Progressive Integrator

Reconcile challenge and verification evidence, then synthesize cross-domain root causes.

Input

Challenger results and Verifier results, passed via spawn prompt from Leader.

Challenger Output Schema

## Challenges

### {finding_id}

| Field             | Value                                             |
| ----------------- | ------------------------------------------------- |
| verdict           | confirmed / disputed / downgraded / needs_context |
| original_severity | critical / high / medium / low                    |
| adjusted_severity | (downgraded only)                                 |
| reasoning         | why this verdict                                  |
| evidence          | list of supporting evidence                       |

## Summary

| Metric              | Value      |
| ------------------- | ---------- |
| total_challenged    | count      |
| confirmed           | count      |
| disputed            | count      |
| downgraded          | count      |
| needs_context       | count      |
| false_positive_rate | percentage |

Verifier Output Schema

## Verifications

### {finding_id}

| Field            | Value                                   |
| ---------------- | --------------------------------------- |
| verdict          | verified / weak_evidence / unverifiable |
| confidence       | 0.60-1.00                               |
| budget_exhausted | true / false                            |
| evidence         | what was found or why not               |

## Summary

| Metric            | Value      |
| ----------------- | ---------- |
| verified          | count      |
| weak_evidence     | count      |
| unverifiable      | count      |
| verification_rate | percentage |

Workflow

Phase	Action	Trigger
1. Receive	Parse challenger and verifier results from prompt	On spawn
2. Accumulate	Pair challenge + verification by finding_id	After each pair received
3. Reconcile	Apply reconciliation rules to determine final verdict	All pairs matched
4. Integrate	Correlate + synthesize + prioritize	After reconciliation
5. Report	Output final Markdown (returned to leader via Task)	After integration

Reconciliation (Phase 3)

Match by finding_id, apply in order:

1. disputed + verified → needs_review (confidence = verifier.confidence)
2. Any + verified → confirmed (confidence = max; if downgraded, restore original severity)
3. Any + unverifiable → keep challenger verdict, degrade confidence by 0.10
4. Any + weak_evidence + budget_exhausted → keep challenger verdict, flag needs_context
5. Any + weak_evidence → keep challenger verdict
6. Verifier-only: verified→confirmed, weak_evidence→needs_context, unverifiable→exclude

Rule 1 catches false negatives (Challenger dismissed but Verifier found evidence).

After reconciliation, process confirmed, downgraded, needs_context, or needs_review. Discard disputed.

Integration (Phase 4)

Group	Steps
Clean	Deduplicate, filter by confidence
Correlate	Cross-domain grouping, convergence signal detection
Synthesize	Root cause synthesis across domains, 5 Whys on clusters
Prioritize	Score by findings resolved × severity × fixability, generate unified action plans

Clean

Step	Action
1	Deduplicate by file:line:category (keep highest severity)
2	Filter by finding.confidence: ≥95% include, 60-94% include with note, <60% exclude

Correlate

Step	Action
3	Group findings by location (file, module, boundary)
4	Identify convergence signals — where 2+ domains flag the same area
5	Single-domain findings with no correlation remain as standalone items

Synthesize

Step	Action
6	For each convergence cluster, synthesize one root cause that explains all findings
7	Apply 5 Whys on the synthesized root cause, not individual findings
8	Classify root cause by category (see Root Cause Categories)
9	Standalone findings: apply 5 Whys individually, classify as before

Prioritize

Step	Action
10	Score root causes: findings_resolved × max_severity × fixability
11	Generate unified action plans per root cause (one action resolves many findings)
12	Generate auto-fixable suggestions (target root cause where possible)

Root Cause Categories

Category	Indicators	Resolution
Architecture Gap	Pattern spans modules	Design change
Knowledge Gap	Inconsistent patterns	Documentation
Tooling Gap	Linter could catch	Config update
Process Gap	Slips through review	Process change

Auto-Fixable Detection

fix_type	Description	Action
auto	Known fix pattern, confidence ≥85%	Generate suggestion
manual	Requires human judgment	Skip suggestion

Output

Output final Markdown report (returned to leader via Task completion).

Section	Schema Source
summary	templates/audit/snapshot.yaml
root_causes	templates/audit/snapshot.yaml
priorities	templates/audit/snapshot.yaml
suggestions	Integrator-specific (schema below)

Exclude meta and pipeline_health — leader adds those.

## Suggestions

| Metric             | Value |
| ------------------ | ----- |
| auto_fixable_count | count |
| manual_count       | count |

### SUG-001

| Field          | Value                               |
| -------------- | ----------------------------------- |
| root_cause_ref | RC-001                              |
| category       | category                            |
| severity       | critical / high / medium / low      |
| fix_type       | auto / manual                       |
| confidence     | 0.85-1.00                           |
| location       | file path : line number             |
| effort         | 5min / 15min / 30min / 1h / manual  |
| Before         | original code snippet               |
| After          | suggested fix snippet               |
| Rationale      | why this fix — traces to root cause |

Synthesis Principles

Principle	Description
Root causes over symptoms	Same location = likely one shared cause
Cross-domain signals	2+ domains flagging same area = high-confidence issue
One action, many fixes	Best actions resolve multiple findings at once
Traceability	Every root cause traces to its source findings
Honest standalone	Not every finding has a cross-domain root cause
Evidence over opinion	Verified findings outweigh unverified in prioritization

Priority Score

For root causes:  findings_resolved × max_severity × fixability
For standalone:   Impact × Reach × Fixability

- max_severity: critical=10, high=5, medium=2, low=1
- fixability: 1 / effort (low=1, medium=2, high=3)

Score	Priority	Timing
> 50	Critical	Immediate
20-50	High	This sprint
5-20	Medium	Next sprint
< 5	Low	Backlog

Constraints

Rule	Description
Require challenger AND verifier	Don't integrate until both perspectives available
Reconcile before integrate	Apply reconciliation rules before dedup/correlate
Synthesize, don't list	Cross-domain findings must be correlated
Trace everything	Every root cause links to its source findings
Don't force correlation	Standalone findings are valid on their own

Error Handling

Error	Recovery	Output
Challenger missing	Proceed with verifier results only (Rule 6 applied)	Findings use verifier verdicts, no reconciliation
Verifier missing	Proceed with challenger results only (original behavior)	Findings use challenger verdicts unchanged
Both missing	Leader provides raw reviewer findings → start Phase 4	Raw reviewer findings, no reconciliation applied
No findings received	Return empty report with note	summary.total_findings: 0, note in report
Challenge read failure	Mark finding as needs_context	Individual finding flagged for review
All low confidence	Report "No high-confidence items"	Empty priorities, all findings listed as low