/gauge

Self-Evolution Subsystem (EVOLVE)

Purpose: Self-modification workflow, triggers, safety levels, and rollback rules for Gauge. Read when: Gauge is evaluating or performing self-evolution. Pattern source: architect/reference/self-evolution.md

Overview

EVOLVE handles Gauge's internal improvement loop — refining detection patterns, checklist criteria, and web sources. Darwin handles ecosystem-wide evolution. They complement — never overlap.

EVOLVE Workflow

RESEARCH → EVALUATE → CLASSIFY → UPDATE → VERIFY → PERSIST

Phase Details

RESEARCH collects:

• Audit results from recent SCAN/CLASSIFY cycles (false positives, false negatives)
• User corrections to Gauge's recommendations
• Web research findings (using queries from web-sources.md)
• Lore pattern insights (if received)
• Detection pattern performance data

EVALUATE scores:

• Relevance: Does this finding address a known detection gap?
• Source quality: What tier(s) support this finding?
• Impact: How many skills would be affected?
• Risk: Could this change cause false positives/negatives?

CLASSIFY assigns safety levels (see Safety Framework below).

UPDATE executes bottom-up:

1. Take snapshot of affected files
2. Update reference/ files first (Level A/B)
3. Only propose SKILL.md changes at Level C (human approval)
4. Log all changes with before/after

VERIFY validates:

• Self-audit: Gauge's own SKILL.md still passes all 19 items
• Detection consistency: No previously PASS items now FAIL on known-good skills
• Reference integrity: All referenced files exist
• On failure: full rollback to snapshot

PERSIST records:

• Journal entry in .agents/gauge.md
• Evolution signal to Darwin (ecosystem awareness)
• Update cumulative evolution log

Self-Evolution Triggers

Scope Definitions

• Lightweight: RESEARCH + EVALUATE only. No modifications. Record observations in journal.
• Medium: RESEARCH → EVALUATE → CLASSIFY. Generate proposal but only execute Level A changes automatically.
• Full cycle: All 6 phases. Execute Level A/B changes, propose Level C for human approval.

Safety Framework

Change Safety Levels

Runaway Prevention Mechanisms

Budget Tracking

Session budget:  3 changes (reference/ or detection patterns)
Monthly budget:  10 changes total
Reset:           Session = per conversation | Monthly = calendar month
Tracking:        Record in evolution log (date, change description, budget remaining)

Templates

Evolution Proposal

## GAUGE_EVOLUTION_PROPOSAL
- Date: YYYY-MM-DD
- Trigger: [GT-XX]
- Research summary: [key findings]
- Proposed changes:
  | # | Target file | Change type | Safety level | Description |
  |---|-------------|-------------|--------------|-------------|
  | 1 | reference/xxx.md | Add/Update | A/B | [description] |
- Source support:
  | Source | Tier | Relevance |
  |--------|------|-----------|
  | [URL] | T{1-4} | [why relevant] |
- Budget check:
  - Session remaining: [X] changes
  - Monthly remaining: [Y] changes
- Risk assessment: [low/medium/high]

Verification Report

## GAUGE_VERIFY_REPORT
- Date: YYYY-MM-DD
- Self-audit: [PASS/FAIL] (own SKILL.md, 19/19 items)
- Detection consistency: [PASS/FAIL] — [notes]
- Reference integrity: [all files exist: YES/NO]
- Result: [ACCEPT/ROLLBACK]
- Rollback executed: [YES/NO/N/A]

Evolution Log Entry

## GAUGE_EVOLUTION_LOG
- Date: YYYY-MM-DD
- Trigger: [GT-XX]
- Scope: [Lightweight/Medium/Full]
- Changes applied:
  - [file]: [summary of change]
- Budget consumed: [session: X / monthly: Y]
- Budget remaining: [session: X / monthly: Y]
- Next scheduled: [date or condition]

Integration Points

Decision Flowchart

Audit task completed?
  └→ Run RESEARCH + EVALUATE (GT-01, Lightweight)
       └→ Issues found?
            ├→ No: Record in journal, done
            └→ Yes: Check trigger table
                 └→ Matching trigger?
                      ├→ Lightweight: Record observations only
                      ├→ Medium: CLASSIFY, execute Level A only
                      └→ Full: CLASSIFY → UPDATE → VERIFY → PERSIST
                           └→ Safety level of changes?
                                ├→ A: Execute immediately
                                ├→ B: Execute + mandatory VERIFY
                                ├→ C: Propose to human, await approval
                                └→ D: REJECT, log attempt as violation