To Grok Grokking: Provable Grokking in Ridge Regression
Theory paper on provable grokking in ridge regression. Hand-driven reproduction that passed the ICML 2026 Agent Reproductions judge end to end — now the pipeline's regression fixture.
6supported
0falsified
0inconclusive
0not audited
0model calls
0tool runs
0wall min
0failures kept
claims
6 audited · click a claim to open the auditor's own summary — metrics, notes, control status
auditor's noteZero-teacher audit (Theorem 4.1): t1=208, grokking at 7792 steps, within the theory's worst-case t2>=9517. All three per-step envelopes hold with zero substantive violations.
grokking_time_empirical7792
t1_empirical208.0
t2_theory_lower9517
substantive_violations0.000
overall_passtrue
auditor's noteRealizable-teacher audit (Theorem 4.2 / Eq. 8): grokking at 7796 steps, zero substantive violations.
grokking_time_empirical7796
substantive_violations0.000
auditor's noteDecomposition into training-convergence / poor-generalization / eventual-generalization phases matches Theorems 4.4-4.6 envelopes on the C1/C2 trajectories; zero violations. Negative control (m~=n) shows the predicted anti-grokking collapse, confirming tightness.
phase_envelope_violations0.000
auditor's noteFigure-2 sweep: lambda-down amplifies t2 (~1/lambda), n-up grows t1, nu^2-up grows t1 logarithmically, m has minor effect. Matches paper trends qualitatively.
lambdarecorded
nrecorded
mrecorded
nu_sqrecorded
auditor's noteTwo-layer ReLU experiments (Figures 3 & 4): clear grokking in the demo run; hyperparameter sweeps show the predicted qualitative dependence beyond the linear setting. For lambda=0.05 the test loss had not crossed threshold by 100K steps; the run was extended to 200K and grokking was caught at t2=109K (documented in notes.md as a near-miss).
demo_t12000
demo_grokking_time5.560e+4
auditor's noteFalsification attempt: condition relaxations were run to find a setting where the paper's prediction breaks; the negative control behaved as the theory predicts. Attempt preserved with full trace.
checkedtrue
why the judge trusts this trail
The verdict is bound to the evidence trail, not the outcome — five dimensions, all checked.
5/5PASS
structureclaims.json (5 claims), audit_report.json present
evidence6 audited claims; issues: none
integrity0 claims skipped as untestable, all declared untestable at extraction
cost disclosureper-claim attempt counts recorded; wall time in run files + trace
trace10 events; legacy hand-driven run — adaptation documented by 10 fixture events (no LLM/tool calls captured by construction)
figure evidence
the trail
0 model calls · 0 tool runs · 0 min wall · append-only, failures preserved
0:00