Files
hack-house/tests/test_sor_confirm.py
T
leetcrypt cb4b6869fc R7/analysis: RQ1/RQ2 confirmatory decision layer (frozen §6 gates in code)
Encodes the four lead-paper confirmatory tests exactly per the frozen prereg §6,
calibrated on synthetic ground truth only (never re-fit to cell data):

- RQ1-P1 leak: correlation AUC + BCa CI; gate = CI excludes 0.5 (D2); 0.60 CI
  lower bound = separate materiality label (material / weak-but-real), not the gate.
- RQ1-P2 padding: paired ΔAUC = AUC(no-pad) − AUC(pad) over circuits; effective
  iff CI > 0.
- RQ2-P1 anonymity set: ΔH = H(federated) − H(single, matched N) with Miller-Madow
  per-circuit entropy; two-sided grow / honest-shrink / inconclusive by CI sign.
- RQ2-P3 mechanism: Spearman ρ(top-3 bridge concentration, per-circuit H) + CI.

apply_holm corrects the reported RQ1/RQ2 subset against the full frozen family of
7 (family_size default 7) — never re-optimised to the 4 reported. Bootstrap
p-values order the Holm step-down only; every decision is a CI gate, never a bare p.

stats.py: bootstrap CIs can now return their resample distribution so the Holm
ordering p-value comes from the same resamples as the CI.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-19 19:36:17 -07:00

150 lines
5.8 KiB
Python

"""RQ1/RQ2 confirmatory gates — validated on synthetic ground truth only.
Each test constructs a distribution whose §6 verdict is known by construction
(a clear leak / a null; padding that works / doesn't; federation that grows /
shrinks / is flat; a funnelling mechanism) and asserts the frozen gate fires the
right way. No confirmatory-cell data is involved. The Holm test pins that the
reported RQ1/RQ2 subset is corrected against the full frozen family of 7.
"""
from cmd_chat.sor.analysis.confirm import (
FROZEN_FAMILY_SIZE,
PairedCircuit,
apply_holm,
rq1_p1_leak,
rq1_p2_padding,
rq2_p1_delta_h,
rq2_p3_funnel,
)
class _LCG:
def __init__(self, seed=0x2468ACE0):
self.s = seed
def u(self):
self.s = (1103515245 * self.s + 12345) & 0x7FFFFFFF
return self.s / 0x7FFFFFFF
def _pairs(hi_mean, lo_mean, n=80, rng=None):
"""n linked pairs near hi_mean + n unlinked near lo_mean, with jitter."""
rng = rng or _LCG()
out = []
for _ in range(n):
out.append((hi_mean + 0.1 * rng.u(), True))
out.append((lo_mean + 0.1 * rng.u(), False))
return out
# --------------------------------------------------------------------------- #
# RQ1-P1 — leak gate + materiality label.
# --------------------------------------------------------------------------- #
def test_rq1p1_material_leak():
t = rq1_p1_leak(_pairs(0.9, 0.1), seed=1, n_resamples=2000)
assert t.decision == "leak"
assert t.label == "material" # CI lower bound >= 0.60
assert t.ci.excludes(0.5)
def test_rq1p1_null_when_scores_overlap():
# Linked and unlinked drawn from the same band -> AUC ~ 0.5, CI spans it.
t = rq1_p1_leak(_pairs(0.5, 0.5), seed=1, n_resamples=2000)
assert t.decision == "null"
assert not t.ci.excludes(0.5)
# --------------------------------------------------------------------------- #
# RQ1-P2 — padding efficacy (paired ΔAUC).
# --------------------------------------------------------------------------- #
def test_rq1p2_padding_effective_when_pad_lowers_auc():
rng = _LCG(0x1111)
circuits = []
for _ in range(30):
nopad = tuple(_pairs(0.9, 0.1, n=4, rng=rng)) # strong linkage
pad = tuple(_pairs(0.5, 0.5, n=4, rng=rng)) # padding blurs it (AUC~0.5)
circuits.append(PairedCircuit(nopad, pad))
t = rq1_p2_padding(circuits, seed=2, n_resamples=2000)
assert t.decision == "padding-effective"
assert t.ci.strictly_greater(0.0)
def test_rq1p2_padding_ineffective_when_no_change():
rng = _LCG(0x2222)
circuits = []
for _ in range(30):
nopad = tuple(_pairs(0.7, 0.3, n=4, rng=rng))
pad = tuple(_pairs(0.7, 0.3, n=4, rng=rng)) # identical regime
circuits.append(PairedCircuit(nopad, pad))
t = rq1_p2_padding(circuits, seed=2, n_resamples=2000)
assert t.decision == "padding-ineffective"
assert not t.ci.strictly_greater(0.0)
# --------------------------------------------------------------------------- #
# RQ2-P1 — ΔH two-sided (grow / shrink / inconclusive).
# --------------------------------------------------------------------------- #
def _uniform_circuits(n_circuits, n_senders, per=50):
return [[per] * n_senders for _ in range(n_circuits)]
def _skewed_circuits(n_circuits, n_senders):
# One dominant sender -> low entropy.
return [[1000] + [1] * (n_senders - 1) for _ in range(n_circuits)]
def test_rq2p1_grow_when_federation_is_more_uniform():
fed = _uniform_circuits(30, 8) # high H (~3 bits)
single = _skewed_circuits(30, 8) # low H
t = rq2_p1_delta_h(fed, single, seed=3, n_resamples=2000)
assert t.decision == "grow"
assert t.ci.strictly_greater(0.0)
def test_rq2p1_honest_shrink_reported():
fed = _skewed_circuits(30, 8) # federation funnels -> low H
single = _uniform_circuits(30, 8) # matched-N single house, high H
t = rq2_p1_delta_h(fed, single, seed=3, n_resamples=2000)
assert t.decision == "shrink"
assert t.ci.strictly_less(0.0)
def test_rq2p1_inconclusive_when_arms_match():
fed = _uniform_circuits(30, 8)
single = _uniform_circuits(30, 8)
t = rq2_p1_delta_h(fed, single, seed=3, n_resamples=2000)
assert t.decision == "inconclusive"
assert not t.ci.excludes(0.0)
# --------------------------------------------------------------------------- #
# RQ2-P3 — funnelling mechanism (Spearman).
# --------------------------------------------------------------------------- #
def test_rq2p3_funnel_negative_rho():
# Higher top-3 concentration -> lower per-circuit entropy.
conc = [i / 20.0 for i in range(20)]
h = [3.0 - c for c in conc]
t = rq2_p3_funnel(conc, h, seed=4, n_resamples=2000)
assert t.decision == "funnel"
assert t.ci.strictly_less(0.0)
# --------------------------------------------------------------------------- #
# Holm over the full frozen family (size 7) while reporting 4.
# --------------------------------------------------------------------------- #
def test_apply_holm_corrects_against_family_of_seven():
t1 = rq1_p1_leak(_pairs(0.95, 0.05), seed=1, n_resamples=1500) # tiny p
t2 = rq1_p2_padding(
[PairedCircuit(tuple(_pairs(0.9, 0.1, n=4)), tuple(_pairs(0.5, 0.5, n=4)))
for _ in range(30)], seed=2, n_resamples=1500)
fed, single = _uniform_circuits(30, 8), _skewed_circuits(30, 8)
t3 = rq2_p1_delta_h(fed, single, seed=3, n_resamples=1500)
t4 = rq2_p3_funnel([i / 20.0 for i in range(20)],
[3.0 - i / 20.0 for i in range(20)], seed=4, n_resamples=1500)
holm = apply_holm([t1, t2, t3, t4])
assert len(holm) == 4
# Smallest-p test gets the full-family multiplier of 7, not 4.
top = min(holm, key=lambda h: h.rank)
assert top.multiplier == FROZEN_FAMILY_SIZE == 7
mults = sorted(h.multiplier for h in holm)
assert mults == [4, 5, 6, 7]