Publications

• Preprint
• Artifact
• Not peer reviewed

Alibi Distance: measuring and closing the silent-failure aperture of a verified exercise corpus

2026-09-18 · DeepForge Research

A corpus-level aperture measurement for a machine-verified exercise bank, with an independent second engine and a machine-verified practice artifact.

• ResultA full-corpus run sampled 106,081 single-edit mutants across 5,730 exercises. 17,502 of them pass every shipped test, 7,727 of those diverge from the reference on the probe bank, and 46.08% of analyzable problems admit a radius-1 alibi.
• ResultA second, separately written CPython engine, run by the same author, agreed with the TypeScript engine to within 0.14 percentage points on the census rates, 0.32 on the length quartiles and 0.44 on the closure estimates. Neither engine, nor its input, nor its output is part of this release, so the census below is a record of a mining run rather than a result an outside reader can re-derive.
• ResultAn in-sample witness closes 59.7% of affected problems and kills 73.2% of mined alibis, but a cross-validated witness kills only 48.4% of held-out alibis. The deployable estimate is therefore about 23.5% of problems still affected, not the in-sample 18.6%.
• ResultTwo planned follow-ups failed and are reported as failures: feature-guided probe selection does not transfer better than random (63.3% against 59.5% conviction within ten probes), and the best two-literal input predicate reaches 0.9 balanced accuracy for only 32.2% of alibis against a permutation control of 0.881.
• ResultA blind audit of ten independently sampled alibis found 10 of 10 plainly wrong.

• HypothesisThat near-miss mutants resemble the slips learners actually make is an assumption, not a measurement. The paper frames the aperture as a property of the corpus and the edit model, and proposes a shadow-logging study that has not been run.
• HypothesisA probe bank derived from the shipped tests is not a proof of semantic inequivalence. Divergence is established only where a probe actually separates the two programs.

• Preprint
• Artifact
• Not peer reviewed

Behavioral Delta Ledger: Per-Edit No-Op Detection on a Hidden Basis Derived from an Exercise's Own Tests

2026-09-18 · DeepForge Research

A count-only per-edit behavioural delta on a hidden probe basis, with the product boundary set by what the measurement failed to support.

• ResultA fresh full-corpus census covers 5,682 analyzable exercises and 88,357 sampled single-edit mutants. 10.23% of mutants are invisible on the 48-probe basis, 16.45% pass every shipped test, and 44.40% of problems carry a test-passing slip the basis can see.
• ResultIn-sample, the sign of the hidden delta matches the sign of the test delta for 90.20% of climb walks. With the basis built from odd-index tests and the delta scored on even-index tests, held-out concordance falls to 84.00% and 82.23%. The held-out figures are the ones reported.
• ResultOn 699 simulated edit walks under three policies, 34% to 51% of edits are exact no-ops by signature.
• ResultFour of nine pre-registered predictions failed or reversed and are reported as failures: the predicted edit-family ordering was reversed, hidden-visible test-passers missed the threshold narrowly, the next-attempt regression ratio reached only 1.14x and 1.36x against a pre-registered 2x, and cold sets were not family-concentrated (1.50%).

• HypothesisThe ledger is retrospective attribution, not forecasting. The pre-registered prediction of a 2x next-test regression risk was falsified, and the routing story implied by cold-set enrichment is post hoc and was cut from the shipped design rather than reported as a finding.
• HypothesisAll learner-behaviour numbers come from simulated edit policies and synthetic cohorts. No learner data was collected, and the delta is not a measure of understanding.

• Preprint
• Artifact
• Not peer reviewed

Refutation-Ledger Values: Auditable Warrant Accounting for Contestable Derived Claims

2026-09-18 · DeepForge Research

A value-level contract in which recorded falsification attempts produce a recomputable warrant state and exact cite-graph demotion.

• ResultAcross 200 frozen seeds and 12 matched conflict pairs per regime, the shipped declared-dependence relation reaches pair-win 1.000 and AUC 1.000 in the correlated regime, where the strongest cheap count baseline reaches 0.000. Audit precision at 12 is 1.000 against 0.000.
• ResultSeed-only churn moves the mean grade by 0.000 and flips 0.000 of pairs.
• ResultThe syntactic tuple variant was declared in advance as a negative arm and was killed by its criterion, as predicted.
• ResultNo individual mechanism is new. Retraction in truth-maintenance systems, evidence fusion, audited ledgers, and attack-family scoring are all established; the contribution is the ledger-to-grade-to-demotion contract that ties them to a runtime value.

• HypothesisThe arena's ground truth is generated from the generator's own blind-spot table, so the result is a property of the contract under that generator and not evidence about educational content.
• HypothesisThe dependence relation is declared, not semantically proved, and a well-formed fabricated ledger can still pass a recomputation audit. Audit is arithmetic consistency plus anchored tamper evidence at the append point; it is never a truth verdict.
• HypothesisWhether a recomputable warrant state is useful to a learner is untested. No study asks.

• Technical report
• Not peer reviewed

Ladder-Graded Spacing: execution-aware scheduling for code practice

2026-09-18 · DeepForge Research

A measurement report for the platform's own review scheduler, run as a simulation against a stated truth model.

• ResultIn a 100-learner, 40-item, 365-day simulation against a memory-based truth model, the shipped variant used 25.63% fewer reviews per retained item than SM-2 at matched first-attempt recall (25.32% in the literal protocol), with a Brier delta of -0.1105.
• Result30 of 30 seeds passed the 10% gate with no seed worse, and no edge cohort exceeded the 1.10x SM-2 review gate.
• ResultAn ablation removed the proposed evidence multiplier: it cost 27% more reviews for no Brier gain, so the shipped variant does not use it.

• HypothesisEvery learner is simulated. This is a report on a scheduler under a stated truth model, not evidence about human memory or learning.
• HypothesisThe baseline is SM-2, the scheduler the platform used before. No published memory model is evaluated on the same protocol, so the result does not establish that this variant is the best available scheduler.
• HypothesisHint debits have no published calibration, and the effort cap is the weakest-chosen default in the model.

• Technical report
• Artifact
• Not peer reviewed

KeyFuse: Falsification-First Cache-Key Auditing with Minimal Collision Witnesses and Conservative Key Repair

2026-09-18 · DeepForge Research

A read-only auditor for build-cache keys. Software-engineering work adjacent to the agenda, recorded as a specification plus an artifact.

• ResultOn a frozen 24-task corpus with brute-force ground truth (408 binary assignments, 651 full product assignments, 1,533 enumerated same-key collision pairs), the exact arm matched ground truth in 69 of 72 task-by-strength cells, with 3 explicit nondeterminism refusals.
• Result83 minimized witnesses were produced, 81 verified 1-minimal with 2 flagged and reported, and repair separated all 101 recorded collision witnesses in the separating check with 0 unseparated.
• ResultMasking is demonstrated rather than assumed: a pinned adversarial covering array misses the (a,b) pair of f = (a AND b) XOR c, and the miss is recorded.

• HypothesisThe anchoring assumption, that every output difference has a witness reachable by a small displacement from the baseline which includes a differing slot, is stated in the code, the paper, and the interface. It is not proved and not verified.
• HypothesisDetection is not soundness and a repaired key is conservative rather than complete. The certificate is exactly: no detected effect of that support at the covered tuples.
• HypothesisCovering-array testing, delta debugging, and cache-key auditing all predate this work, and the frozen corpus is 24 synthetic tasks rather than a production build graph.

• Technical report
• Artifact
• Not peer reviewed

REPROGPU: Cross-Adapter Bit-Reproducible WebGPU Kernels and a Conformance Harness for a Declared Integer Subset

2026-09-19 · DeepForge Research

A conformance harness for a declared integer WGSL subset, recorded as a narrow artifact contribution rather than a new method.

• ResultThe CPU-only gate reproduces the pinned reference hashes byte-for-byte across five of five criteria, matches three published Philox known-answer tests and the FIPS SHA-256 vectors, pins five WGSL source hashes, and finds no float type, atomic, discard, or unmasked runtime shift in the integer kernels.
• ResultThe subset is exact by construction: integer wrapping arithmetic, bitwise operations, comparisons, bitcast between i32 and u32, and host-shared little-endian layout. WGSL leaves rounding direction, reassociation, fusion, and subnormal flushing unspecified for floating point.
• ResultA float matmul kernel is carried as an explicit negative control with no pass criterion, and its divergence is the expected outcome.

• HypothesisCross-adapter equality is demonstrated only by the browser lab, on one machine, at one time. The gate cannot compile WGSL or run a GPU, and no CI artifact is evidence of GPU agreement.
• HypothesisAdapter identity is self-reported and not attested, and the hardware sample is small.
• HypothesisExact integer GPU execution hashed and compared already exists, and reproducible floating-point BLAS work is mature. The contribution is the narrow protocol and the pinned subset, not a new capability.

Census by real execution

The reference solution and every sampled mutant are executed in CPython against a deterministic per-problem probe basis. No model, no network, no learner telemetry is involved.

It measures this corpus under one stated edit model and one probe basis. A program that agrees with the reference on every probe is not thereby proved equivalent.

Alibi Distance · Behavioral Delta Ledger

Independent re-execution

A second engine, written separately, recomputes the census from scratch and reports its own rates.

It is an internal replication by the same author on the same corpus. It is not external replication, and agreement between two implementations still leaves both open to a shared mistake in the edit model.

Alibi Distance

Simulation against a stated truth model

Synthetic learners are scheduled for a fixed horizon against a memory-based truth model, with paired-by-learner bootstrap gates repeated over 30 seeds and explicit edge-learner cohorts.

It cannot say anything about human learners, because there are none. It is a property of the scheduler under the chosen truth model.

Ladder-Graded Spacing

Leakage-controlled scoring

The probe basis is built from one subset of the shipped tests and the delta is scored on the disjoint remainder, so the reported concordance is not measured in sample.

It removes basis and score overlap. It does not remove the blind spots the probe basis inherits from the shipped tests.

Behavioral Delta Ledger

Synthetic arena with frozen seeds and a pre-declared negative arm

A fixed generator produces matched conflict pairs across regimes, run over 200 frozen seeds against nine graders, one of which was declared as a negative arm before the run and is expected to fail.

Its ground truth is read off the generator's own blind-spot table, so it measures the contract under that generator rather than the world.

Refutation-Ledger Values

Brute-force ground truth

Every assignment over a finite typed slot universe is enumerated and every same-key collision pair is listed, so a detector can be scored against a complete answer rather than a sample.

The universe is 24 synthetic tasks, not a production build graph, and exhaustive enumeration is affordable only at that size.

KeyFuse

CPU reference gate plus browser demonstration

The gate recomputes pinned reference hashes, published known-answer tests, and source hashes on the CPU. The browser lab executes the kernels and records a canonical manifest of per-adapter output hashes.

The gate cannot compile WGSL, execute a GPU, or compare two adapters. Only the browser run is a demonstration, and only on the machine that ran it.

REPROGPU

Permanent gates in continuous integration

Every instrument ships a gate that recomputes its pinned aggregates and fails the build when they drift, so a change that silently invalidates a published number cannot merge.

A gate re-derives a published number from the same repository at the same commit. It is a drift alarm, not an independent reproduction.

Alibi Distance · Behavioral Delta Ledger · KeyFuse · Refutation-Ledger Values · REPROGPU

• Unpublished
• Not peer reviewed

Shadow-logging study

Do the mined near-miss programs resemble the errors learners actually make?

No data collected. The coupling between a mutation and a learner slip is assumed throughout the Alibi Distance paper, and a shadow-logging study is proposed there. It needs learner telemetry, consent, and ethics review before it can start.

• Unpublished
• Not peer reviewed

Assisted-tutoring comparison

In the age of generative AI, does evidence-bearing feedback change outcomes against an unassisted control?

No protocol is written, registered, or run. No comparison, no effect size, and no claim about AI-assisted learning exists in this portfolio. The agenda names the setting; the measurement has not happened.

• Unpublished
• Not peer reviewed

External replication

Do the corpus-level rates hold on an exercise corpus authored by someone else?

Not started. Every figure here comes from a corpus whose author also wrote the measurement code, so the main threat to validity has not been addressed.

• Unpublished
• Not peer reviewed

Interventional study of the behavioural delta

Does surfacing per-edit behavioural change help a learner, or does it become surveillance?

Not started. The ledger is deliberately opt-in, count-only, and outside assessment; an intervention would change that boundary and would need a design that a learner can refuse.