v8: ticket #000012 Phase 1a — ForkScore consumes the new bench substrate
Pure scoring function over (parent, child) BatteryResult bundles. Closes
the scoring half of fox's 2026-05-08 frontier note ("how does an organism
mutation become canonical?") — the canonicalization half (validator
state, acceptance, fork choice) stays under #000012 as the v8 paper.
Formula:
ForkScore = α·Δ5S + β·Δ5T + γ·Δ5F + δ·SelfModelCalibration
+ ε·AuditCompleteness + ζ·ValidatorDiversity
- η·RegressionPenalty - θ·CapitalCostPenalty
- ι·SecurityRisk - κ·Complexity - λ·MemoryInvalidation
Consumes every metric this session shipped:
- 5S/5T/5F sub-battery rates → Δ-rate per battery (mean over subs)
- adaptation_efficiency_mean_finite + adaptation_efficiency_infinite_count
→ 5F efficiency-aware bonus (damped + capped via INFINITE_BONUS_CAP)
- adaptation_efficiency_neg_infinite_count > parent → NEG_INF_REGRESSION
flag → automatic REJECT (free regression unsafe)
- capital_delta from #000020 ledger
- memory_invalidation_count from #000017
Verdict thresholds:
- ACCEPT when score >= SIGNAL_FLOOR (5pp; matches docs/bench-maxing.md)
- MARGINAL [0, SIGNAL_FLOOR)
- REJECT on negative score OR hard-regression OR neg-inf efficiency
Hard-regression flag fires if any single sub-battery rate drops by
>= 5pp parent→child, regardless of net score. CLI exits 1 on REJECT
so CI gates run `arborist v8 score` directly.
Surface:
- arborist/v8/{fork_score,weights}.py
- WeightSet dataclass with α…λ + DEFAULT_WEIGHTS (single-validator
tuned: ζ=0, ι=0, κ=0; η=2.0 weighted heavier than improvement
weights; θ=0.5 modest cost penalty)
- weights_from_dict accepts "lambda" key (Python reserved word)
- bench_result_to_metrics adapter from runner --all JSON
- CLI: arborist v8 score --parent P.json --child C.json [--weights W.json]
with override flags for SelfModelCalibrationGain, AuditCompleteness,
capital_delta, memory_invalidation_count, etc.
Reference: docs/v8-fork-score.md (formula + term semantics + verdict
matrix + Phase-1a vs Phase-1b boundary).
Tests: tests/test_v8_fork_score.py (25 cases)
- adapter from runner JSON
- all verdict paths (ACCEPT / MARGINAL / REJECT)
- hard-regression flag
- neg-inf efficiency rejection
- inf-bonus capping
- weight tuning (alpha scales 5s, eta scales penalty)
- breakdown completeness (11 terms) sums to score
- CLI smoke + explicit weights + REJECT exit code
- Determinism: same inputs → same output
Full suite: 1186 passed, 36 skipped.
#000012 status: in progress (Phase 1a landed, consensus paper still open).
russell@unturf.com
ea39455d75