Contact

Correspondence is welcome. Solicitation is not possible.

Research correspondence, requests for the full text of a paper, and corrections to anything we have published: research@axoquant.com

AxoFund trades its own capital. There is no capital to raise, no mandate to win, and no recruiting funnel, which removes most of the reasons a firm answers its post. What is left is method. Method we will discuss at length.

What we answer

The register below is not a filter for politeness. It is the list of questions we can answer completely, which is a shorter list than the one we would enjoy answering.

Replication. Every published paper states its apparatus: sample size, the number of independent windows, and the procedure used to control false discovery. If that is not enough to rebuild the result independently, the defect is in the paper and we would like to hear about it.
Method in the abstract. Evidence protocol, the gate chain we retired in August 2026 and why, what we take a verdict to mean, and where we think the reasoning is weakest. We will argue about method for as long as you like.
Corrections. Including misreadings of the literature on our reading list. Mischaracterising someone else's work is our error, not theirs.
Counterparty and operational enquiries. Answered where the answer does not require describing a position we hold.
Not capital. We manage none and can accept none. That is a description of what the firm is, not a stage it is passing through.

Full texts

Abstracts are public. Full texts carry operating parameters (the settings, budgets and thresholds a result was actually produced under) and are shared on request with calibration redacted. The redaction policy is itself published: every item in the regime library ends with the list of what was withheld from it and why, so a reader can see which class of value was removed before deciding whether the redacted text is worth having.

Redactions are declared at the point they occur. Where a value is withheld the sentence says so and names what kind of value was withheld, rather than being quietly written around. An omission would read as though the number never existed; a declared redaction tells you it exists, that we hold it, and that we chose not to print it.

Corrections

Reports of errors in anything we have published (a wrong number, a broken derivation, a paper on the reading list we have characterised badly) are acted on in public. An item in the regime library is never silently edited: the correction is issued as a new item that cites the original, and the original's status line is changed to name it. A correction costs us a paragraph. Being wrong for a year costs us the corpus that was built on top of it.

The same discipline runs internally, on an append-only research journal. It is the reason we can find our own errors at all.

Dated and chronological. Every substantive question and finding gets an entry on the day it happened, in order.
Append-only. A prior entry is never edited to fix it. A wrong entry is corrected by a later entry that supersedes it, and both stay.
Surprises recorded. Each entry states what did not behave as expected and which assumptions turned out to be wrong. Those are the two fields that decay fastest from memory and matter most in review.
Adverse notes required. Each entry lists what a reviewer should sanity-check: leakage risks, slices too thin to carry the claim resting on them, and framing choices that flatter the result.
Artefact over record. Where a written record and the artefact it describes disagree, the artefact wins — and the disagreement is itself a finding, not a typo to be tidied.

The rule is only worth stating if it costs something, and it does. A routine consistency probe found that a journal entry disagreed with the campaign record written the same day from the same artefacts, on four of the five counts the two documents shared. The artefacts settle it. Under the house rule the entry does not get patched: the correction is a later entry that supersedes it, and the original stays. Both readings remain on the record, which is the only reason the discrepancy can be pointed at rather than quietly reconciled. An editable journal would have absorbed it silently, and we would have learned nothing about our own transcription error rate on same-day material.

Documents carry the same treatment on their face. Campaign charters here are headed with dated correction notices naming what was withdrawn and the audit that withdrew it, and the withdrawn claim stays in the document as the case study of its own refutation.

A retraction

In August 2026 we retracted one of our own working papers. Its conclusions had already circulated internally, and several of them we liked.

The corpus underneath it had been optimised to roughly a fifth of production's dimensionality-scaled search depth, and the shortfall was not uniform: it widened with the number of parameters a strategy carried. That makes the sample invalid rather than merely underpowered, because the ranking it produced could have been generated by the budget alone: more complex strategies were more starved of search, and complexity is correlated with everything else we were measuring. No amount of additional cells fixes a confound that sits in the design.

So: the paper stays up with the withdrawal on its face rather than being deleted, the corpus was archived under a name that marks its era rather than being dropped, and every conclusion in it — including the ones we liked, and specifically those — is treated as unvalidated pending a redo at the correct budget.

The confound was named in that paper's own limitations section. That is the point worth taking away. A confound named in limitations is a scheduled experiment, not a disclaimer; here the experiment was scheduled, it ran, and it retracted the paper that scheduled it. A limitations section that never becomes an experiment is a way of being wrong on the record with a clean conscience.

Item status vocabulary, as printed on the face of every item in the library. Closed set of five — each item carries exactly one.
Status	Meaning
current	The item as published stands.
revised	Text changed, conclusions unchanged. Carries the revision month.
superseded	A later item replaces it. This one stays up, and names its successor.
corrected	A specific claim is wrong. The correction is a separate item that cites this one; the original is not edited.
retracted	The evidence underneath it does not support it. The item stays up with the retraction on its face.

What we cannot answer

Three classes, stated plainly so that nobody has to spend a letter finding out.

Live positions. A refuted result is publishable, and we publish them. A live one is not. This extends further than it first looks: we name the mechanism a book trades — a rebalance mandated by a prospectus, an administered price limit, a defended currency band, the legal identity between two share classes — and not the market it trades in. The set of markets we screen is information even when every result in it is dead.
Calibration. Score floors, cut-offs, tuned thresholds, detection floors in absolute units, and the discrimination of our own instruments. We publish the ordering, the direction, the sample size and the policy; the values themselves are withheld. The commercial reason is obvious. The technical one is better: a metric's calibration is a property of the pipeline that produced it, not of the metric. We have watched a statistic move from one end of that ordering to the other when the search budget underneath it was corrected: same statistic, same corpus, different pipeline. Published without its pipeline, a calibration value would mislead an honest reader as efficiently as a dishonest one.
Performance. No returns, no per-book statistics, no capacity or dollar figures, at any level of aggregation, in any correspondence. This is policy rather than modesty, and it is what makes the last sentence of the footer below true.