Two tracks, and the withdrawal that set the standard for both.
One track is academic and faculty-advised: whether a model’s hand in a company’s own prose can be seen at all, and if it can, whether the market prices it. The other is an allocation framework I designed, built and operate myself. This page says what each one is, where it actually stands, and what would show it to be wrong. The work is written up in two Case Studies, and this page is not a summary of them.
I withdrew my own published result.
In June 2026 I compiled a report carrying three headline findings and sent it out. I then checked it against its own source data and against filings I pulled fresh, found three defects, and superseded it: a dated correction, saying the report was superseded and not to be cited, to go to every recipient I could identify. I cannot be certain where the original traveled, and that is a large part of why a correction exists at all rather than a quiet replacement. The findings themselves are not repeated anywhere on this site, including on this page.
All three are named and dated in the Case Study. One of them governs everything I have done since: the headline effect turned up at dates I had invented, where nothing whatever had happened. None of that was carelessness, which is the uncomfortable part. Every call underneath it had been made while the result was already on screen, and not one had been logged as a call, because none of them felt like one at the time.
Which is why more care is not the remedy: care pointed at an expected answer is what built the defect. The remedy has to be a commitment entered into while the numbers do not yet exist, in a form able to disqualify the answer I would rather have.
The Gate is what that commitment looks like.
Before any detector was run against the labeled data, the pass bar for the rebuilt study was written into a file with a date on it. It is a real bar and it is worth stating as a number: an AUC of at least 0.70 at the moderate rung of edit intensity, with false positives held at or under 5% against genuine pre-2020 footnotes. Four candidate detectors were named in the same file. One that does not reach the bar is disqualified, and published as disqualified.
What the bar buys is that the bar cannot move. The rungs are not re-cut afterwards, a fifth detector is not added because the first four disappointed, and the panel of filings is not scored at all until something clears it. That is what turns a negative answer into a reportable finding instead of an embarrassment — and the same commitment, one track over, is why the allocation framework’s measured results are absent from this site rather than printed with a caveat beside them.
Why the withdrawn findings are not restated on this page
The superseded report is the one piece of work here whose conclusions are deliberately unavailable, and that is a decision rather than an oversight. Restating them would put three figures back in front of a reader who has no way to re-derive them, attached to a sentence explaining that they are wrong — and a figure travels a great deal further than the sentence qualifying it.
What is published instead is the defect behind each one, which is the part somebody else can use: it says what to look for in work of the same shape, and it can be checked against the pipeline that replaced it.
What each one is, and where it stands.
Neither track is finished, and they are unfinished in different ways: one has its instrument built and a result it is not yet permitted to compute, the other has a system that runs on a schedule and a history far too short to judge it by. Both are set out below in the same three slots, and the third slot is the one that says what would show the work to be wrong.
Exhibit 1 Both tracks have their machinery settled, and neither has yet met the test that could prove it wrong
Select a slot for a link to that track’s standing in full; each track ends in its Case Study.
View the data
| Research Track | Settled | In progress | What would falsify it | Case Study |
|---|---|---|---|---|
| Track 01 — the filings study | A corpus that predates the models and traces to its source text; an instrument scored at three edit intensities | Four detectors, named in advance, measured on the labeled set; real filings left unscored until one clears the Gate | No detector clears the Gate, or the return effect vanishes once a time trend is in | The Research |
| Track 02 — the allocation framework | An unattended loop on a timer, separable components, and a replay harness that shares the live code | Backfilling the stored history so a replay spans several market environments | No edge over holding the broad market once trading costs are charged, or regimes that move on a re-fit | MacroSense |
Can a model’s hand in a filing be seen — and does it get priced?
Independent · faculty-advised · Engagement 04
A great deal of finance research reads corporate disclosure as text: how hedged it is, how readable, how far it has moved since last year. All of it assumes a person chose the words. Where an issuer put the paragraph through an assistant first, those measures are partly measuring the assistant, and a company that reads as newly careful may just have changed its drafting habits.
How often that happens is unknown, for a narrow reason: the detectors that exist were built to separate web prose from model output, and the notes to a set of financial statements resemble neither. None of them has been shown to work on this kind of writing at all. So the study runs in two halves and the second waits on the first. Half one asks whether the edit is visible, against text whose authorship is beyond doubt. Only if something passes does half two ask whether the market prices it. Scoring real filings with an instrument that has not passed produces a number either way, which is the trap the withdrawn version walked into.
- Settled
- The corpus is extracted, and each passage carries the filing it came from, the note it sits in and its character range, so any figure traces back to text a reader can open for themselves. Human authorship is certain by construction: all of it predates the models. The instrument reports accuracy at three levels of edit intensity rather than as one score, and a passage’s level is measured against its own unedited original instead of being read off the prompt that asked for the edit.
- In progress
- The four detectors named in advance are being measured against the labeled validation set. The panel of real filings is not scored, and is not authorised to be scored, until one of them clears the Gate. Around that work sit three things that are not results: an application for research funding, submitted and not yet decided; a conference as the intended venue whichever way the Gate falls; and a faculty relationship that is a working one rather than a name on a cover page.
- What would falsify it
- If no detector clears the Gate, the study stops at the validation stage, and the reportable finding is that editing of this kind cannot be detected with the public methods available — published with the accuracy curve attached, so that filers not doing it stays distinguishable from an instrument that cannot see it. And if a detector does clear the bar but the return effect vanishes once a time trend sits in the specification, that answer goes the same way the first one did.
The Research — the full Case Study, with the three defects named and dated
Can one person run the process a committee is supposed to run?
Self-directed · Engagement 06
An allocation decision taken in a meeting leaves nothing behind that anybody can run again. Six months later there is no process to interrogate, only a sequence of views and whatever the market did next. MacroSense is that decision built as a machine instead: on a schedule it reads public macro and market series, calls the environment as one of four regimes with a probability for each, allocates across sector and credit funds under constraints that change with the regime, scores company commentary out of public filings and earnings calls, and writes its own note in plain English explaining the change it has just made.
Every input is free and public, which was a design constraint rather than a budget. A process whose entire claim is that it repeats should be reproducible by the reader least inclined to believe it.
- Settled
- The loop runs end to end on a timer with nobody watching it — fetch, classify, allocate, explain, log — and it has been doing that rather than having been demonstrated once. The components are separable, each runs on its own behind a signature that was fixed early and left alone, and the replay harness calls the same classifier and the same optimizer the live system does rather than a second copy of the strategy.
- In progress
- Backfilling the stored history so the system can be replayed across several years and several market environments instead of one short stretch of one. Until that is done the harness’s output describes its window rather than the thing it is pointed at, so no performance figure for this framework appears here or anywhere else on the site.
- What would falsify it
- A history long enough to contain all four regimes, exercised against real returns and charged for the cost of trading, in which the framework does not beat simply holding the broad market. Or the regimes themselves proving unstable — labels that move when the model is re-fit are not a classification you can hang a constraint set on. Either result is publishable. The absence of a figure today is not a space reserved for a flattering one.
Why the company layer has never been in a replay
Two layers, one history, and it only reaches one of them. Every replay run so far has therefore exercised the macro half of the system while looking from the outside like a test of all of it. The Case Study says what the stored rows do and do not hold.
It belongs beside the short window rather than inside it, because a reader told the history is short would reasonably take it to be short on the whole thing. Reconstructing what had been filed and said as at each past date is the larger half of the backfill, and the reason it is not finished.
What is deliberately missing here.
Three things a reader might expect on a research page are absent from it, and each absence is the same decision: the withdrawn findings, which are superseded and so are not restated as results; the research funding, which is an application submitted and not a decision received; and any measured result for the allocation framework, which has only one short market environment to be measured over.
None of the three is left to my judgment at the moment of writing. Each has an entry in a registry of refused patterns that fails the build if one reaches a shipped file — and that registry is itself checked, by injecting every pattern into a scratch copy of the site and requiring the check to go red. What a pattern can refuse is the form a claim travels in: a figure pasted, an award verb beside the word grant. None of them refuses the same claim restated from memory in a sentence of my own. The registry is not the decision. It is what stops the decision quietly coming undone at one in the morning.