Library health audit findings: guardrail compression, determinism fixes, structural cleanup

[Alan-Jowett]

Source

Identified by the audit-library-health template run on 2026-04-08 (full library, all 4 passes, 197 components).

High Priority

P3-001: Compress self-verification Rule 2 (7,400 token savings)

• Files: protocols/guardrails/self-verification.md
• Issue: Rule 2 (Citation Audit) restates content already in anti-hallucination Rules 1-4. Amplified 74x across every template.
• Fix: Replace Rule 2 body with cross-reference to anti-hallucination. Est. savings: 7,400 tokens.

P4-001: review-code L85 has no observable criteria

• Files: templates/review-code.md
• Issue: "Are abstractions appropriate (not too much, not too little)?" is subjective.
• Fix: Decompose into 3 observable checks (depends on >3 unrelated concepts? duplicates sibling logic? unclear responsibility boundaries?).

Medium Priority — Guardrail Determinism (74x blast radius)

P4-005 + P4-006: anti-hallucination uses subjective language

• Files: protocols/guardrails/anti-hallucination.md
• Issue: L27 "reasonable conclusion" and L48 "low confidence" are unanchored.
• Fix: Replace "reasonable" with "stated chain of logical steps". Anchor "low confidence" to ASSUMED premise count from Rule 1.

P4-008: self-verification determinism check is non-deterministic

• Files: protocols/guardrails/self-verification.md
• Issue: Rule 6 L97 asks "structurally similar" without defining what that means.
• Fix: Operationalize as "same section headings, same number of items per section (+-20%), same classification labels."

P3-002 + P1-001: Coverage template duplicated in two guardrails

• Files: protocols/guardrails/operational-constraints.md, protocols/guardrails/self-verification.md
• Issue: Identical 4-field coverage statement (Examined/Method/Excluded/Limitations) in both. Changes to one will drift.
• Fix: Keep in operational-constraints Rule 9, replace self-verification Rule 3 with cross-reference. Saves 1,140 tokens.

P4-002 + P4-003: review-code maintainability checks subjective

• Files: templates/review-code.md
• Issue: L83 "clear and readable" and L86 "SOLID, DRY, or other design principles" have no observable criteria.
• Fix: Replace with enumerated checklists (4 readability indicators, 4 design violations).

Medium Priority — Structural

P1-002: adversarial-falsification applicable_to is stale

• Files: protocols/guardrails/adversarial-falsification.md
• Issue: Lists templates that don't use it, missing 5 templates that do.
• Fix: Update to reflect actual usage.

P1-003: 4 orphan protocols with zero template consumers

• Files: definition-of-done, tool-reliability-defense, input-clarity-gate, fixed-point-verification
• Issue: Defined in manifest but no template references them.
• Fix: Either add to relevant templates or document intended composition patterns.

P1-004: plan-implementation / plan-refactoring near-duplicate

• Files: templates/plan-implementation.md, templates/plan-refactoring.md
• Issue: Identical composition (persona, protocols, format), differ only in input framing.
• Fix: Merge into single template with mode parameter.

Low Priority

• P1-005: triage-issues / triage-pull-requests near-duplicate
• P1-006: review-code / review-cpp-code composition candidate
• P1-007: Non-standard epistemic markers in hardware protocols
• P1-010: protocol-engineering pipeline missing consumes chain
• P1-011: 10 templates missing format field
• P3-005: audit-library-health template bloated (62% density)
• P3-006: spec-extraction-workflow repeated boilerplate

Informational (no action needed)

• P1-008: Three confidence scales (intentional design)
• P1-009: 44% templates use investigation-report (monitor)
• P1-012: validation-plan / protocol-validation-spec overlap (by design)
• P3-003: 14 unreferenced protocols (composable by design)
• P3-004: CWE taxonomies large (expected)
• P3-007: 3 unreferenced formats (delivery wrappers)
• P2-003: Contoso placeholder (harmless)
• P2-004: Academic citations lack DOIs (nice-to-have)