Metrics

February 27, 2026 · View on GitHub

Quantitative results from the D&D 3.5e referee engine project — the proving ground for this framework.

Last Updated: 2026-02-27


Verification Results

DomainFormulasCorrectWrongAmbiguousUncited
A: Attack Resolution57434→22→49
B: Combat Maneuvers4534551
C: Spells & Saves21153→11→32
D: Conditions3828442
E: Movement & Terrain4236321
F: Character Progression77655→33→54
G: Play Loop2216222
H: Targeting1816110
I: Geometry & Feats4938344
Total33825530→~222825

Arrows indicate changes from research cross-referencing (Domain A re-verified, others estimated).

First-pass accuracy: 75.4% (CORRECT / total) With AMBIGUOUS as non-bugs: 83.7% ((CORRECT + AMBIGUOUS) / total) Estimated genuine bugs after cross-ref: ~22 of 338 = 93.5% accuracy


Bug Classification

30 bugs categorized into 8 error patterns:

PatternCountExample
Missing modifier/multiplier6STR grip multiplier not applied
Wrong threshold/floor4min damage 0 instead of 1
Condition not differentiated4Prone AC flat instead of melee/ranged
Missing field/parameter3Concentration DC missing spell level
Inverted condition2Soft cover applied to melee instead of ranged
Incorrect die/formula3Water fall d6 instead of d3, sunder hardcoded 1d8
Incomplete enumeration2SIZE_ORDER missing 3 categories, Colossal footprint
Design decision misidentified6Cover values flagged as bugs, were intentional

Agent Coordination Metrics

MetricValue
Total agent sessions100+
Parallel agent groups (max)7 simultaneous
Silent agent failure rate3/7 (43%) in one parallel dispatch
WOs requiring reclassification1 of 13 (WO-FIX-11, code structure mismatch)
Schema cascade underestimation1 of 13 (WO-FIX-03, 3 files scoped → 6 touched)
Cross-file consistency failures2 (Domain C verification, Domain A checklist)
Research cross-ref reclassifications4 confirmed (Domain A), ~8-10 estimated (all domains)
Total WOs dispatched (all types)100+ (fix, feature, research, governance, audit, framework)
Builder debriefs archived50+
Research documents produced30
Delivery batches completed25+
Gate test suite size8,521+ tests

Fix Execution Metrics

MetricValue
Fix WOs dispatched13 (12 active, 1 retired) — Phase 1 only
Fix WOs completed11 of 12
Fix WOs needing reclassification1 (WO-FIX-11)
Fix WOs partially completed1 (WO-FIX-12, BUG-F2/F3 unverified)
Tests passing after all Phase 1 fixes5,277
Tests passing after all Phase 2 batches8,521+
Tests updated (old wrong behavior)6
Gold master files regenerated4
Total commits for fix session9

H1 WO Batch Metrics

MetricValue
H1 WOs completed7
Tests passing after H1 batch5,804+
Builder commit failures recovered (one batch)4 (3/7 agents silently failed to commit)
Integration Constraint PolicyCodified — no new infrastructure WOs until canary runs
Integration break points found by canary4 (all invisible to unit tests)

Phase 2 Batch Delivery Metrics

MetricValue
Batches completed (Phase 2)20+ (Batches I through R+)
Average gate tests per batch8 (range: 6-11)
Average accepted WOs per batch4
Regression failures introduced0 (zero regressions across all accepted batches)
Ghost WOs dispatched (feature already implemented)~3 (identified via pre-dispatch verification)
Parallel path drift incidents caught1 (F-011, 21-modifier divergence, discovered by sweep audit)
Regression spiral incidents1 (F-ML-004, agent burned context on pre-existing failures)
PM commits sweeping staged builder code1 (F-ML-005, caught post-hoc; audit trail corrected)
Spec authority gap incidents1 (F-012, community variant shipped over specification value)

Enforcement Tier Effectiveness

TierStickinessExample
Tier 1: Test-enforced~100%Boundary law tests, WorldState immutability
Tier 2: Process-enforced70-80%Dispatch templates, handoff checklists
Tier 3: Prose-enforced40-60%PM inbox 10-item cap (violated 2.3x)

Context Window Observations

ObservationData Point
Largest agent token consumption~170K tokens (WO-FIX-01/02 attack resolvers)
Typical WO agent consumption30K-80K tokens
PM summary effective length7 items / ~500 tokens
Full debrief length~2,000-4,000 tokens
Compression ratio (full → summary)4:1 to 8:1