fittingcorpus / results /dn1_design_nulls.json
Tarush-AI's picture AnonymousInterpScience's picture
Duplicate from AnonymousInterpScience/pythia-jlens-artifacts
3b151ef
Raw History Blame Contribute Delete
15.5 kB
{
"experiment": "DN1 \u2014 null distributions of three already-registered bars",
"prereg": "docs/experiments/preregs/DN1_design_nulls.md",
"status": "PRE-REGISTERED",
"measures_nothing_new": true,
"scope": "Calibrates thresholds, not measurements. Cannot overturn CV6, CV7 or R9; prices the bars they were judged against.",
"n_sim": 1000000,
"n_perm": 100000,
"A1_design_null": {
"definition": "R = range(group means) / sqrt(mean(per-group sample var)), G=5 groups of n=3",
"distribution": "iid normal, scale-free in sigma and location-free in mu",
"mean": 1.4558214009529349,
"sd": 0.6683661758236041,
"percentiles": {
"p50": 1.3504426642962697,
"p90": 2.3197482513624257,
"p95": 2.685370632595311,
"p99": 3.539988075003452,
"p99.9": 4.87893156220862
},
"P_R_ge_10": 2e-06,
"P_R_ge_10_exceedances_observed": 2,
"note": "A one-sided exceedance of 0 in n_sim draws is reported as < 1/n_sim, not as 0. The bar is calibrated against this column."
},
"C1_scale_freedom": {
"required": "|p99(sigma=1) - p99(sigma=100)| <= 0.02",
"p99_sigma_1": 3.539988075003452,
"p99_sigma_100": 3.5461058294242056,
"abs_diff": 0.006117754420753574,
"fires": true,
"falsification_witness": {
"what": "same code with the denominator replaced by the constant 1.0",
"p99_sigma_1": 2.6567914459131163,
"p99_sigma_100": 265.1280524472683,
"abs_diff": 262.4712610013552,
"control_would_fire": false,
"note": "this must be False, or C1 cannot distinguish a scale-free statistic from any other"
}
},
"A2_permutation_null_410m": {
"source": "results/d3_corpus_by_family_410m.json -> per_arm.<corpus>|s<seed>.<family>.z",
"unit": "15 z values per family, permuted across 5 corpus labels, 3 per label",
"by_family": {
"multihop": {
"observed_R": 25.361005428912275,
"null_mean": 1.4354219513761668,
"null_sd": 0.8281876972206043,
"null_p50": 1.277465107028326,
"null_p95": 2.515198551887543,
"null_p99": 5.61098563909792,
"null_max": 25.43264217979271,
"n_null_ge_bar": 23,
"P_null_ge_bar": 0.00023,
"n_null_ge_observed": 2,
"p_observed_one_sided": 2.999970000299997e-05
},
"multilingual": {
"observed_R": 19.426787847831193,
"null_mean": 1.4472181163211184,
"null_sd": 0.7043406387033073,
"null_p50": 1.3324022652684364,
"null_p95": 2.660570482190145,
"null_p99": 3.8541903041621643,
"null_max": 15.066763195787532,
"n_null_ge_bar": 10,
"P_null_ge_bar": 0.0001,
"n_null_ge_observed": 0,
"p_observed_one_sided": 9.99990000099999e-06
},
"order-ops": {
"observed_R": 24.632511354365725,
"null_mean": 1.4600481090175972,
"null_sd": 0.724071411525985,
"null_p50": 1.341316486899336,
"null_p95": 2.7854524342316607,
"null_p99": 3.4035854960502476,
"null_max": 24.632511354365725,
"n_null_ge_bar": 10,
"P_null_ge_bar": 0.0001,
"n_null_ge_observed": 1,
"p_observed_one_sided": 1.999980000199998e-05
},
"poetry": {
"observed_R": 42.686224720210525,
"null_mean": 1.4207336216576643,
"null_sd": 1.1226321014484546,
"null_p50": 1.0953311671022932,
"null_p95": 2.289972367428842,
"null_p99": 8.799193358965976,
"null_max": 42.686224720210525,
"n_null_ge_bar": 571,
"P_null_ge_bar": 0.00571,
"n_null_ge_observed": 1,
"p_observed_one_sided": 1.999980000199998e-05
},
"typo": {
"observed_R": 27.014210896207846,
"null_mean": 1.4510950885052967,
"null_sd": 0.7238398592296635,
"null_p50": 1.3394268870958137,
"null_p95": 2.644590252011299,
"null_p99": 4.196756766414129,
"null_max": 11.924406661562124,
"n_null_ge_bar": 11,
"P_null_ge_bar": 0.00011,
"n_null_ge_observed": 0,
"p_observed_one_sided": 9.99990000099999e-06
}
}
},
"C2_pooled_sd_recomputes": {
"required": "<= 1e-12 on all 5 families",
"by_family": {
"multihop": {
"recomputed": 0.017453481001782855,
"stored": 0.017453481001782855,
"abs_diff": 0.0
},
"multilingual": {
"recomputed": 0.019695881058294912,
"stored": 0.019695881058294912,
"abs_diff": 0.0
},
"order-ops": {
"recomputed": 0.012061420042855183,
"stored": 0.012061420042855183,
"abs_diff": 0.0
},
"poetry": {
"recomputed": 0.005578110253433368,
"stored": 0.005578110253433368,
"abs_diff": 0.0
},
"typo": {
"recomputed": 0.019451275697770446,
"stored": 0.019451275697770446,
"abs_diff": 0.0
}
},
"max_abs_diff": 0.0,
"fires": true
},
"C3_R_recomputes": {
"required": "<= 1e-12 on all 5 families",
"by_family": {
"multihop": {
"recomputed": 25.361005428912275,
"stored": 25.36100542891225,
"abs_diff": 2.4868995751603507e-14
},
"multilingual": {
"recomputed": 19.426787847831193,
"stored": 19.426787847831193,
"abs_diff": 0.0
},
"order-ops": {
"recomputed": 24.632511354365725,
"stored": 24.632511354365725,
"abs_diff": 0.0
},
"poetry": {
"recomputed": 42.686224720210525,
"stored": 42.686224720210525,
"abs_diff": 0.0
},
"typo": {
"recomputed": 27.014210896207846,
"stored": 27.014210896207846,
"abs_diff": 0.0
}
},
"max_abs_diff": 2.4868995751603507e-14,
"fires": true
},
"C4_identity_in_support": {
"required": "the unpermuted assignment reproduces the observed R exactly",
"by_family": {
"multihop": {
"identity_R": 25.361005428912275,
"observed_R": 25.361005428912275,
"abs_diff": 0.0
},
"multilingual": {
"identity_R": 19.426787847831193,
"observed_R": 19.426787847831193,
"abs_diff": 0.0
},
"order-ops": {
"identity_R": 24.632511354365725,
"observed_R": 24.632511354365725,
"abs_diff": 0.0
},
"poetry": {
"identity_R": 42.686224720210525,
"observed_R": 42.686224720210525,
"abs_diff": 0.0
},
"typo": {
"identity_R": 27.014210896207846,
"observed_R": 27.014210896207846,
"abs_diff": 0.0
}
},
"max_abs_diff": 0.0,
"fires": true
},
"C5_design_vs_permutation": {
"required": "the two 99th percentiles agree to within a factor of 1.25",
"p99_design": 3.539988075003452,
"p99_permutation_worst_family": 8.799193358965976,
"ratio": 2.485656214804451,
"fires": false
},
"B_tau_null_exact": {
"n": 5,
"n_orderings": 120,
"bar": 0.6,
"n_orderings_clearing_bar": 14,
"P_single_family_tau_ge_bar": 0.11666666666666667,
"distinct_tau_values": [
-1.0,
-0.8,
-0.6,
-0.4,
-0.2,
0.0,
0.2,
0.4,
0.6,
0.8,
1.0
],
"P_k_of_5_families_clear_under_independence": {
"0": 0.5378028459362139,
"1": 0.3551528227880658,
"2": 0.09381395318930041,
"3": 0.012390522119341565,
"4": 0.0008182420267489713,
"5": 2.1613940329218108e-05
},
"P_at_least_3_of_5_under_independence": 0.013230378086419754,
"DEPENDENCE_CAVEAT": "the five families share the same five operators and one activation cache, so they are not independent; the independence figure is an UPPER bound on the evidence, not an estimate of it"
},
"C_r9_count_null_exact": {
"design": "8 corpora, 15 own-derangement draws each",
"P_one_corpus_clears_all_draws_under_H0": 0.0625,
"expected_count_under_H0": 0.5,
"P_count_equals_zero_under_H0": 0.5967194738332182,
"observed_count": 0,
"reading": "a count of 0 is the modal outcome under the exact null, so the count alone carries almost no information; the direction does",
"sign_test_on_direction": {
"statistic": "number of corpora whose real z_vs_null is negative, min aggregation",
"n_negative": 7,
"n_total": 8,
"z_by_corpus": {
"Wikipedia_en": -1.6709216685460513,
"USPTO_Backgrounds": -0.7050665067014444,
"Pile-CC": -1.9874216306081265,
"StackExchange": -0.09090121377251528,
"Github": 0.593144297659444,
"OOD_News_2024": -0.7490796022292463,
"OOD_arXiv_2023": -0.0881460749226426,
"OOD_CommonPile": -0.7064942600047919
},
"p_one_sided_exact_binomial": 0.03515625,
"DEPENDENCE_CAVEAT": "the eight corpora share one activation cache and one 551-item battery; the sign test assumes independence and therefore understates the p-value"
},
"positive_control": {
"what": "persist on the same draws returns 8 of 8",
"n_beating": 8,
"why_it_matters": "the design CAN return the positive branch, so min's 0 of 8 is a failure of the statistic and not a lack of power in the null"
}
},
"controls_fired": {
"C1_scale_freedom": true,
"C2_pooled_sd_recomputes": true,
"C3_R_recomputes": true,
"C4_identity_in_support": true,
"C5_design_vs_permutation": false
},
"VERDICT": "UNCLEAR \u2014 a control did not fire: C5_design_vs_permutation",
"provenance": {
"schema_version": 1,
"experiment": "DN1",
"utc": "2026-08-31T23:19:44+00:00",
"script": {
"path": "experiments/dn1_design_nulls.py",
"sha256": "06e93d9dbd43feadaf2975d5854fd7252ba55264109a66909b2f4df18ea7f6e6",
"n_lines": 305
},
"git": {
"commit": "9987d82887ed339bf69b5e7fa1e4fc98d5bbb4ab",
"commit_short": "9987d82887ed",
"git_repository_present": true,
"branch": "pythia",
"dirty": true,
"dirty_files": [
"CONTEXT.md",
"NEELNANDA_REP.md",
"docs/experiments/COVERAGE.md",
"docs/experiments/METRICS.md",
"docs/experiments/descriptions/D1.md",
"docs/experiments/descriptions/E33.md",
"docs/experiments/descriptions/E34A.md",
"docs/experiments/descriptions/E34B.md",
"docs/experiments/descriptions/E36.md",
"docs/experiments/descriptions/E37.md",
"docs/experiments/descriptions/E44.md",
"docs/experiments/descriptions/E58.md",
"docs/experiments/preregs/E66d_s_quantization_limited.md",
"docs/experiments/preregs/superseded/PREREG_PYTHIA_T7_v2.md",
"docs/reproducibility/ARTIFACTS.md",
"docs/reproducibility/PAYLOAD_REDACTIONS.md",
"docs/reproducibility/SCRIPT_PROVENANCE.md",
"docs/reproducibility/review-snapshot-2026-08-19/MANIFEST.sha256",
"docs/reproducibility/review-snapshot-2026-08-19/PROVENANCE.md",
"docs/reproducibility/review-snapshot-2026-08-19/REDACTIONS.md",
"docs/reproducibility/review-snapshot-2026-08-19/review/EXTERNAL_REVIEW.md",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/AGENT_REPORT.md",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/CONTEXT_UNDER_REVIEW.md",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/FINDINGS.md",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/JUSTIFIABILITY.md",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/LEDGER.md",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/PHASE_A_ACCOUNT.md",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/SUSPICIONS.md",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Github_s0.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Github_s1.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Github_s2.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Pile-CC_s0.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Pile-CC_s1.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Pile-CC_s2.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_StackExchange_s0.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_StackExchange_s1.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_StackExchange_s2.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_USPTO_Backgrounds_s0.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_USPTO_Backgrounds_s1.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_USPTO_Backgrounds_s2.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Wikipedia_en_s0.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Wikipedia_en_s1.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Wikipedia_en_s2.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/balance_test.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/check_payload.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/coverage2.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/derange.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/derangement_result.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/logit_base.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/logit_baseline.json",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/my_anova.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/my_e52.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/my_rescore.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/prereg2.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/prereg_audit.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/rescore_all.py",
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/rstrip_anova.py",
"docs/validity/AUDIT_READINESS.md",
"docs/validity/audit-2026-08-22/DIGEST.json",
"experiments/d1_min_union_diagnostic.py",
"experiments/e28_eval_410m.sh",
"experiments/e28_ladder_410m.py",
"experiments/exploratory/descriptions/T0.md",
"experiments/exploratory/descriptions/T12.md"
],
"dirty_diff_sha256": "aff7ed44a7f13d5d00e837c12a7446366431d5a5ef0602ddfcfb051e2a0ddd20",
"describe": "9987d82-dirty"
},
"argv": [
"experiments/dn1_design_nulls.py"
],
"cwd": ".",
"inputs": [
{
"path": "results/d3_corpus_by_family_410m.json",
"sha256": "795676289554af8111a4b4464617203271848f210fe1ce29a9946d44ab5dad5a",
"bytes": 23102,
"exists": true
},
{
"path": "results/r9_permutation_calibrated_min.json",
"sha256": "773e2a6423db3022d96f71373c859fd0ae313f529abf569c396fc7de56461b89",
"bytes": 22358,
"exists": true
}
],
"env": {
"python": "3.14.2",
"torch": "2.13.0",
"transformers": "5.14.1",
"numpy": "2.5.1",
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
"machine": "arm64",
"tf32": {
"allow_tf32_matmul": false,
"allow_tf32_cudnn": true,
"NVIDIA_TF32_OVERRIDE": null
},
"threads": {
"torch_num_threads": 10,
"torch_num_interop_threads": 14,
"OMP_NUM_THREADS": null,
"os_cpu_count": 14
}
},
"payload_sha256": "201a8d6db3dfd8bcb12d81b23aa253b0ee6e7795a039ddb802b39ef668518dfe"
}
}