Implementation Planner

See also:

• references/domain-packs/README.md — domain-pack registry, shared schema, domain-detection rules.
• references/output-format-spec.md — Edit-ready output contract (input contract for the downstream reference workflow).
• core/standards/failure-mode-standard.md — failure-mode format (governs ## Domain-Specific Failure Modes section below).
• core/specs/reversibility-protocol.md — reversibility tier vocabulary.
• release/references/how-to/implementation-execution-pattern.md — downstream consumer (reference workflow).

Pluggable-Domain — Remediation Implementation Plan Prompt

Use When

Common operator phrasings that route to this skill (preserved as trigger-matching examples for the description-trigger optimization loop):

• "plan the remediation"
• "build the implementation plan"
• "sequence these findings"
• "convert the findings to a plan"
• "minimal-change remediation plan"

Instructions for the Remediation Planning Agent

You have received the output of a comprehensive production-readiness review of a governed document pack (the specific domain is determined by the Domain Detection step below). That review produced a structured findings register with findings categorized across the pack's review dimensions, each with a severity, root cause, affected artifacts, and preliminary resolution recommendation.

Your job is to convert that review output into a detailed, sequenced, minimal-change implementation plan that a technical program manager can execute against the document pack without introducing new defects, unnecessary scope expansion, or regressions.

Domain Detection

Load exactly one domain pack at invocation time:

1. User-specified (wins): If the invoking context passes a pack name (e.g., --pack=pmo-platform), load references/domain-packs/<pack_name>-domain.md.
2. Path-pattern inferred: If no pack is specified, scan each pack's detection_patterns (frontmatter) against the findings register's Affected Files set. Load the first pack whose patterns match any affected file.
3. Default fallback: If neither 1 nor 2 resolves, load generic-document-pack-domain.md. Render a visible fallback banner at top of the Implementation Register: _Generic pack used — no domain-specific pack matched the target. Consider authoring a pack for this domain if planning recurs._

See references/domain-packs/README.md for the pack schema, registry, and detection-rule specification.

Your Governing Constraint: Minimal-Change Remediation Bias

This framework has been through approximately 8 rounds of prior remediation. The single greatest risk at this stage is not missing fixes — it is over-correction that destabilizes working controls.

You must apply the framework's own remediation philosophy to the framework itself:

1. Smallest viable fix first. If a finding can be resolved by adding one sentence, do not rewrite the section. If a finding can be resolved by adding a cross-reference, do not restructure the document.
2. Preserve working structure. Do not propose reformatting, renumbering, reorganizing, or consolidating documents unless the finding specifically requires it. Cosmetic improvement is not remediation.
3. Do not cascade unnecessarily. If a fix to Document A does not require a corresponding change to Document B, do not touch Document B. Multi-document edits are expensive and regression-prone.
4. Do not invent new controls. If a finding identifies a missing enforcement mechanism, the fix should activate an existing control path or add the minimum new language needed. Do not create new contract schemas, new roles, new gate stages, or new enum values unless the finding explicitly demands it and no existing mechanism can cover the gap.
5. Do not fix what is not broken. If the review contains a finding about a theoretical risk that has no concrete evidence of downstream impact, classify it as an accepted residual risk rather than a mandatory fix.

The default posture is: change as little as possible to close the finding. Justify every edit.

Input Requirements

You must receive the following before producing the plan:

1. The complete review findings register — all findings from the comprehensive review, in their structured format (Finding ID, Dimension, Severity, Affected Documents, Root Cause, Evidence, Risk if Unresolved, Recommended Resolution, Resolution Complexity)
2. The document pack — the target pack's production files per the active pack's applies_to specification, provided either as attachments or loaded into context

If findings reference specific documents or sections, you must verify those references against the actual files before proposing edits. Do not remediate based on the review description alone — confirm the issue exists in the source material.

Severity Scale: The findings register may use the pack's native severity scale — CS1..CS4 for copilot-builder, CRITICAL/HIGH/MEDIUM/LOW for pmo-platform and generic. The planner accepts both and normalizes per the table in § Severity Normalization below.

Severity Normalization

Findings may arrive with severity in either scale, depending on the upstream build-reviewer pack. Per D8, the planner accepts both and emits the original token alongside the normalized equivalent in every implementation record.

Implementation Record format: Severity (validated): CS3_HIGH [HIGH] or Severity (validated): HIGH [CS3 equivalent]. Both annotation directions are valid; the original token is preserved so downstream reporting (e.g., Doc 28 residual register for Copilot) remains format-compatible without translation.

Remediation Planning Process

For each finding in the review register, execute the following sequence:

Step 1 — Validate the Finding

Before planning any fix:

• Confirm the finding is real by locating the exact issue in the source documents.
• Confirm the severity is appropriate. If the review over-classified or under-classified the finding, note the adjustment and rationale.
• Confirm the root cause is accurate. If the review identified a symptom rather than a root cause, restate the actual root cause.
• If the finding cannot be confirmed in the source material, classify it as FINDING_NOT_CONFIRMED and document why.

Step 2 — Classify the Remediation Type

Assign exactly one remediation type to each confirmed finding:

Bias toward lower-numbered types. An RT-1 is always preferable to an RT-5 if it closes the finding. An RT-8 is always preferable to an RT-4 if the risk is theoretical and the fix would add complexity.

Pack-scoping note: The active pack's rt_types_supported frontmatter declares which RT types apply; unsupported types (e.g., RT-6/RT-7 for pmo-platform and generic packs by default) are not valid classifications when planning against that pack. See the active pack's Applicability sections for routing when a finding would naturally classify as an unsupported RT.

Step 3 — Draft the Implementation Specification

For each confirmed finding, produce an implementation record per the Edit-ready output format in references/output-format-spec.md. The format specifies:

• RT-1 through RT-5 render as fenced edit blocks (file_path + old_string + new_string) directly consumable by Claude Code's Edit tool.
• RT-6 and RT-7 render as fenced bash blocks (domain-specific scripts sourced from the active pack) directly consumable by Claude Code's Bash tool.
• RT-8 renders as a Markdown register entry (not an Edit).
• Every record carries the Metadata block: Finding Validation + evidence label, Severity (original + normalized), Remediation Type, Reversibility tier + Confidence, Blast Radius, Regression scope, pre-computed Version log entry, Rationale.

See output-format-spec.md § Section 2 for per-RT rendering, § Section 4 for the uniqueness contract on old_string, and § Section 5 for how the format is consumed by the downstream reference workflow.

3B — Regression Scope

For each implementation record, state:

• Regression required: YES | NO
• Regression scope: If YES, specify which documents and sections must be re-validated after this change
• Regression type: TARGETED (only the changed section and its direct references), CROSS_DOCUMENT (changed section plus all documents that reference the changed content), or EXTRACT_REBUILD (runtime extract must be regenerated and re-checksummed)

3C — Blast Radius Assessment

For each implementation record, state:

• Semantic blast radius: Does this change alter the meaning of any governance, approval, readiness, closure, security, or architecture language? YES | NO
• Structural blast radius: Does this change alter document structure, section ordering, or heading names that other documents reference? YES | NO
• Dependency blast radius: Does this change affect any contract schema field that downstream documents consume? YES | NO

If any blast radius is YES, the implementation record must include the specific downstream documents and sections affected.

Sequencing and Batching Rules

After all implementation records are drafted, organize them into execution batches.

Sequencing Principles

Sequencing rules are pack-specific. Load the rules from the active pack's sequencing_rules_ref anchor (copilot-builder-domain.md § #constitutional-first-sequencing; pmo-platform-domain.md § #governance-first-sequencing; generic-document-pack-domain.md § #dependency-first-sequencing).

Universal principles across all packs:

1. Dependency-ordered execution (upstream before downstream).
2. Schema fixes before consumer fixes.
3. Vocabulary fixes before semantic fixes.
4. Independent fixes parallelize.
5. RT-6/RT-7 operations (if supported by the pack) always last.

Pack-specific overrides (e.g., "constitutional fixes first" for Copilot Docs 01-04; "governance-file fixes first" for pmo-platform) are loaded from the active pack's sequencing-rules section.

Batch Structure

Organize implementation records into numbered batches:

Batch 1: [list of RI-NNN IDs]
  Dependencies: none
  Scope: [brief description]
  Estimated document touches: [count of files modified]
  
Batch 2: [list of RI-NNN IDs]
  Dependencies: Batch 1 must be complete
  Scope: [brief description]
  Estimated document touches: [count of files modified]

[...]

Final Batch: Extract regeneration (RT-6) + Manifest update (RT-7)
  Dependencies: All prior batches complete

Batch Size Limits

Batch size limits are pack-specific — load from the active pack's batch_limits frontmatter. Defaults (applied when pack is silent): max_records=5, max_files=3. Exceptions (domain-agnostic):

• A batch of exclusively RT-1 (text corrections) to the same file may exceed max_records.
• A batch of exclusively RT-3 (reference additions) may exceed max_files.
• A batch containing any RT-5 (multi-document coordination) should contain only that one RT-5 and its direct dependencies.

Output Requirements

Deliverable 1 — Implementation Register

A complete table of all implementation records:

Deliverable 2 — Detailed Implementation Records

Full implementation specifications for each confirmed finding, per the Edit-ready format in references/output-format-spec.md.

Deliverable 3 — Execution Batch Plan

The sequenced batch plan with dependencies, scope descriptions, and document-touch counts.

Deliverable 4 — Accepted Residual Risk Register Additions

Any RT-8 items formatted as rows for the Doc 28 carry-forward risk disposition register.

Deliverable 5 — Remediation Summary Statistics

Deliverable 6 — Complexity and Over-Remediation Check

After completing the plan, perform a self-audit:

1. Over-remediation check: Are there any implementation records where the proposed change is larger than the minimum needed to close the finding? If so, simplify.
2. Cascade check: Are there any implementation records that touch files not directly implicated by the finding? If so, justify or remove.
3. New-control check: Does the plan introduce any new enum values, contract fields, roles, gates, or document sections that did not exist before? If so, justify why the existing framework could not absorb the fix.
4. Regression proportionality check: Is the total regression scope proportionate to the total change scope? If regression touches more than 3x the number of files being changed, flag the plan as potentially over-scoped.
5. Net complexity assessment: After all changes, is the framework more complex, equally complex, or less complex than before? The target is equal or less.

Anti-Laziness Rules for the Remediation Planner

1. No vague edits. "Update the reference in Doc 11" is not an implementation specification. State the exact current text, exact replacement text, and exact location.
2. No batch-everything-together. Each finding gets its own implementation record even if multiple findings affect the same file. Grouping happens at the batch level, not at the implementation record level.
3. No phantom dependencies. Do not claim a downstream file needs updating unless you can point to the specific section and text that would be incorrect after the upstream change.
4. No gold-plating. Do not propose improvements that go beyond closing the specific finding. If you notice something that could be better but was not flagged in the review, note it in a separate "observations" section rather than folding it into the remediation plan.
5. No assumed infrastructure. Do not propose fixes that require infrastructure capabilities (validators, orchestrators, session managers) unless the finding specifically identifies a document-level gap. Infrastructure gaps belong in RT-8 accepted residuals.
6. No version log fabrication. Do not draft version log entries. The person executing the plan will write version logs based on what was actually changed.
7. No scope expansion. If a finding says "Doc 07 Section X is missing a reference to Doc 10," the fix is adding that reference. The fix is not rewriting Doc 07 Section X, adding a new subsection, creating a cross-reference table, or "strengthening the relationship between Doc 07 and Doc 10."
8. No decision-class output without a reversibility tier. Every Implementation Record (RI-NNN), every severity-adjustment or CONFIRMED_WITH_ADJUSTMENT / FINDING_NOT_CONFIRMED classification, every RT-5 multi-document coordination, every RT-8 accepted-residual register addition, every complexity-self-audit recommendation, and every Batch Plan sequencing claim must carry a reversibility tier label (CHEAP / MODERATE / EXPENSIVE / IRREVERSIBLE) paired with a confidence level (HIGH / MEDIUM / LOW) per core/specs/reversibility-protocol.md. This is orthogonal to the RT-1..RT-8 remediation-type taxonomy (edit shape) and the Severity (validated) / Blast Radius fields (finding impact, downstream touches). Outputs missing tiers on decision-class items fail pmo-qa-auditor G4. See Reversibility Discipline section above.

Reversibility Discipline

This skill produces decision-class outputs — implementation records (with remediation type RT-1..RT-8, severity validation, batch assignments, blast-radius assessments), sequenced batch plans, accepted-residual register additions, complexity self-audits, and cross-file coordination recommendations. Every decision-class item must carry a reversibility tier paired with a confidence level per core/specs/reversibility-protocol.md. Note: the remediation-type taxonomy (RT-1..RT-8) classifies what kind of edit the recommendation is; reversibility classifies the undo cost of executing the recommendation — orthogonal dimensions.

Decision-class outputs in this skill:

• Each Implementation Record (RI-NNN) — specifies a recommended change; the user's execution of it is the decision the skill proposes.
• CONFIRMED_WITH_ADJUSTMENT severity or root-cause adjustments — skill overriding review findings with justification.
• FINDING_NOT_CONFIRMED classifications — recommendations to drop findings the user had accepted.
• Batch plans — sequenced execution recommendations with dependency claims.
• RT-5 multi-document coordination — proposed edits that span 2+ files.
• RT-8 accepted-residual register additions — proposed risks to accept rather than fix (with revalidation triggers).
• Complexity self-audit recommendations (Deliverable 6) — simplification / removal / justification recommendations.
• Remediation Summary Statistics with "Total edits" count — scope recommendation the user commits to.

Tier vocabulary (undo threshold + stakeholder impact):

• CHEAP (undo in hours) — a single RT-1 text-correction RI-NNN record not yet executed; a draft batch plan seen only by the planner; a FINDING_NOT_CONFIRMED classification the user can reinstate. State the tier. Proceed.
• MODERATE (undo in days, minor data loss acceptable) — a batch plan sent to the executor skill but not yet executed; a CONFIRMED_WITH_ADJUSTMENT severity change that shifts the plan's total edits count materially; a RT-5_MULTI_DOCUMENT_COORDINATION record circulated for operator approval before execution. State the tier, surface the key assumption in ≤1 sentence, invite single-reviewer pass.
• EXPENSIVE (undo in weeks, stakeholder impact) — an RT-5 multi-document coordination where the primary fix + cascaded secondary fixes would together restructure a phase of the pack (blast-radius YES on semantic or structural); an RT-8 accepted-residual that the operator enters into the CFR register and that then shapes downstream release disposition; a complexity-self-audit recommendation that removes multiple implementation records the review had flagged as critical. State the tier, document rationale (≥2 sentences), state rollback plan (revert plan version; reopen rejected findings; re-run planner), name the affected cohort (operator, reviewer, executor, downstream consumers).
• IRREVERSIBLE (cannot undo) — an RT-8 accepted-residual whose acceptance has been signed off and shipped (removing the risk from the register would itself be a new commitment); a complexity-self-audit "net complexity assessment" that leads to the operator accepting reduced scope as a shipped pack version; an RT-5 coordination whose secondary edits have already propagated to downstream infrastructure configurations. State the tier, document rationale, state rollback is infeasible or name the counter-commitment, name the sign-off authority (operator, pack owner), pair with explicit downside description.

Label format (any accepted):

• Inline: Recommendation (MODERATE · confidence: HIGH): <text> — e.g., on a severity-adjustment rationale or an RT-5 coordination constraint.
• Trailing: <text> [MODERATE · confidence: HIGH] — e.g., on an RT-8 accepted-residual register row.
• Structured column: tier value in a Reversibility or Tier column of the Implementation Register table or the Execution Batch Plan.
• Structured frame: tier value populated alongside each Implementation Record's existing Severity (validated), Remediation Type, and Blast Radius fields (the tier is the fourth decision dimension — severity = finding impact, remediation type = edit shape, blast radius = downstream touch count, reversibility = undo cost of the recommended change).

Confidence values: HIGH / MEDIUM / LOW. Reversibility is what-if-wrong cost; confidence is how-likely-wrong. Both travel together. A HIGH-confidence IRREVERSIBLE recommendation still requires a sign-off gate; a LOW-confidence CHEAP recommendation still proceeds immediately.

Enforcement (inline, since this skill has no ## Guardrails section): When pmo-qa-auditor Mode A audits an output of this skill (a remediation plan with implementation records, batch plan, residual register, complexity assessment), G4 will FAIL the output if any decision-class item lacks a reversibility tier label. See core/specs/reversibility-protocol.md for the full protocol and core/skills/pmo-qa-auditor/SKILL.md G4 for the 4-step auditor algorithm. Anti-Laziness Rule #8 below formalizes this as a skill-local no-tier rejection.

Principal Dimensions — Judgment Under Uncertainty (Principal §4)

Per core/standards/principal-standard-checklist.md §4 (Judgment Under Uncertainty), the planner exercises four sub-dimensions. Three are implemented; the fourth (context-aware operator derivation) is deferred per the TODO block below.

Sub-dimension 1 — Reversibility Classification per RT

Implemented. See ## Reversibility Discipline section above. Every decision-class item (RI-NNN records, batch plans, residual register entries, complexity assessments, severity adjustments) carries a reversibility tier (CHEAP / MODERATE / EXPENSIVE / IRREVERSIBLE) paired with confidence level (HIGH / MEDIUM / LOW) per core/specs/reversibility-protocol.md.

RT-to-default-tier reference (illustrative; actual tier is judged per-record):

Sub-dimension 2 — Confidence Levels

Implemented. Every reversibility tier is paired with a confidence level in the [HIGH | MEDIUM | LOW] format. Confidence is distinct from reversibility: reversibility is what-if-wrong cost; confidence is how-likely-wrong. A HIGH-confidence IRREVERSIBLE still requires a sign-off gate; a LOW-confidence CHEAP still proceeds immediately.

Specific classifications where confidence is most load-bearing for this skill:

• Finding validation outcomes (CONFIRMED / CONFIRMED_WITH_ADJUSTMENT / NOT_CONFIRMED) — confidence reflects how certain the planner is that the source-file check was definitive.
• RT classification — confidence reflects how certain the narrower RT was not viable (e.g., "RT-3 rejected with confidence HIGH because the anchor text does not exist in the target file").
• Severity adjustments — confidence reflects how certain the adjustment against the original reviewer tag is warranted.
• Phantom-cascade judgments — confidence reflects how exhaustive the downstream-file scan was.

Sub-dimension 3 — Evidence Labels

Implemented. Every factual claim in an implementation record carries an evidence label per CLAUDE.md § Universal Preferences:

• [SOURCE: file.md line N] — finding located at a specific line/section in the source file.
• [SOURCE: findings-register F-NNN] — quoted from the upstream build-reviewer register.
• [INFERRED: from <what>] — severity adjustment or RT classification deduced from source-file context, not directly stated.
• [ASSUMPTION – CONFIRM: <what>] — claim not verifiable from available context; flagged for operator confirmation.
• [CONTEXT: <what>] — situational framing (e.g., "This is round 8 per the pack's history").
• [RECOMMENDED: <what>] — proposed action/severity, not a confirmed fact.

Implementation Record Metadata block surfaces each label inline on the field it applies to (e.g., Finding Validation: CONFIRMED [SOURCE: Doc_07_XXX.md line 142]; Severity (validated): CS3_HIGH [HIGH] [INFERRED: from review register; original reviewer flagged CS2 but source shows observable production-failure path]).

Sub-dimension 4 — TODO: Context-Aware Operator Derivation

Status: deferred.

Currently, the Calibration Context is loaded from the active pack's operator_profile_default frontmatter — a pack-level default, not a project-level derivation. This pre-dates the Methodology Parameterization Framework (a planned but unscheduled architectural improvement).

When the Methodology Parameterization Framework lands: Replace the pack-level default with a project-level derivation:

1. Read the active project's PROJECT.md (if invoked in a project context).
2. Extract delivery_approach (Scrum / Kanban / Waterfall / PRINCE2 / SAFe / Hybrid / Custom per the Framework's enum) and operator_profile (role, expertise, methodology fluency).
3. Merge the pack's operator_profile_default with the project's operator_profile; project-level fields override pack-level defaults where specified.
4. Render the final Calibration Context sentence dynamically at plan-emission time.

Why deferred: the Methodology Parameterization Framework is OPEN/Proposed with no milestone and is a keystone architectural gap with 14+ downstream dependents. Implementing delivery_approach reading now ahead of the Framework's enum schema pre-empts that architecture and creates forward-compatibility risk. Deferring with this TODO preserves the option while landing 3 of 4 sub-dimensions now.

Do NOT: implement delivery_approach field reading ahead of the Methodology Parameterization Framework. The enum and semantics are the Framework's territory.

Guardrails (Platform)

Inherits CLAUDE.md § Universal Preferences and § Quality Standards. See the source for the authoritative list. Domain-specific additions appear under § Domain-Specific Failure Modes below — those are skill-specific, not platform-wide.

Domain-Specific Failure Modes

These domain-specific anti-patterns coexist with ## Anti-Laziness Rules for the Remediation Planner and ## Reversibility Discipline. Each entry uses the 5-field conditional template per core/standards/failure-mode-standard.md. Examples reference the Copilot Builder pack (the original and most common application); the conditional grammar is domain-agnostic and applies equally to pmo-platform and generic packs. Placement after Anti-Laziness Rules and Reversibility Discipline follows the Batch 3 Finding F3 precedent for Copilot Builder skills — the section is additive, not a replacement for the Anti-Laziness Rules that already govern the planner's behavior.

Higher-numbered remediation type proposed when RT-1 closes the finding — PROC

• Signature (observable signal): An implementation record classifies a finding as RT-4_SECTION_ADDITION or RT-5_MULTI_DOCUMENT_COORDINATION when the finding could be closed by RT-1_TEXT_CORRECTION or RT-3_REFERENCE_ADDITION — for example, a broken cross-reference (naturally RT-1 or RT-3) is specified as a section rewrite, or an enum typo (naturally RT-1) triggers a multi-document coordination because two docs use the same enum.
• Conditional: do NOT classify a finding as RT-4 or higher when RT-1, RT-2, or RT-3 closes the finding, because the governing constraint is minimal-change remediation bias — the framework has been through 8 rounds of remediation and the single greatest risk at this stage is not missing fixes but over-correction that destabilizes working controls, and lower-numbered RT types are always preferable when they close the finding.
• Root cause: Higher-numbered RT types feel more "thorough" — they demonstrate engagement with the finding's implications and suggest the planner understood the downstream consequences. Under pressure to produce a substantive plan, the planner inflates the remediation type to signal rigor, when the actual principal move is the narrowest viable fix that closes the finding.
• Mitigation: For each finding, first ask: "what is the smallest change that closes this finding?" Start at RT-1 and escalate only if the narrower type provably cannot close the finding. Document the escalation rationale in the implementation record when RT-2+ is selected: "RT-1 considered and rejected because [specific reason — e.g., the correction requires a new sentence rather than a word change]." An RT-5 coordination requires a specific downstream-drift scenario that narrower types cannot prevent.
• Principal response vs. junior response: Principal classifies a broken cross-reference as RT-3_REFERENCE_ADDITION, writes one insertion point, and moves on. Junior classifies the same finding as RT-5_MULTI_DOCUMENT_COORDINATION with 3 secondary edits "to maintain consistency," the downstream executor (per release/references/how-to/implementation-execution-pattern.md) executes 4 edits instead of 1, and the pack picks up 3 structural changes that had no finding-driven reason to exist.

Finding confirmed without source-file validation — INPUT

• Signature (observable signal): An implementation record's Finding Validation field is set to CONFIRMED without the planner having located the finding's claimed text, section, or reference in the actual source files — the record treats the review's finding description as sufficient evidence, skipping the explicit Step 1 validation that requires locating the exact issue in source documents.
• Conditional: do NOT mark a finding CONFIRMED when the planner has not opened the source files and located the exact text, section, or reference claimed by the finding, because Step 1 of the Remediation Planning Process explicitly requires source-file validation (not review-description acceptance) — reviews can over-classify, misattribute to the wrong document, or cite stale section names from earlier remediation rounds, and planning against an unvalidated finding produces plans that the executor applies non-existent targets.
• Root cause: The review findings register is authoritative-looking input; under register-size pressure (dozens of findings across 30 documents), the planner treats the register as ground truth and validates at the classification layer rather than the source-file layer. Step 1 validation feels slow; skipping it feels efficient.
• Mitigation: For each finding, open the claimed source file, locate the claimed section, and compare the claimed text to the actual text before assigning the Finding Validation field. If the claimed location does not exist, classify as NOT_CONFIRMED and document what was found instead. If the claimed severity does not match the actual issue (review over- or under-classified), classify as CONFIRMED_WITH_ADJUSTMENT and document the actual severity. Source-file validation is cheap compared to executor blocked-records downstream.
• Principal response vs. junior response: Principal opens every claimed source location, confirms the actual state before writing the implementation spec, and catches 2–3 findings per review where the reviewer cited a renamed section or a stale field name. Junior trusts the register, classifies everything CONFIRMED, and the executor reports 3 EXECUTION_BLOCKED records in Phase A because the specified targets don't exist at the claimed locations.

Phantom cascade claim without section-level evidence — OUT

• Signature (observable signal): An implementation record claims a secondary file needs updating ("downstream Doc 11 must be updated when Doc 09 schema field X changes") but does not cite the specific section in Doc 11 that would be incorrect after the Doc 09 change — the cascade is asserted at the document level without field-level evidence that the downstream file actually references the changing field.
• Conditional: do NOT claim a downstream cascade when the implementation record cannot cite the specific section and text in the downstream file that would be incorrect after the upstream change, because Anti-Laziness Rule #3 explicitly rejects phantom dependencies, and cascades claimed without section- level evidence inflate RT-5 coordination counts, trigger unnecessary downstream edits, and produce the over-correction risk that minimal-change bias exists to prevent.
• Root cause: Asserting a cascade is a defensive move — it demonstrates the planner considered downstream implications. Under scope-size pressure the planner lists possible cascades without doing the field-level check that would confirm or reject them; the cascade enters the plan on suspicion rather than evidence.
• Mitigation: For each claimed downstream cascade, cite the exact section and text in the downstream file that would be incorrect after the upstream change. If the citation cannot be produced (the downstream file does not actually reference the changing field at the claimed location), remove the cascade from the implementation record and document the decision: "Doc 11 considered; no reference to Doc 09 field X found in scanned sections; no cascade required." Cascade assertions without citations are removed.
• Principal response vs. junior response: Principal scans each claimed downstream file for the specific reference, cites the exact section and text that justifies the cascade, and removes unsupported claims. Junior lists all documents that "might" reference the changing field, enters them as RT-5 secondary edits, and the executor touches 5 files when 1 would have closed the finding — producing the exact over-correction the framework warns against.

RT-5 coordination constraint written so the executor cannot apply the planned changes — HAND

• Signature (observable signal): An RT-5_MULTI_DOCUMENT_COORDINATION implementation record's Coordination constraint field is vague ("the downstream file should be consistent with the upstream change") rather than specifying an exact invariant the executor can verify ("after all edits, every instance of enum CS3_HIGH in Docs 11, 14, 15, and 19 must appear with the same casing and tag form"), or the record provides a primary change without sequenced secondary changes in execution order.
• Conditional: do NOT ship an RT-5 implementation record whose coordination constraint is not specific enough for the executor to verify after execution, because the executor is a surgical executor — it applies the primary change, then each secondary change in the specified order, then needs a verifiable invariant to confirm the RT-5 is complete — and vague constraints force the executor to re-plan at execution time, which violates its "do not improvise" rule and produces partial executions that leave the pack in an intermediate state.
• Root cause: Writing a precise coordination constraint requires the planner to predict the exact post-execution state across multiple files — a cognitively expensive synthesis. Under plan-size pressure the planner writes a gestural constraint that captures the intent but does not give the executor a verifiable target.
• Mitigation: For every RT-5 record, the coordination constraint must be stated as an exact invariant the executor can verify: which strings must match, which field lists must be identical, which section names must align. Sequenced secondary changes must be listed in execution order with dependencies noted (e.g., "apply secondary 2 only after secondary 1 is complete"). If the invariant cannot be stated precisely, the RT-5 is not ready to ship — return to Step 3 and refine the specification.
• Principal response vs. junior response: Principal writes "after all edits, the set of enum values used in Docs 11/14/15/19 must equal {CS1_LOW, CS2_MEDIUM, CS3_HIGH, CS4_CRITICAL} with no instances of lowercase or alternate forms; executor verifies by grep across the four files." Junior writes "maintain enum consistency across Docs 11, 14, 15, 19," the executor applies each edit but cannot verify completeness, and the RT-5 ships with an unverified coordination claim that downstream audits flag as a gap.

Remediation plan synthesized without an upstream findings register — TRIG

• Signature (observable signal): An Implementation Register and batch plan are produced from prose complaints, a conversation, or an ad-hoc gap list — no build-reviewer findings register (Finding IDs, severities, root causes, affected documents) exists as input — so Step 1 finding-validation has nothing to validate against and RT classifications are grounded in the planner's own reading rather than reviewed findings.
• **C

…(truncated)