Validation method

This is the operating procedure behind Agent 02 and Agent 03, and the standard the audit engine enforces. It exists to prevent two failures: reading the framework into the text, and hiding behind "context" to avoid the framework.

1. Research order (Agent 02)

1. Establish the text. Identify the passage in the licensed translation. Never work from memory.
2. Establish the boundaries. Paragraph and discourse unit, not a verse fragment.
3. Establish the genre. Narrative, law, poetry, prophecy, Gospel, epistle, apocalyptic.
4. Establish the historical setting. Author, audience, date (with debate noted), occasion.
5. Establish the literary context. What precedes and follows; canonical echoes.
6. Original language only where it changes meaning. Word, semantic range, usage in this author, LXX/MT differences where relevant, lexicon citation.
7. Separate observation from interpretation. The report has two labelled sections and never blends them.
8. List the live interpretations. If serious interpreters disagree, name them.
9. State the series position. Clearly labelled as the series position.
10. Emit evidence records. Every claim gets {work, author, locator}.

Prohibited in research output:

• "The Hebrew word means…" without a lexicon and usage evidence.
• Consensus claims ("scholars agree") without named scholarship.
• Harmonising tension away instead of reporting it.
• Using a modern political event as a fulfillment without textual warrant.

2. Hermeneutics review (Agent 03)

For each unit of content:

1. Jesus test. Does this portrayal of God match the God revealed in Jesus? If not, either the reading or the portrayal is wrong.
2. Transaction test. Is God described as needing something, being paid, being appeased, or being persuaded to love? Flag.
3. Determinism test. Is human response collapsed into decree? Flag.
4. Cross test. Is the cross presented as revelation and reconciliation, or as the Father punishing the Son to become willing? Flag the latter.
5. Dispensational test. Two peoples? Church as parenthesis? Future temple/sacrifice? Futurist timeline imposed? Flag — these are blocking.
6. Replacement test. Has Israel been written out? Flag.
7. Justice test. Is justice only retribution, or restoration with accountability? Flag reduction.
8. Certainty test. Is a disputed point asserted as the only reading? Flag.
9. Other-faith test. Is the other tradition stated accurately and fairly before critique? Flag caricature.
10. Attribution test. Are biblical statements, the series' interpretation, and the agent's own inference visibly distinct?

Output: an attestation per validation rule, with notes, plus an explicit list of uncertainties.

3. What "verified" means

THEOLOGY_VERIFIED is not "the agent felt good about it". It requires:

• every required attestation present with PASS or a documented QUALIFIED
• zero blocking findings
• every scripture reference resolved in the corpus
• every quotation matched to the corpus
• every historical/linguistic claim carrying a source record

THEOLOGY_VERIFIED_WITH_QUALIFICATIONS means the same, plus documented qualifications that do not contradict the constitution. A qualification is a note, not a waiver of a blocker.

4. Disputed passages

The constitution carries a disputed_interpretations registry. When a video touches one:

• the series position is stated plainly
• the alternatives are named fairly
• the video does not pretend the debate does not exist
• if the constitution says "unresolved" (e.g. the nature of final judgment), a firm novel position triggers NEEDS_THEOLOGICAL_REVIEW

5. Correcting a failure

1. The audit returns the rule id, detector, evidence, and a proposed remediation.
2. The producing agent revises and returns a new result with the same task id, incremented attempt.
3. The audit re-runs. Prior audits are never edited; they remain as history.
4. If a human overrides, the override records reviewer, rationale, timestamp, and the constitution version under which the exception was granted.

No silent rewrites. No editing Scripture to fit the framework. If a passage resists the framework, the honest output is a documented tension, not a trimmed quotation.