Blog · arXiv Analysis · Last reviewed July 10, 2026

The Intervention Claim Becomes the Confidence Sequence

Amir Asiaee's July 2026 arXiv paper turns mechanistic-interpretability interventions into auditable causal estimands, then asks whether the claimed fidelity survives uncertainty, repeated peeking, and adaptive counterexample hunting.

Published: July 10, 2026 Modified: July 10, 2026 Last reviewed: July 10, 2026 Tags: arXiv, mechanistic interpretability, AI governance, evaluation, confidence sequences, causal claims

The Paper

The paper is Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretability, arXiv:2607.08349 [cs.LG]. The arXiv record lists Amir Asiaee as author, with version 1 submitted on July 9, 2026. The arXiv page says the paper was accepted at UAI 2026, the PDF is 18 pages, and the HTML lists the Department of Biostatistics at Vanderbilt University Medical Center.

The paper belongs beside mechanistic interpretability, circuit variance, interpretability receipts, sparse circuit auditing, false edit levers, and AI audit trails. Its fresh angle is statistical: an explanation should not become a claim until the intervention, target distribution, metric, and stopping rule are auditable.

The Causal Claim

Mechanistic-interpretability papers often test explanations by intervening inside a model: swapping hidden states, patching activations, ablating components, tracing causal paths, or comparing a compressed mechanism to the original network. Asiaee's critique is that these experiments are often compressed into a single point estimate, even when researchers watch the run, change the evaluation set, or search for failures as the experiment proceeds.

Certified Interventional Fidelity, or CIF, rewrites the reported quantity as a causal estimand: an expectation of a bounded score over a declared input distribution and a declared intervention distribution. That declaration separates the scientific target from the sampling procedure. A circuit-completeness score, an interchange-intervention accuracy, or an activation-patching recovery number is not just a number. It is a population claim over specified inputs and specified interventions.

The Sequence

The "anytime-valid" part is the governance hinge. A fixed confidence interval protects a fixed experiment; it does not protect a workflow where an evaluator checks after 100, 500, or 1,000 interventions and stops when the evidence looks strong. CIF uses confidence sequences, which remain valid under repeated monitoring and optional stopping. It also supports adaptive failure-directed sampling through bounded mixture importance weighting, so the evaluator can spend more effort on suspected weak points without silently changing the estimand.

The paper instantiates the framework with transparent Hoeffding-style sequences and variance-adaptive betting sequences. The betting sequences matter because Hoeffding bounds are simple but conservative. The paper reports 10-30x certification-cost reductions for betting sequences in its experiments, turning some certifications from thousands of forward passes into tens or low hundreds.

Experiments

The experiments cover a controlled MNIST neural-abstraction benchmark and GPT-2 Small on the Indirect Object Identification task. For GPT-2 Small, the paper specifies a 12-layer, 12-head model with a 768-dimensional residual stream and roughly 124 million parameters, then evaluates IOI circuits from ACDC, attribution patching, AtP*, and a hand-identified circuit.

The strongest practical numbers are about uncertainty cost. On the IOI task, at 2,000 samples, the paper reports Hoeffding radii around 0.070 and betting radii around 0.005. Its table reports that the full 13-head IOI circuit certifies normalized patching recovery of at least 0.90 at 102 samples and at least 0.95 at 357 samples under betting; under Hoeffding, the corresponding requirements are 1,875 samples and more than the 2,000-sample budget. Even the 3-head name-mover circuit certifies at least 0.90 recovery in 110 samples under betting.

The paper also validates coverage over 500 independent runs in the MNIST setting under i.i.d. sampling, adaptive mixture sampling, and aggressive peeking. Hoeffding is conservative in every reported configuration; betting stays at or above nominal coverage while producing tighter intervals. The published repository includes the code, saved result CSVs, notebooks, and paper-to-code map, under an MIT license.

Limits

CIF certifies an estimate for a chosen metric and chosen distribution. It does not prove that the metric captures mechanistic truth, nor does it solve the identification problem of whether the proposed abstraction is the right causal structure. It makes overclaiming harder, but it does not make interpretability automatic.

The framework also depends on bounded scores or bounded clipping, and multiple-comparison control still matters when scanning many components or circuits. The paper discusses Bonferroni correction and future e-value false-discovery-rate approaches for larger circuit searches.

The Spiralist reading is that interpretability should stop treating a causal diagram as a finished artifact. The claim is not complete until the intervention distribution, uncertainty process, and stopping behavior are part of the record.

The Receipt

An interventional-interpretability receipt should record the model, layer or component set, explanation class, input distribution, intervention distribution, intervention operator, bounded score, estimand, sample schedule, peeking rule, confidence level, confidence-sequence construction, importance-weight cap, adaptive proposal, paired-comparison rule, multiplicity correction, stopping time, raw point estimate, interval or confidence sequence, failed certifications, code commit, random seeds, and final claim boundary.

The audit question is not "did the patching score look high?" It is "what causal population claim was tested, how often did the evaluator look, and what uncertainty survived that workflow?"

Sources

Return to Blog