a4d68682df
- Promote tests/test_register_substantive_consumption.py into the smoke suite (local core/cli_test.py + mirrored .github/workflows/smoke.yml): the falsifiable register-axis (ADR-0069/0071/0077) contract was red on main outside every gate for the 2026-07-22..24 window, masked by an unrelated flake label. 36 tests, ~4.6s. - Found in passing: test_core_test_deductive_suite_expands_to_entailment_lane was already red on clean main (an exact-tuple pin from when the `deductive` suite had one file; it has since grown to 12). Fixed to a contains-style assertion so it can't go stale the same way again. - Two new capability-index adapters: deduction_serve_existential_result (Band v6-EX alone — the one band among six that changes the decision procedure itself) and curriculum_serve_result. Breadth 11 -> 13; baseline re-frozen (a deliberate re-freeze per its own docstring). - Promotion sweep: confirmed via the existing wrong=0 lane gate (run repeatedly this session across 166/166 deduction-serve cases) that no split beyond the already-applied ds-mem-0020 has a stale declined-gold case a current band now decides differently.
130 lines
5.3 KiB
Python
130 lines
5.3 KiB
Python
"""Cross-domain capability index — AGI-roadmap Phase 1 (MEASURE).
|
||
|
||
The yardstick that gates every later "more capable" claim. Two honest axes —
|
||
**accuracy** (of committed answers; wrong stays 0 in assert mode) and **coverage**
|
||
(attempted-not-refused) — aggregated across domains so it CANNOT be gamed by a
|
||
narrow per-domain win: the headline coverage is the GEOMETRIC MEAN across domains,
|
||
which only rises if *every* domain rises. A hack that maxes one lane and leaves
|
||
the rest at zero leaves the geomean ~0.
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
from evals.capability_index.index import (
|
||
DomainResult,
|
||
aggregate,
|
||
deterministic_digest,
|
||
)
|
||
|
||
|
||
def _d(domain: str, correct: int, wrong: int, refused: int) -> DomainResult:
|
||
return DomainResult(domain=domain, correct=correct, wrong=wrong, refused=refused)
|
||
|
||
|
||
def test_domain_result_axes() -> None:
|
||
r = _d("logic", correct=8, wrong=0, refused=2)
|
||
assert r.total == 10
|
||
assert r.attempted == 8
|
||
assert r.coverage == 0.8
|
||
assert r.accuracy == 1.0 # of committed answers
|
||
|
||
|
||
def test_aggregate_axes_micro() -> None:
|
||
idx = aggregate([_d("a", 6, 0, 4), _d("b", 2, 0, 8)])
|
||
assert idx.wrong_total == 0
|
||
assert idx.coverage == 0.4 # (6+2)/(10+10) micro
|
||
assert idx.accuracy == 1.0 # no wrong
|
||
assert idx.breadth == 2 # both domains have some coverage
|
||
|
||
|
||
def test_geomean_coverage_resists_narrow_gaming() -> None:
|
||
# A NARROW hack: one domain maxed, the rest at zero coverage.
|
||
narrow = aggregate(
|
||
[_d("gamed", 10, 0, 0), _d("x", 0, 0, 10), _d("y", 0, 0, 10)]
|
||
)
|
||
# A BALANCED engine: every domain partially covered.
|
||
balanced = aggregate(
|
||
[_d("gamed", 4, 0, 6), _d("x", 4, 0, 6), _d("y", 4, 0, 6)]
|
||
)
|
||
# Micro-coverage is similar (~0.33 vs 0.40), but the geomean exposes the hack:
|
||
assert narrow.coverage_geomean == 0.0 # any zero-coverage domain -> geomean 0
|
||
assert balanced.coverage_geomean > 0.39
|
||
# The capability score (geomean × accuracy) refuses to reward the narrow hack.
|
||
assert narrow.capability_score == 0.0
|
||
assert balanced.capability_score > 0.39
|
||
|
||
|
||
def test_balanced_progress_moves_the_score_monotonically() -> None:
|
||
low = aggregate([_d("a", 2, 0, 8), _d("b", 2, 0, 8)])
|
||
high = aggregate([_d("a", 6, 0, 4), _d("b", 6, 0, 4)])
|
||
assert high.coverage_geomean > low.coverage_geomean
|
||
assert high.capability_score > low.capability_score
|
||
|
||
|
||
def test_wrong_is_a_hard_gate() -> None:
|
||
# In assert mode wrong MUST be 0; any wrong invalidates the index (score 0)
|
||
# and is surfaced — never averaged away.
|
||
idx = aggregate([_d("a", 8, 1, 1), _d("b", 5, 0, 5)])
|
||
assert idx.wrong_total == 1
|
||
assert idx.assert_mode_valid is False
|
||
assert idx.capability_score == 0.0 # wrong=0 is non-negotiable in assert mode
|
||
|
||
|
||
def test_digest_is_deterministic_and_bites() -> None:
|
||
a = aggregate([_d("a", 6, 0, 4), _d("b", 2, 0, 8)])
|
||
b = aggregate([_d("a", 6, 0, 4), _d("b", 2, 0, 8)])
|
||
assert deterministic_digest(a) == deterministic_digest(b)
|
||
moved = aggregate([_d("a", 7, 0, 3), _d("b", 2, 0, 8)])
|
||
assert deterministic_digest(moved) != deterministic_digest(a)
|
||
|
||
|
||
def test_empty_index_is_well_defined() -> None:
|
||
idx = aggregate([])
|
||
assert idx.coverage == 0.0
|
||
assert idx.coverage_geomean == 0.0
|
||
assert idx.breadth == 0
|
||
assert idx.capability_score == 0.0
|
||
|
||
|
||
def test_real_lanes_compose_into_the_index_with_wrong_zero() -> None:
|
||
# The baseline: three structured-input reasoning lanes PLUS the comprehension lanes
|
||
# (prose -> MeaningGraph -> projection -> independent oracle / pack-named predicate)
|
||
# compose into the cross-domain index with zero wrong commits. The one-hop
|
||
# relational_inference lane (#775) is the 10th domain; the transitive
|
||
# relational_transitive lane (B2) is the 11th. deduction_serve_existential
|
||
# (ADR-0261, Band v6-EX) is the 12th; curriculum_serve (ADR-0262, generalization
|
||
# arc Tier S) is the 13th. (This assertion was stale at 9 from the #596 era —
|
||
# #775's 10th domain was never added here; corrected with B2, then again here.)
|
||
from evals.capability_index.adapters import collect_domain_results
|
||
|
||
collection = collect_domain_results()
|
||
assert collection.not_covered == () # every adapter ran (no silent drop)
|
||
idx = aggregate(list(collection.results))
|
||
assert idx.wrong_total == 0
|
||
assert idx.assert_mode_valid
|
||
assert idx.breadth == 13
|
||
assert {d.domain for d in idx.domains} == {
|
||
"deductive_logic",
|
||
"dimensional",
|
||
"relational_metric",
|
||
"comprehension_set_membership",
|
||
"comprehension_syllogism",
|
||
"comprehension_total_ordering",
|
||
"comprehension_propositional",
|
||
"comprehension_relational_metric",
|
||
"comprehension_relational_predicate",
|
||
"comprehension_relational_inference",
|
||
"comprehension_relational_transitive",
|
||
"deduction_serve_existential",
|
||
"curriculum_serve",
|
||
}
|
||
assert idx.capability_score > 0.5 # real, non-trivial cross-domain capability
|
||
|
||
|
||
def test_index_report_is_deterministic_across_runs() -> None:
|
||
# The capability number is reproducible — improvement is a replayable curve.
|
||
from evals.capability_index.adapters import collect_domain_results
|
||
|
||
a = deterministic_digest(aggregate(list(collect_domain_results().results)))
|
||
b = deterministic_digest(aggregate(list(collect_domain_results().results)))
|
||
assert a == b
|