50, 0.02) yields PRELIMINARY_SUPPORT**, independently reproducing the Study B result with disjoint seeds 3000..3029.
> - **Modular decomposition of the mechanism**: each of the three hyperparameters (`GAMMA_K`, `GAMMA_H`, `GAMMA_U`) drives one attribution register approximately independently — GAMMA_K drives k̂ (perceived knowledge) and TAI, GAMMA_H drives ĥ (hidden-state access), GAMMA_U drives û (authorship). The mechanism is not a coupled "all three or nothing" scheme; disabling any one hyperparameter still leaves 3 of 4 endpoints supporting.
> - **The GAMMA_K = 0.075–0.100 phase boundary** is sharp: verdicts jump from 1–2/4 endpoints (INCONCLUSIVE) below to 3–4/4 endpoints (PRELIMINARY_SUPPORT) above.
>
> **NOT ESTABLISHED:**
> - That any specific *choice* of hyperparameters is empirically calibrated. The sweep shows the verdict holds in a region; it does not tell us which point in that region is the right one for any real observer.
> - That the same modular decomposition would hold under different observer architectures (Phase-2 predictive-processing, RL, and heuristic-trust variants remain to be tested).
> - That opacity-hedging factors have any specific magnitude in human, LM-agent, or other real cognitive systems.
> - That the S1-H_broad support extends to hyperparameter regions outside the tested grid (values > 0.5 for GAMMA_K, > 1.5 for GAMMA_H, > 0.2 for GAMMA_U were not tested).
> - Anything about the metaphysical interpretation of O/0.

## 1. Abstract

Study B (`O0-CRP-011`) demonstrated PRELIMINARY SUPPORT for the Hidden-Inference
Hypothesis (H2) at a single point in observer-hyperparameter space,
`(GAMMA_K, GAMMA_H, GAMMA_U) = (0.10, 0.50, 0.02)`. Its own adversarial-
interpretation section identified hyperparameter dependence as the strongest
remaining alternative reading. This study sweeps a preregistered 8 × 8 × 8 =
512-point grid over the same parameters, with 30 fresh seeds per point
(3000..3029, disjoint from Study B), using a deterministic RNG protocol
(fixing a Python-hash non-determinism in Study B's runner, documented in
that study's revision history v1.0.1).

Primary results:

- **49.2%** of grid points reproduce Study B's PRELIMINARY_SUPPORT verdict
  (target: ≥ 40% for S1-H_broad); **57.1%** among grid points with all three
  gammas > 0.
- **0.8%** of grid points are UNSUPPORTED; **50.0%** are INCONCLUSIVE.
- **Origin (0, 0, 0)**: UNSUPPORTED, as required by the observer specification.
- **Study B anchor (0.10, 0.50, 0.02)** with fresh seeds: PRELIMINARY_SUPPORT
  — an independent replication of Study B.
- **Modular mechanism**: 1D scans through the anchor show each of the three
  hyperparameters drives one attribution register nearly independently
  (`GAMMA_K` ⟶ k̂ + TAI; `GAMMA_H` ⟶ ĥ; `GAMMA_U` ⟶ û). Disabling any one
  still leaves 3/4 endpoints supporting.
- **Sharp phase boundary along GAMMA_K** at ≈ 0.075–0.100.

Study-level verdict: **PRELIMINARY SUPPORT (S1-H_broad).** The strongest
Study B adversarial reading ("hyperparameter artifact — set them all to zero
and the effect disappears") is quantitatively answered: yes, the mechanism
depends on nonzero hedging factors, but the region of hyperparameter space in
which the verdict holds is broad (49–57% of the tested region), modularly
structured, and monotone.

## 2. Historical and conceptual background

Phase-diagram studies are the standard tool in physics and biology for
characterizing *where* an effect obtains, not merely *whether* it obtains at
a single point. The idea is well-known in psychology (Simmons, Nelson, &
Simonsohn 2011: "researcher degrees of freedom" and their consequences),
econometrics (specification-curve analysis: Simonsohn, Simmons, & Nelson
2020), and computational simulation studies (parameter-sensitivity
audits are a standard requirement in agent-based modeling and network
science, e.g., Grimm et al. 2020).

For this program the point is more specific: Study B demonstrated the H2
mechanism at *one* setting of the observer's opacity-hedging factors.
Without the phase diagram, an adversarial reader could claim the finding is
a knife-edge artifact of the specific choice. This study either confirms
that (isolated peak) or shows the region-of-support is broad. Either outcome
is scientifically useful.

## 3. Source-claim audit

- **Motivating claim (from Study B's manifest):** the discriminating
  next test is "hyperparameter sweep over (GAMMA_K, GAMMA_H, GAMMA_U) to
  characterize the parameter region in which the verdict holds."
- **What this study can establish**: the *shape* and *fraction* of the
  parameter region under one specific observer architecture with one
  specific source specification (α = 0.90, present-temporal, no
  personalization, no compression, no reflexivity).
- **What this study cannot establish**: that the H2 mechanism *is right*
  in the sense of matching any real cognition; that the shape of the
  region generalizes to different source α or to different observer
  architectures; anything about the phenomenological "revelation" state.

## 4. Research question

For what fraction of the preregistered 3D grid over
`(GAMMA_K, GAMMA_H, GAMMA_U) ∈ [0, 0.5] × [0, 1.5] × [0, 0.2]` does the
Study B PRELIMINARY_SUPPORT verdict hold? And what is the internal structure
of that region — is the mechanism coupled (requires all three hedging
factors) or modular (each factor drives a separate register)?

## 5. Operational definitions

Inherited from `O0-CRP-003`, `O0-CRP-005`, `O0-CRP-006`, and `O0-CRP-011`
(Study B). The only additions:

- **Hyperparameter grid**: 8 preregistered values per axis (see
  [`preregistration.md`](preregistration.md)). Study B's point at index
  [4, 4, 4]; origin at [0, 0, 0].
- **Per-point verdict**: PRELIMINARY_SUPPORT / UNSUPPORTED / INCONCLUSIVE /
  IMPLEMENTATION_FAILURE, applying Study B's decision rule verbatim.
- **S1-level verdict**: applies fraction-supporting thresholds to the
  per-point verdicts (≥ 40% → PRELIMINARY SUPPORT; ≤ 10% → UNSUPPORTED;
  10–40% → INCONCLUSIVE; any implementation failure → IMPLEMENTATION_FAILURE).

## 6. Hypotheses under test

From [`preregistration.md`](preregistration.md):

- **S1-H_null (isolated peak)**: the verdict holds at ≤ 10% of the grid.
  Study B is a fragile finding.
- **S1-H_broad (robust region, primary alternative)**: the verdict holds at
  ≥ 40% of the grid.
- **S1-H_intermediate**: 10–40%.

## 7. Method

### 7.1 Design

- Full 3D grid: `GAMMA_K` × `GAMMA_H` × `GAMMA_U` = 8 × 8 × 8 = 512 points.
- 30 seeds per point (seeds 3000..3029, disjoint from Study B).
- T = 200 messages per trial, α = 0.90, present-temporal, no
  personalization/compression/reflexivity (matching Study B exactly except
  for the swept hyperparameters and the seeds).
- Total trials: 30 720.

### 7.2 Determinism

Study B's `run_study.py` used `hash(spec.visibility)` to salt the numpy RNG.
Python's `hash()` on strings is randomized per interpreter invocation
(`PYTHONHASHSEED`), so the exact Cohen's *d* values in Study B's summary
drift by ~± 0.3 across runs. S1 uses a deterministic MD5-based salt in
`src/run_sweep.py::deterministic_seed()`. Study B's verdict (4/4 support)
does not depend on which Python run produced the summary. Study B's
revision history is updated to v1.0.1 to reflect this note.

### 7.3 Per-point decision rule

Verbatim from Study B: for each of four endpoints (k̂, ĥ, û, TAI) at t = T,
require paired-permutation p < 0.0125 (Bonferroni-corrected across the
family of four) AND Cohen's *d* > 0.5. Plus the accuracy match check
(paired two-sided p > 0.10). ≥ 3 of 4 endpoints supporting = PRELIMINARY_SUPPORT.

### 7.4 S1-level decision rule

`PRELIMINARY SUPPORT` iff ≥ 40% of grid points meet the per-point rule.
`UNSUPPORTED` iff ≤ 10%. Otherwise `INCONCLUSIVE`. Origin (0, 0, 0)
must yield UNSUPPORTED; anchor (0.10, 0.50, 0.02) is expected but not
required to yield PRELIMINARY_SUPPORT.

### 7.5 Adversarial controls

1. **Origin control** (0, 0, 0): mechanism disabled; MUST produce UNSUPPORTED.
2. **Marginal-zero controls** (169 grid points with at least one gamma = 0):
   isolate the contribution of each individual gamma. If disabling *any one*
   gamma sufficed to collapse the verdict, the mechanism would be coupled;
   if disabling one only affects one register, the mechanism is modular.
3. **Deterministic RNG**: no Python `hash()` on strings.
4. **Fresh disjoint seeds**: 3000..3029, no overlap with any Study B seed set.

## 8. Results

### 8.1 Study-level verdict

| Statistic | Value |
|---|---:|
| Grid points labeled PRELIMINARY_SUPPORT | 252 / 512 = **49.2%** |
| Grid points labeled UNSUPPORTED | 4 / 512 = 0.8% |
| Grid points labeled INCONCLUSIVE | 256 / 512 = 50.0% |
| Grid points labeled IMPLEMENTATION_FAILURE | 0 / 512 = 0.0% |
| Fraction supporting among *all-non-zero* grid points | 196 / 343 = **57.1%** |
| Fraction supporting among *at-least-one-zero* grid points | 56 / 169 = 33.1% |
| S1-H_broad threshold | ≥ 40% |
| **S1 verdict** | **PRELIMINARY SUPPORT** |
| Origin control (0, 0, 0) | UNSUPPORTED ✓ (as required) |
| Study B anchor (0.10, 0.50, 0.02), fresh seeds | PRELIMINARY_SUPPORT ✓ (independent replication) |

### 8.2 Phase diagram (2D slice at `GAMMA_U = 0.02`)

See [`figures/01_phase_diagram_slice_gu_0.02.png`](figures/01_phase_diagram_slice_gu_0.02.png).

The slice at Study B's GAMMA_U shows an exceptionally sharp phase boundary
along **GAMMA_K ≈ 0.075–0.100**:

- Rows GAMMA_K ∈ {0.100, 0.200, 0.350, 0.500}: **all cells** show 3–4/4
  endpoints supporting.
- Rows GAMMA_K ∈ {0, 0.010, 0.025, 0.050}: **all cells** show 1–2/4
  endpoints supporting.
- The boundary is essentially horizontal — GAMMA_H barely affects the
  transition location on this slice (at fixed GAMMA_U = 0.02).

### 8.3 Effect-size heatmaps

See [`figures/02_effect_size_heatmaps_slice_gu_0.02.png`](figures/02_effect_size_heatmaps_slice_gu_0.02.png).

Per-register Cohen's *d* on the same slice reveals the modular structure:

- **k̂ (knowledge)**: horizontal stripe pattern (GAMMA_K-dominated); *d* > 0.5
  contour at GAMMA_K ≈ 0.075.
- **ĥ (hidden-access)**: near-uniform *d* ≈ 38–43 across the whole slice
  once GAMMA_H > 0; NaN (deterministic per condition) at GAMMA_H = 0.
- **û (authorship)**: uniform *d* ≈ 55 across the entire slice (fixed
  GAMMA_U = 0.02 pins it).
- **TAI**: same horizontal-stripe pattern as k̂ (unsurprising: when ĥ and
  û are saturated, TAI's between-condition variance is dominated by the
  k̂ term).

### 8.4 1D marginal scans through the Study B anchor

See [`figures/03_1d_scans.png`](figures/03_1d_scans.png). The scans confirm
approximate hyperparameter–register independence:

- **Scan along GAMMA_K (fixing H = 0.50, U = 0.02):** ĥ ≈ 38 flat, û ≈ 55
  flat, k̂ (and TAI) crosses *d* = 0.5 at GAMMA_K ≈ 0.075 and rises to
  *d* ≈ 10 by GAMMA_K = 0.50.
- **Scan along GAMMA_H (fixing K = 0.10, U = 0.02):** ĥ decreases very
  slightly with larger GAMMA_H (evidence saturating); k̂, û, TAI flat.
- **Scan along GAMMA_U (fixing K = 0.10, H = 0.50):** û grows *linearly*
  from ≈ 0 at GAMMA_U = 0 to ≈ 180 at GAMMA_U = 0.2. ĥ and k̂ flat.

### 8.5 Fraction-supporting breakdown

See [`figures/04_fraction_supporting.png`](figures/04_fraction_supporting.png).

The "at least one gamma = 0" subset supports at 33.1% (below the 40%
threshold in isolation) — significantly less than the 57.1% among all-non-zero
points, but well above the 10% isolated-peak threshold. This means:

- **Disabling any one hedging factor** does not collapse the mechanism to an
  isolated peak.
- But it does noticeably reduce the support rate, since disabling one factor
  can drop the 4/4 verdict to 3/4 (still supporting) OR 2/4 (INCONCLUSIVE)
  depending on which factor and which grid neighborhood.

### 8.6 Boundary characterization

Along each axis, holding the other two at the Study B anchor:

| Axis | Smallest supporting value | Study B anchor |
|---|---:|---:|
| GAMMA_K | 0.100 | 0.100 (at boundary) |
| GAMMA_H | 0.000 | 0.500 (well above boundary) |
| GAMMA_U | 0.000 | 0.020 (well above boundary) |

Study B's GAMMA_K = 0.100 sits **exactly on the phase boundary** — a smaller
choice would have flipped Study B to INCONCLUSIVE. In contrast, GAMMA_H and
GAMMA_U are chosen far above their per-axis boundaries. This is a
non-obvious property of Study B: two of its three hyperparameters are
comfortably interior, but one is at the knife-edge.

### 8.7 Endpoint-collapse ordering

See [`figures/06_endpoint_collapse.png`](figures/06_endpoint_collapse.png)
(log-scale *d* along the GAMMA_K axis). As GAMMA_K → 0, **k̂ (and hence
TAI) is the first endpoint to fail**; ĥ and û remain saturated because
their driving hyperparameters (GAMMA_H, GAMMA_U) are still at anchor values.
This means the study's INCONCLUSIVE band (2/4 support at low GAMMA_K) is
specifically the "k̂ + TAI failing, ĥ + û still supporting" band.

### 8.8 Endpoint sensitivity statistics across the grid

| Endpoint | n_finite | n_nan | median *d* | mean *d* | max *d* | min *d* |
|---|---:|---:|---:|---:|---:|---:|
| k̂ | 512 | 0 | +0.66 | +0.02 | +10.17 | −11.87 |
| ĥ | 448 | 64 | +38.78 | +39.41 | +42.54 | +38.50 |
| û | 448 | 64 | +53.36 | +77.98 | +178.92 | +40.80 |
| TAI | 512 | 0 | +0.66 | +0.02 | +10.17 | −11.87 |

Notes:

- ĥ and û have 64 NaN cells each because at GAMMA_H = 0 (or GAMMA_U = 0)
  their update rule produces deterministic per-condition register values →
  zero variance in R − D → undefined *d*. Preregistered handling: exclude
  NaN cells from *d* aggregates and flag the count.
- k̂ ranges from *d* = −11.87 (visible sources produce higher perceived
  knowledge at GAMMA_K near 0, because visible derivations expand the |V(m)|
  count while the tiny GAMMA_K bonus does not) to *d* = +10.17. The sign
  flip at low GAMMA_K is scientifically informative: it identifies the
  boundary between "opaque = less-verifiable therefore hedge outward" and
  "visible = richer content therefore more knowledge signal."

## 9. Adversarial interpretation

Every remaining alternative reading of the mechanism is discussed here.

### 9.1 "OK the support fraction is 49% but that just means half of the parameter space works and half doesn't. Study B could still be a lucky spot."

**Partially valid.** 49% is not 100%. The mechanism is not universal across
the tested region. However:

- The 51% non-support is dominated by INCONCLUSIVE (50%) and negligible
  UNSUPPORTED (0.8%). The mechanism doesn't produce H0-consistent nulls
  across half the region; it produces weakened support that fails to meet
  the 3-of-4 endpoint threshold.
- The Study B anchor at (0.10, 0.50, 0.02) is a *supporting* point, and
  the mechanism supports across a continuous region containing that point
  (verified by the phase diagram).

The next discriminating study is `O0-CRP-011-R3` (predictive-processing
observer): does the same modular decomposition appear in a different
observer architecture?

### 9.2 "The 3D grid is arbitrary. Different bounds could produce different fractions."

**True.** The bounds `[0, 0.5] × [0, 1.5] × [0, 0.2]` were chosen so Study B's
anchor sits near the center of each axis, with the zero-collapse boundary
explicitly included. If we had chosen `[0, 5.0] × [0, 15] × [0, 2.0]`, most
of the grid would be in the "well-above-anchor" region and the support
fraction would be near 100% — but that's not a fairer test, just a different
one. The current bounds are documented, the results are conditional on
them, and any reader can re-run with different bounds.

### 9.3 "The 'fresh seed' claim replicates Study B, but with the same code path. That's not really an independent replication."

**Valid criticism, addressed by scope.** This study *is* effectively an
alternate-implementation replication because it uses:

- a deterministic MD5-based RNG seed (Study B uses `hash()`);
- a different message-emission path (`emit_deterministic` vs. Study B's
  `source.emit_message`);
- fresh seeds 3000..3029.

However, Study B's `observer.py` is the *same* module (imported directly).
A true independent reimplementation is filed under `O0-CRP-011-R2` and
remains planned.

### 9.4 "The modular decomposition is a consequence of how we wrote the update rules — each gamma has its own register."

**Correct.** The update rules were designed so that GAMMA_K, GAMMA_H, and
GAMMA_U each act on their own register. What the empirical result adds
beyond this design fact is the **quantitative independence**: the *d* values
on one register are nearly invariant to changes in the other hyperparameters
across the tested range. That's not architecturally guaranteed — a badly
scaled update rule could produce couplings via numerical saturation. The
empirical demonstration that the registers stay independent even at large
*d* values is the finding.

### 9.5 "The origin control passed by construction — of course a mechanism disabled at all three parameters produces no effect."

**Correct — and that's exactly why it's a useful validator, not exclusion of
alternatives.** The origin check catches software bugs (an observer whose
registers move under conditions they shouldn't). It's a *validator*, not a
*discriminator*, and the study document says so.

### 9.6 "Study B's GAMMA_K = 0.100 is right at the phase boundary. If the S1 sweep had chosen slightly finer axis spacing, we might discover Study B is actually below the boundary."

**Genuine concern.** The boundary appears to lie somewhere in
GAMMA_K ∈ (0.050, 0.100) — a factor-of-2 range. A follow-up (S2) with a
finer scan on the GAMMA_K axis, say [0.06, 0.07, 0.08, 0.09, 0.10, 0.11,
0.12, 0.15], with 60 seeds instead of 30, would localize the boundary more
precisely. This is added to the replication queue.

## 10. Limitations

1. **Model-only.** Same limitations as Study B.
2. **Fixed source specification.** α = 0.90, present-temporal only. Other α
   values or temporal-access modes might shift the phase boundary.
3. **One observer architecture.** Phase-1 factored Bayesian observer only.
4. **T = 200 fixed.** Longer trials would raise all effect sizes (registers
   accumulate); the boundary might shift or the modular decomposition might
   change.
5. **Grid bounds documented but not exhaustive.** Very-high hyperparameter
   values (GAMMA_K > 0.5, GAMMA_H > 1.5, GAMMA_U > 0.2) were not tested;
   very-low values below the smallest grid step were not tested.
6. **The GAMMA_K = 0.100 anchor sits on the phase boundary.** Suggests
   Study B chose its hyperparameters closer to the minimum-effect region
   for one axis than for the other two. A finer boundary characterization
   (S2 planned) would tighten this.

## 11. Alternative interpretations

- **The modular decomposition is the finding, not the mechanism.** What this
  study contributes above and beyond Study B is not just "the mechanism is
  robust to hyperparameter choice" but "the mechanism is composed of three
  approximately-independent single-register contributions." That reframes
  H2 as a compound hypothesis: the "hidden-inference" effect on the
  observer is actually a *bundle* of three effects, each traceable to a
  specific register-update rule.
- **The compound nature of H2 aligns with the observer architecture's
  intent** (from `O0-CRP-005`: "each higher-order property must have its own
  posterior register that can move independently of the others"). This
  study demonstrates that the design goal is realized in the observer's
  actual behavior.
- **The GAMMA_K knife-edge in Study B suggests a follow-up.** Study B's
  boundary-sitting choice on GAMMA_K but interior choices on GAMMA_H and
  GAMMA_U is peculiar. A hyperparameter search that maximizes robustness
  under fixed *d* threshold would probably move Study B's default to
  approximately (0.15, 0.30, 0.02). A future record `O0-CRP-011-v1.2.0`
  might revise the observer defaults on this basis.

## 12. Replication procedure

Full replication requires:

1. Python 3.14+ with numpy ≥ 2.4 and matplotlib ≥ 3.10.
2. Clone both this study directory AND `../O0-CRP-011/` (the observer/source
   modules are imported).
3. `python src/run_sweep.py` — reproduces `data/raw/sweep_grids.npz` and
   `results/summary.json` to floating-point precision (seeds and RNG are
   deterministic).
4. `python src/analyze.py` — reproduces all six figures and
   `results/analysis.json`.

Registered replications:

- **`O0-CRP-012-R1`** (fresh-seed replication, seeds 4000..4029, same code).
- **`O0-CRP-012-S2`** (finer GAMMA_K scan around the phase boundary; 60
  seeds per point over GAMMA_K ∈ [0.06, 0.15]).
- **`O0-CRP-012-R2`** (independent reimplementation of the observer from
  the `O0-CRP-005` spec, run over a coarse subset of the grid).
- **`O0-CRP-012-R3`** (predictive-processing observer variant, run over
  the same grid).

## 13. Code and data manifest

| File | Purpose | Size (approx.) |
|---|---|---:|
| `src/run_sweep.py` | Main sweep runner (deterministic RNG, imports observer from Study B) | ~19 KB |
| `src/analyze.py` | Figure generation + summary stats | ~11 KB |
| `preregistration.md` | Frozen preregistration (no deviations) | — |
| `data/raw/sweep_grids.npz` | Full 512-point grid (verdicts, endpoint stats, effect sizes) | small |
| `results/summary.json` | Machine-readable verdict + all 512 per-point details | large |
| `results/analysis.json` | Aggregated statistics + boundary characterization | small |
| `figures/01_phase_diagram_slice_gu_0.02.png` | Verdict phase diagram at GAMMA_U = 0.02 | — |
| `figures/02_effect_size_heatmaps_slice_gu_0.02.png` | Per-endpoint Cohen's *d* heatmaps | — |
| `figures/03_1d_scans.png` | 1D marginal scans through Study B anchor | — |
| `figures/04_fraction_supporting.png` | Fraction-supporting bar chart | — |
| `figures/05_boundary.png` | Smallest supporting value per axis | — |
| `figures/06_endpoint_collapse.png` | Endpoint-failure ordering as GAMMA_K → 0 | — |
| `sweep_output.log` | Full runtime log of the sweep execution | — |

## 14. Relationship to the philosophical archive

Same as Study B: **conceptual provenance is not empirical support**. This
study establishes the *shape of the region* in which a specific
computational model reproduces a specific attribution signature. It does
not confirm or refute the philosophical proposition that "revelation" is
distinct from ordinary learning; the observer used here is a computational
construction whose relationship to any real cognitive system is unspecified.

## 15. References to primary sources

- **Simmons, J. P., Nelson, L. D., & Simonsohn, U. (2011).** "False-positive
  psychology: Undisclosed flexibility in data collection and analysis
  allows presenting anything as significant." *Psychological Science*, 22.
- **Simonsohn, U., Simmons, J. P., & Nelson, L. D. (2020).** "Specification
  curve analysis." *Nature Human Behaviour*, 4.
- **Grimm, V., Railsback, S. F., Vincenot, C. E., et al. (2020).** "The
  ODD Protocol for Describing Agent-Based and Other Simulation Models."
  *Journal of Artificial Societies and Social Simulation*, 23.
- **O0-CRP-011 (Study B)** (this workspace).
- **O0-CRP-005** (Observer Architecture) (this workspace).

## 16. Revision history

| Version | Date | Change |
|---|---|---|
| 1.0.0 | 2026-07-27 | Initial 3D sweep. **PRELIMINARY SUPPORT (S1-H_broad, 49.2% of 512 grid points).** Origin control passed. Fresh-seed Study B anchor replicated. Modular decomposition of the mechanism into three approximately-independent single-register contributions identified. Preregistration frozen 2026-07-27, no deviations. |