GESA Framework

Appearance

GESA × Fetch ​

Learning Which Decisions Produce Good Outcomes ​

Every Fetch decision is a GESA episode input. Over time, GESA learns which Fetch thresholds in which contexts produce good outcomes — and can recommend calibration adjustments.

Fetch Decisions as Episodes ​

Each Fetch decision maps to a specific episode type:

Fetch → Execute  →  Episode stored as "action taken"
Fetch → Confirm  →  Episode stored as "action deferred pending validation"
Fetch → Queue    →  Episode stored as "action queued, conditions not met"
Fetch → Wait     →  Episode stored as "no action, insufficient signal"

All four are valuable:

  • Execute episodes show what happened when the system acted with high confidence
  • Confirm and Queue episodes show what happened when the system was cautious
  • Wait episodes show what happened when no action was taken

The comparison between Execute and Wait episodes for similar contexts reveals whether high-confidence action actually produces better outcomes than waiting — which is not guaranteed.

Threshold Calibration ​

The Fetch formula produces a score on a continuous scale. The thresholds (1000/500/100) are defaults. GESA can recommend domain-specific threshold adjustments based on episode history.

Fetch Threshold Recommendation = f(domain, historical_outcome_by_threshold)

Example: In a specific trading domain, GESA may find that Execute-threshold actions (>1000) produce positive outcomes only 55% of the time, while Confirm-threshold actions (500–1000) that proceed after confirmation produce positive outcomes 78% of the time. This would suggest raising the auto-execute threshold in this domain.

GESA does not modify the Fetch formula. It recommends calibration of the domain-specific thresholds — a human or system operator makes the final threshold change.

Confidence Learning ​

GESA tracks whether the Fetch confidence score is well-calibrated for the domain:

Calibration(domain) = correlation(fetchScore, outcomeSuccess)
  • Perfect calibration: high fetch scores reliably predict success; low fetch scores reliably predict failure
  • Overconfident: high fetch scores frequently lead to poor outcomes
  • Underconfident: low fetch scores frequently pass up successful actions

Calibration feedback is surfaced as a GESARecommendation with type: 'threshold_adjustment'.

The Dependency Preserved ​

The architectural hierarchy is maintained in both directions:

Fetch depends on DRIFT (cannot compute without a gap)
GESA depends on Fetch (cannot learn without decisions)
Fetch does not depend on GESA (operates purely on current state)

GESA enriches the Fetch layer's decision history. It does not change how Fetch computes its score. The layers remain independently valid.

Episode Density and Fetch Learning ​

Learning threshold calibration requires sufficient episode density per context:

  • Minimum 10 Execute episodes in a domain before calibration recommendations are reliable
  • Minimum 5 Wait/Queue episodes for absence-of-action calibration
  • Compare across similar DRIFT magnitudes (a Fetch=1200 with DRIFT=10 is different from Fetch=1200 with DRIFT=80)

→ GESA × 6D Foraging