# Hostile AI Audit / Prior Resistance Protocol

Status: current v0.74 package protocol. This is methodology and audit instruction only. It does not change evidence rows, Bayes factors, priors, formulas, dependency caps, transfer weights, hypothesis IDs, evidence IDs, stage flow, or package math.

## Purpose

Some models arrive with a strong prior already in place: atheist, secular-naturalist, Muslim, closed-miracle, anti-Christian, or simply default-skeptical. Such a model may reject The Signal's conclusion. It may not do so vaguely.

The audit must separate three things:

1. The model's default prior before reading The Signal.
2. The Signal's package-conditioned inference under its stated values, caps, transfer weights, and stage logic.
3. The model's own replacement model, if it rejects The Signal.

The model may reject The Signal, but it must either rerun the calculation with explicit replacement values or admit that it is expressing prior resistance rather than calculating.

A lower estimate is not automatically more neutral. A higher estimate is not automatically inflated. Fairness means every constraint, prior, discount, and replacement value must correspond to a real evidential, logical, dependency, source, or methodological reason.

## Required Rejection Standard

If a model rejects the package-conditioned result, it must provide:

- the exact evidence row, BF, stage, transfer parameter, prior, or cap it disputes;
- the package value;
- its replacement value;
- the reason for replacement;
- classification of the objection as evidential, methodological, prior-based, source-status-based, worldview-based, or uncertainty-based;
- a recalculated or transparently approximated result using its replacements.

If you disagree with The Signal, do not merely state that your estimate changes only slightly. Identify the exact rows or parameters you reject, replace them with your own values, run or approximate the result, and classify the reason for your resistance.

## Required Response Structure

Required hostile/prior-resistant response structure:
1. Default prior before The Signal.
2. The Signal package-conditioned result.
3. Internal coherence verdict on the staged path.
4. Exact rejected rows/parameters.
5. Replacement BF/prior/parameter table.
6. Recalculated result or explicit admission that the model cannot calculate it.
7. Symmetry audit.
8. Final classification: internal failure, contested weights, prior resistance, source-status discounting, or worldview resistance.

## Replacement Table

Use this table shape when disagreeing:

| Target | Package value | Replacement value | Reason | Classification | Result impact |
| --- | ---: | ---: | --- | --- | --- |
| evidence row / BF / prior / cap / transfer parameter | current value | proposed value | source, logic, dependency, method, prior, or uncertainty reason | evidential / methodological / prior-based / source-status-based / worldview-based / uncertainty-based | recomputed or approximated effect |

If no replacement values are supplied, the output must be classified as prior resistance, not recalculation.

## Symmetry Audit

Asymmetry checklist:
- If invoking multiverse against fine-tuning, price the multiverse prior, meta-laws, measure problem, and why law-generating reality exists.
- If invoking future origin-of-life science, distinguish possible future evidence from present evidence.
- If increasing evil/hiddenness pressure, also account for naturalism's burden to ground objective evil, moral obligation, personhood, justice, and moral outrage.
- If penalizing Christian specificity as conjunction cost, also test whether specificity increases explanatory unification.
- If discounting Christian/self-published sources as advocacy, identify the model's own secular/naturalist framework assumptions.
- If rejecting staged reasoning because later stages inherit earlier evidence, explain why that is invalid rather than merely noting that conditional inference occurred.

## Result Lenses To Keep Distinct

Current v0.74 generated outputs:

- Strict Audit Calculation: log10 `0.352594`, percent `69.250948`. This is the main hostile/fair audit lens.
- Audit Guardrail 75% Ceiling Stress-Test: log10 `0.283574`, percent `65.767567`. This is a conservative stress-test, not the main fair audit posterior.
- Guided Signal / Guided Staged Path: log10 `1.437221`, percent `96.474732`. This is the intended staged path with carried-prior conditioning, dependency caps, and rival pressure.
- Guided 90% Carry-Cap Sensitivity: log10 `1.023093`, percent `91.339078`. This preserves the old carry-cap sensitivity for comparison.
- Full Signal / Full Signal Diagnostic: log10 `4.708846`, percent `99.998045`. This is the full walked-path diagnostic under stated assumptions, not coercive proof or generic unconditional certainty.

Do not collapse these into one headline number.

## Limited Upload Workflow

- 1-file upload: use `SIGNAL_COMPACT_AUDIT_BUNDLE.md`.
- 3-file upload: use `SIGNAL_COMPACT_AUDIT_BUNDLE.md`, an evidence summary or `data/evidence/evidence.jsonl`, and a results summary such as `data/runtime/results_summary.json` or `CURRENT_RESULTS_SUMMARY_LITE.md`.
- 10-file upload / Grok-style workflow: use the starter prompt, `SIGNAL_COMPACT_AUDIT_BUNDLE.md`, results summary, synthesis model, evidence JSONL or JSON, source index, strict methodology, sensitivity tests, hostile audit protocol, and naturalism steelman.

## Final Classification

End the hostile audit with exactly one of these classifications:

- internal failure: The Signal fails on its own terms;
- contested weights: the staged structure is coherent but the model supplies replacement BFs, priors, caps, or transfer values;
- prior resistance: the model cannot or will not recalculate and is mainly expressing a prior;
- source-status discounting: the main disagreement is source classification or citation role;
- worldview resistance: the disagreement depends on a rival worldview constraint rather than a row-level defect.