Independent research

Peer-reviewed and in print.

1. Journal of High School Science (JHSS) · Volume 10, Issue 2, pp. 521–545 · 2026

   Restoring grokking under structured label noise: audit efficiency and optimization-history dependence

   When mislabeled training data gets cleaned up, a model's training history matters — not just the corrected dataset it ends up with.

   Awarded the Certificate of Excellence in Scholarship for the best original manuscript of the issue.

   Neural networks sometimes "grok": long after they appear to have simply memorized their training data, generalization arrives suddenly. Mislabeled data suppresses that. This paper asks not whether label noise breaks generalization, but how and when generalization can be restored once you start fixing labels — and how to spend a limited annotation budget doing it.

   Two results. First, recovery is path-dependent: correcting labels partway through training and continuing recovers generalization faster than retraining from scratch on exactly the same corrected data, which introduces optimization history as a variable in grokking dynamics. Second, a small review budget spent on the examples the model finds hardest restores generalization far more efficiently than reviewing examples at random.

   The conditional half of the result is the part worth keeping: region-targeted auditing helped only when the error geometry it assumed matched the noise that was actually there. Under uniform noise it bought nothing at all.

   What it does not show The testbed is synthetic modular arithmetic — predicting (a + b) mod 97 — not real-world data, and the efficiency figures belong to the specific audit budgets they were measured at. It studies how generalization recovers; it does not explain or solve grokking.

   Read the paper ↗ Code ↗

2. Journal of High School Science (JHSS) · Volume 10, Issue 3 · 2026

   Target-family differences in stereochemical discrimination: a same-assay analysis of enantiomer pairs in ChEMBL

   How sharply different families of proteins tell a molecule apart from its own mirror image, measured across thousands of pairs that were tested side by side in the same experiment.

   Shireesh Apte of JHSS called it a "strong contribution."

   Many drug molecules are chiral: they come in two mirror-image forms, called enantiomers, which are identical in composition but cannot be superimposed — like a left and a right hand. A protein binding pocket is itself handed, so the two forms usually differ in potency. The ratio between them is the eudysmic ratio, and it is normally quoted one compound at a time.

   The trouble with measuring it at scale is noise: potency values pooled from different laboratories scatter by roughly 0.68 log units, which is larger than many of the effects of interest. This paper removes that by construction, comparing two enantiomers only when both were measured under the same assay identifier — same conditions, same protein preparation. From ChEMBL 37 that yielded 47,350 pair measurements, collapsing to 23,086 unique structural pairs, of which 12,344 map cleanly onto one of five protein families.

   Across those pairs the median log eudysmic ratio was 0.48, with 27.7% exceeding tenfold and 7.5% exceeding a hundredfold. Grouped by family the medians ordered as protease 0.98, nuclear receptor 0.77, GPCR 0.59, ion channel 0.40 and kinase 0.39 — with kinase and ion channel statistically equivalent. The effect survived controls for the ligand chemistry each family attracts, and aggregating to individual proteins produced a target-level index that held up across entirely separate chemical series.

   What it does not show This is an association with the class of target a pair was assayed against, not a demonstrated causal property of the binding pocket. Effect sizes are modest and the distributions overlap heavily, so it is a shift in typical value and tail weight rather than a categorical divide — and while protease and nuclear receptor sit reliably above kinase and ion channel, their order relative to each other is not stable across weightings. The same-assay design removes differences between laboratories, not the ordinary error within a single assay, which was not separately quantified.

   Read the paper ↗

Submitted, with decisions pending. Nothing here is published yet.

1. Journal of Cheminformatics · submitted July 2026 · 2026

   Carried but Conditionally Used: A Frozen Held-Out Audit of Whether Chirality-Aware Molecular Fingerprints Predict Measured Same-Assay Enantioselectivity

   A molecular fingerprint can encode which mirror image a molecule is. That is a different question from whether a model actually uses it to predict which one works better.

   My entry for the Regeneron Science Talent Search.

   Chemists represent molecules to software as "fingerprints" — bit patterns summarising which fragments a structure contains. This paper separates two things that are easy to conflate: whether a fingerprint carries stereochemical information at all, and whether a model uses that information to predict measured differences in potency between mirror images.

   Carry and use were audited separately, under a protocol frozen and hashed before any held-out modeling took place, with the targets used during development excluded from the held-out set. Stereochemical information was carried essentially completely at fingerprint radius 3. Use was far more selective: 5 of 11 held-out targets came out positive under scaffold extrapolation.

   The title is the finding. The information is carried; it is only conditionally used.

   What it does not show Under review — not accepted and not published. The positive targets are specific results under one frozen protocol, not a general claim that chirality-aware fingerprints work or fail.

   Reproducibility archive ↗

2. Journal of High School Science (JHSS) · 2026

   A Completeness-Corrected Rotation Census of the Jupiter Trojan Swarms: The Apparent L4/L5 Asymmetry is a Survey-Coverage Artifact

   An apparent difference between Jupiter's two asteroid swarms turns out to be a record of how unevenly they were observed, not a difference between the swarms.

   Jupiter shares its orbit with two great clouds of asteroids — the L4 swarm running ahead of the planet and the L5 swarm trailing behind. How fast they spin carries a record of their collisional history, so any genuine difference between the two swarms would be telling us something real about how they formed.

   Rotation periods were recovered for 811 numbered Trojans from public Zwicky Transient Facility photometry, and the raw numbers do look asymmetric: periods came out for 31% of L4 objects against just 21% of L5. But at matched brightness, L4 objects had been observed about twice as often — a median of 439 measurements against 222.

   To test whether anything survived that, synthetic light curves were injected into the real observation times of each swarm to measure how often a period could be recovered at all. Compared only where the two swarms are equally well observed, no statistically significant difference remains (KS p = 0.24; difference in fast-rotator fraction −0.01 ± 0.06). The asymmetry is an artifact of coverage.

   What it does not show This is a null result: no detectable difference under matched completeness, which is not the same as proving the two distributions are identical. Slow rotators beyond about 100 hours cannot be recovered from sparse ground-based photometry at all, and are excluded from the comparison rather than accounted for.

   Code and data ↗

3. Journal of High School Science (JHSS) · 2026

   Are Catalytic Metal Sites More Distorted? An Archive-Wide, Confound-Matched Test of the Entatic-State Hypothesis in Zinc and Iron

   A 60-year-old idea about how enzymes hold their metal atoms, tested across the entire structural archive — and a warning about how much archive-wide geometry depends on the software that produced it.

   The entatic-state hypothesis, proposed by Vallee and Williams in 1968, holds that enzymes grip their catalytic metal ions in strained, awkward geometries that leave them poised to react. It has been influential for six decades. This paper tests it at archive scale: geometric distortion was computed for every zinc and iron site in the Protein Data Bank, each site was labelled catalytic or structural using independent curation that never looks at geometry, and the two groups were compared while matching on metal, coordination number, donor atoms and resolution.

   No robust distortion difference survives. Every apparent signal proves unstable to at least one analysis choice — how resolution is adjusted for, which refinement software produced the structure, how donor sets are matched, or how the labels are defined.

   That instability turned into the second half of the paper. Of 27 geometric quantities of the kind routinely reported in archive-wide studies, 9 are materially dependent on which program refined the structure: median iron–oxygen distance differs by 0.078 Å between REFMAC- and PHENIX-refined entries, and even iron-bound water — chemically identical everywhere — differs by 0.098 Å. In a survey of 74 archive-wide analyses from 2015–2025, 63.5% never name the refinement program at all.

   What it does not show A null result under this matching scheme. It does not disprove the entatic-state idea for any particular enzyme; it shows that the signature does not appear archive-wide once confounds are matched, and that geometry drawn from a structural archive carries a software fingerprint most analyses never check.

   Code and data ↗

Written up but not yet submitted.

1. Independent research · manuscript in preparation · 2026

   Do Lens Detectors Need Look-Alike Negatives? A Controlled, Preregistered Test of Hard-Negative Training in Strong Gravitational-Lens Searches

   A popular fix for false positives in gravitational-lens detectors works mainly because it adds matched training data — not because of the look-alike shapes everyone assumed were doing the work.

   Strong gravitational lenses — rare alignments where a galaxy bends the light of a more distant one into arcs — are hunted automatically, and the hunt drowns in false positives. The standard fix is to retrain on hard negative examples that look like lenses, usually spiral galaxies. This study ran the control that had never been run: 90 training runs comparing spirals against ordinary galaxies matched one-to-one on brightness and distance.

   Both cut false positives by almost exactly the same amount. A preregistered follow-up then failed its own success criterion, written down and frozen before the experiment ran, and development stopped as promised.

   What it does not show Manuscript in preparation. No lenses were discovered in this work, and the detector is not deployment-ready.

   Read about it ↗

2. Manuscript in preparation · not yet released · 2026

   Schistosomiasis egg detection in urine-filtration microscopy

   Machine-learning detection of parasite eggs in microscopy images, aimed at the conditions where the diagnosis actually has to be made.