# Source Coherence and Repair Protocol

**Status:** Governing compiler and editorial protocol
**Authority:** Controls diagnostics and proposed source repair; does not add or approve theory claims
**Required loop:** `source snapshot -> identity field -> diagnostic -> proposed repair -> human approval -> new source snapshot`

## 1. Purpose

The identity field must reveal where the source is disconnected, duplicated, ambiguously named, unevenly sequenced, or stronger than its proof state. It must also support a more compact canonical explanation without erasing relations, source lineage, guardrails, or real uncertainty.

This is the return path of the source compiler. It makes the publication correctable; it does not make the compiler its own authority.

## 2. Language layers

| Layer | Candidate signal | Required question |
|---|---|---|
| Identity expression | Noun or noun phrase | Does one recoverable referent or bounded identity own this name at the declared scale? |
| Change expression | Verb or verb phrase | What changed, for which identity, through which carrier, relative to what prior condition? |
| Relation expression | Preposition, conjunction, case, dependency, or explicit relation phrase | What separates and connects the participants, in which direction, with what consequence? |
| Qualification | Modifier, modal, number, tense, status, evidence, or guardrail | Which scale, time, uncertainty, proof state, or boundary does this qualify? |
| Assertion | Sentence or bounded proposition | What exactly is stated, who authored it, and where is its stable locator? |
| Identity cell | Normalized I/O/R and Y-Engine signature | Are inside, outside, relation, carrier, residue, return, mechanism, and closure declared? |
| Field | Cross-source topology | Which records duplicate, conflict, depend on, translate, contain, or correct one another? |

Parts of speech are retrieval signals, not ontological verdicts. Nominalization, pronouns, metaphor, ellipsis, polysemy, and language-specific grammar require contextual review.

## 2.1 The source relation comes before conflict

The Living Source is a versioned relational result, not a winner chosen by filename, date, or rhetorical confidence. Every admitted proposition must first be reconstructed against the same normalized identity and conditions.

An explicit author correction governs the proposition and scope it actually clarifies. It is not a global newest-file override: coherent material outside that scope remains in the field unless the correction explicitly supersedes it or an incompatibility survives the complete reconciliation gate.

Before creating a conflict edge, compare:

- exact wording and immutable source address;
- referent identity and boundary;
- inside, outside, relation, carrier, and direction;
- scale, dimension, reference frame, clock, and conditions;
- structural, cognitive, phenomenological, physical, mathematical, theological, or implementation scope;
- claimed consequence and proof mode;
- whether a source-licensed translation, refinement, or narrower reading resolves the difference.

A local cognitive definition of *mind*, for example, does not contradict universal structural *Mind* merely because both use the same word. A model implementation is not a competing ontology merely because its class name is `Identity` or `Observer`.

Use `contradicts` only when two propositions remain mutually incompatible about the same normalized subject under the same declared conditions and proof mode. Otherwise prefer a typed relation such as `supports`, `derives`, `refines`, `narrows`, `translates_across_scale`, `corresponds_structurally`, `implements`, `duplicates`, or `remains_unresolved`.

## 2.2 Coherence and persistence

New knowledge enters the same relation field and may reinforce, constrain, clarify, extend, split, weaken, or retire candidate structure. Persistence is earned when a proposition remains recoverable and coherent through several consequential relations—not when the same bytes have merely been copied many times.

The engine must therefore distinguish:

- one exact source expression and its aliases;
- byte-distinct versions of one source family;
- non-duplicate formulations that independently preserve a relation;
- derived restatements whose support depends upon an upstream source;
- implementation traces and generated feedback that cannot serve as independent theory evidence.

Every candidate persistence update must retain the supporting and constraining source spans, dependency direction, scope/reference frame, before/after state, and the rule that produced the update. Nonduplicate coherent relations strengthen structural persistence when that lineage remains visible. Repetition and duplicate wording may improve recoverability or show recurrence; they do not multiply independent support or proof.

## 2.3 Proof is a vector, not one label

Report these axes independently for every substantial proposition:

1. **internal derivation** — whether it follows from declared Y-Theory premises and transformations;
2. **cross-source reinforcement** — whether non-duplicate source formulations preserve the same normalized relation;
3. **structural correspondence** — whether a relation-preserving map into another scale or domain is declared;
4. **external empirical support** — whether the native-domain consequence is measured or independently supported;
5. **runtime predictive result** — whether a named preregistered model or experiment predicted unseen observations better than its rivals.

A proposition can be internally derived while externally untested. A structural correspondence can be exact as a mapping without proving material or causal identity. A failed software replay does not disprove an internal derivation unless that replay actually tested the derivation's declared consequence.

## 2.4 Source-level semantic review

Exact-sentence and similarity matching are queue-building instruments. They do not prove that a source has or has not been understood. A chapter can be an explicit source for a canonical claim while every sentence in that chapter still appears unmatched to a lexical candidate index.

`atlas/SOURCE_SEMANTIC_REVIEW_REGISTRY.json` is the governed bridge between file-level inventory and identity-cell compilation. One review record must:

1. bind one exact source ID, repository path, byte hash, and line count;
2. divide the complete source into ordered, contiguous section ranges with no missing or overlapping lines;
3. summarize each range without silently strengthening it;
4. relate each surviving decision to exact canonical claim IDs or leave it explicitly open;
5. distinguish an already addressed thesis, candidate guardrail, candidate wisdom, open question, and non-distinct transition;
6. preserve proof boundaries and unresolved obligations;
7. remain an attributed agent proposal until the canonical human steward approves a source change.

A completed agent review establishes recoverable semantic disposition for that exact snapshot. It does not make its paraphrases canonical, validate an external scientific claim, establish a referent identity, fill an identity cell, or approve a cross-domain relation. Those later transformations retain their own gates and lineage.

## 3. Diagnostic classes

Every diagnostic must have a stable ID, source locator, field record IDs, severity, evidence, proposed action, and review state. Supported classes include:

- same term with competing definitions;
- apparent conflict not yet normalized to one identity, scope, reference frame, and proof mode;
- different terms that appear to name one identity;
- assertion with no recoverable referent;
- change expression with no declared carrier or affected identity;
- relation with unresolved endpoints, direction, or scale;
- identity missing an outside, boundary, native mechanism, return, or closure condition;
- duplicated explanation or circular dependency;
- sequence gap or use-before-definition;
- cross-domain literalization of a structural correspondence;
- wording stronger than authority or evidence state;
- source repair that would remove a guardrail, rival, falsifier, or unresolved question;
- field record not yet represented in canonical teaching;
- canonical text not yet represented in the field.

## 4. Optimization objective

An accepted repair should improve:

- stable terminology;
- identity and relation coverage;
- dependency and teaching order;
- explicit scale and reference frame;
- proof/status consistency;
- guardrail retention;
- source traceability;
- semantic density with reduced duplication.

Word count is a secondary observation, never the primary objective. A shorter source that loses a distinction or falsifier is a regression.

## 5. Repair transaction

1. Freeze and hash the source and identity-field snapshots.
2. Generate deterministic structural diagnostics and attributed model-assisted candidates separately.
3. Content-hash deduplicate exact aliases before counting support.
4. Normalize candidate propositions by identity, scope, scale, reference frame, conditions, and proof mode.
5. Link every finding to exact source locators and field IDs.
6. Draft a minimal proposed patch with before/after meaning and affected descendants.
7. Test for lost claims, relations, guardrails, sources, proof states, and open questions.
8. Require human review and explicit authority for canonical mutation.
9. Preserve rejected and superseded proposals.
10. Create a new source snapshot and rebuild the field and all manifestations.

The website/VR may display diagnostics and collect proposed corrections. It never mutates the source or field directly.

## 6. Digital Mind boundary

The loop exhibits retained state, present-to-history comparison, mismatch detection, proposed correction, and governed return. These are useful Mind-structure operations. They do not establish feeling, moral judgment, observerhood, selfhood, or subjective consciousness. Moral orientation requires additional values, consequence models, lived or represented stakeholders, approval rules, and tests.