e831ed2615
Phase 2a r2/r3/r4 of the redefined plan: the general comprehension reader now
reads THREE independent-gold reasoning domains end-to-end (prose -> MeaningGraph
-> projection -> independent oracle -> answer vs gold), all wrong=0, and all
three are wired into the capability index.
reader.py — new domain-agnostic templates (function words + order; parse-or-refuse):
- categorical E/I/O: "no Xs are Ys"->disjoint, "some Xs are Ys"->intersects,
"some Xs are not Ys"->some_not (A "all Xs are Ys"->subset already existed)
- "therefore <categorical>" -> conclusion QUERY (same neutral predicate vocab)
- comparative facts: "<X> [is] <comp> [than] <Y>" -> less(...), closed
less/greater comparator lexicon, elided-copula support
- sort query ("sort ascending|descending", "... order from <low> to <high>")
and compare query ("compare <X> with <Y>")
- clause-splitting on commas / leading and|or for multi-clause sentences
projectors.py — to_syllogism (premises + validity conclusion, finite-model size 3)
and to_total_ordering (less-facts + sort/compare). Both return None when nothing
is honestly askable of their oracle (caller treats as refusal).
capability_index — wire 3 comprehension lanes into ADAPTERS; re-freeze baseline
breadth 3->6, capability_score 0.919641->0.814356 (geomean falls BY DESIGN as
honest partial-coverage domains join; wrong_total stays 0). digest 0a98b9b4...
Scores: set_membership 8/8, syllogism 6/8, total_ordering 4/8 — all wrong=0.
Multi-word NP handling is DEFERRED on purpose, not missed: the gold lanes
canonicalize multi-word NPs three contradictory ways ("North station"->"north",
"Level one"->"level_one", "metal objects"->"metal"), so no single general rule is
wrong=0-safe. The reader refuses multi-word NPs until the gold lanes carry a
canonicalization contract. Every refusal is a genuine harder phenomenon
(multi-word NP, adjectival predicate, trailing tokens) — never a readable case
silently dropped.
Tests: reader templates, projector unit tests, syllogism/total_ordering
end-to-end wrong=0 with pinned counts, capability breadth 3->6. 138 targeted +
87 smoke green. Lane SHAs 8/9 (sole miss = public_demo env wall-clock flake).
118 lines
4.6 KiB
Python
118 lines
4.6 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 three
|
||
# comprehension lanes (prose -> MeaningGraph -> projection -> independent
|
||
# oracle) compose into the cross-domain index with zero wrong commits.
|
||
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 == 6
|
||
assert {d.domain for d in idx.domains} == {
|
||
"deductive_logic",
|
||
"dimensional",
|
||
"relational_metric",
|
||
"comprehension_set_membership",
|
||
"comprehension_syllogism",
|
||
"comprehension_total_ordering",
|
||
}
|
||
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
|