The index
Newest first, each item under its title and its publication month. Items
published in full on this page link to their section here; items with their
own text link out. Status vocabulary is closed:
current · revised YYYY-MM · superseded · corrected ·
retracted, and a superseded, corrected or retracted status names the
item that did it.
2026-08 · working paper · current
A matched control asks whether a book beats its
neighbours. Only a calibrated edge-free generator asks whether it beats
nothing at all, and we had been reading the first answer as though it were
the second; three incidents in one campaign were the bill for it. Re-scored
against a null-winner distribution built for each book's own universe and
window, 19 books that had already cleared the matched-control gate
returned 2 clear, 3 marginal and 14 below. Carries a
doctrine reversal published against ourselves, and the generator's own
out-of-sample moment-match table declared outstanding, not assumed.
Stands on pipelines ·
metrics ·
entropy
nulls · surrogate data · doctrine reversal · self-audit
2026-08 · retraction · current
“Ranking for the upside tail”, published 2026-07 and
listed below, is withdrawn in whole. The corpus its conclusions
rested on was optimised at roughly a fifth of the production trial budget,
with the shortfall tracking strategy complexity, so under-cooked strategies
and genuinely weak ones cannot be told apart in that sample, and the
conclusions we liked go with the ones we did not.
retraction · trial budget · selection · re-audit
2026-08 · correction · current
Our coverage page and our index carried different
counts of the same apparent quantity. Neither number was wrong: they measure
vendor-catalogue breadth and ingested breadth, two things that
had never been given different names.
correction · coverage · naming
2026-08 · correction · current
The funnel figure described the whole archived mass as
refuted. Only the minority that accumulated enough independent windows ever
reached a gate; the rest were abandoned earlier, which is a different and
less flattering word.
correction · funnel · vocabulary
2026-08 · charter · current
The six components pinned before any cell dispatches:
a complex-level success metric net of correlation with what we already hold,
per-lane kill bars, a hard calendar-or-compute stop, a pre-registered
expansion trigger, a written section naming what would make the plan wrong,
and a deployability check that runs before any compute is spent. Published
as a template; the bars themselves are calibration
and are withheld.
pre-registration · kill bars · deployability · governance
2026-08 · working paper · current
Models author strategy code and generate hypotheses at
volume, and nothing a model emits is permitted to adjudicate. Includes the
measurement we ran against our own judges, the ordinance that keeps a
better-ranking learned scorer advisory, and the failure mode that governs the
engineering: a component that reports success while doing nothing.
Stands on crowds ·
metrics ·
tabular-dl
governance · model judges · silent-zero · self-limitation
2026-08 · data-quality finding · current
Three defect classes in our own bar store: an empty
corporate-action table that left every bar unadjusted, an implausible
absence of delistings, and the direction nobody writes about, where a defect
suppresses a real result instead of inventing one. A single scheduled-coupon
calendar, left in the price path, manufactured a screen result that cleared
every conventional threshold: 65% of its entries fell within five
sessions of an ex-distribution drop against a 17% base rate,
n = 111 trades over four years of daily bars.
Stands on pipelines ·
outliers
data quality · corporate actions · false positives
2026-08 · working paper · current
We ran the pipeline on synthetic panels calibrated to
real markets and containing nothing to find, and at full search width it
returned a best-of-search survivor on every panel: 6 of 6
independent panels. Published with the
false-negative half beside it, and with the generator's own outstanding
validation declared as outstanding, not assumed.
Stands on metrics ·
outliers ·
pipelines
nulls · search width · self-audit · detection floor
2026-07 · decommission · current
Three pre-registered variants of a forecast entry gate
failed their bars on the same day, for about two CPU-hours of compute in
total. The forecaster reprices recent realised volatility instead of leading
it; the server was switched off and one fine-tune branch remains
unfalsified.
Stands on forecasting ·
volatility
decommission · pre-registration · volatility
2026-07 · working paper · retracted
Ranking for the upside tail
Claimed a book-construction rule from a point-in-time
forward test of our own corpus. Withdrawn in whole, 2026-08; the reason and
what remains open are in the retraction notice
published against it. Its text is not quoted forward, here or anywhere.
selection · point-in-time · withdrawn
2026-06 · working paper · revised 2026-07
A relationship with an obligated enforcer (a
prospectus rebalance mandate, a defended currency band, share-class identity,
contract convergence, an index mandate) survives specificity controls;
co-movement without one does not. Revised after a venue-authenticity audit
withdrew one class of supporting evidence, leaving the principle resting on
the remainder and the audit itself as the paper's strongest exhibit.
Stands on cointegration ·
microstructure ·
regime
mechanism · specificity · enforcement · controls
Corrections and retractions
Three items, kept here instead of on a page of their own: three
corrections do not fill a page, and a near-empty corrections page reads as a
claim rather than a record. Each entry names the item it acts on by title and
date, and states what was published, what was wrong, and what supersedes
it.
2026-08 · correction · corrects the funnel figure on the
cover
“Refuted” overstated what we had actually tested
Published. A funnel figure describing the whole
archived strategy mass as refuted.
Wrong. Only the minority of strategies that
accumulated enough independent test windows ever faced the specificity gate.
The remainder were abandoned earlier (at authoring validation, at a build,
or on an activity floor) and never met a null at all. Abandoned before test
and refuted by test are different claims, and the second flatters us.
Supersedes. The figure now separates authored,
gate-tested and survived as three distinct counts, and the archive language
distinguishes abandonment from refutation. The counts themselves live on the
index, which is their home of record.
2026-08 · correction · corrects the exchange-breadth
figures on coverage and the
cover
Two different breadths were published under one name
Published. Two different counts of exchange
breadth, on two pages, presented as though they were the same quantity.
Wrong. Not the numbers but the naming. One counts
what our vendor catalogue offers; the other counts what has been ingested
and is read by live research. Both are true and they are not the same thing. A
reader had no way to know that, and the discrepancy read as sloppiness
rather than as the distinction it actually was.
Supersedes. Coverage
now names both quantities and prints each against the thing it counts.
Coverage is the home of record for estate figures; other pages link to it
instead of restating them. Restatement is what let the two drift apart in
the first place.
2026-08 · retraction · withdraws
Ranking for the upside tail, 2026-07
Retraction of “Ranking for the upside tail”, in whole
Published. Ranking for the
upside tail, a working paper of 2026-07 proposing a
book-construction rule, whose conclusions were already circulating
internally and had begun to shape how books were assembled.
Wrong. The corpus underneath it was optimised
at roughly a fifth of the production trial budget, and the shortfall tracked
strategy complexity, so the strategies most likely to be under-cooked were
exactly the ones the paper's ranking treated as weak. Under-optimisation and
genuine weakness are not separable in that sample, which makes the sample
invalid for the question the paper asked. Every conclusion goes, including
the ones we liked; a retraction that keeps the flattering half is not a
retraction.
Supersedes. Nothing, yet. The campaign is being
redone at the production budget. The corpus is archived under a distinct
name, not deleted, so the redo can be compared against it, and the
paper stays listed in the index with status retracted
so that anyone who read it internally can find out that it is gone.
The rule that caught it. The confound was named
in a companion paper's limitations section before anyone acted on it. The
standing rule now is that a confound named in a limitations section is a
scheduled experiment, not a disclaimer: it goes on the register with a date,
or it does not go in the limitations section.
Standing consequence · re-audit doctrine
Re-audit from the artefacts, not from the records
Both corrections and the retraction have the same shape: a written record
that had never been checked against the thing it described. So the rule.
Every claim carried forward from a closed campaign is re-derived from the
stored run artefacts or re-run from the pinned script, never quoted from our
own write-up. Where a record and its artefact disagree, the artefact wins and
the disagreement is itself a finding. Where an artefact cannot be located or
re-run, the claim is void, not probably fine.
The probe that set the budget for the most recent review was a single spot
check on a one-day-old record: the campaign document matched its artefact,
and the journal entry written the same day disagreed with both on four of the
values it shared with them. One day old, one probe, one discrepancy. That is
the base rate a review has to plan for, and it is why the review is budgeted
in weeks, not in an afternoon.
The self-critique belongs here too, because it is the obvious way for this
to fail. A re-audit that re-runs the same scripts against the same store and
reports agreement has confirmed determinism, not correctness. Testing the
instrument rather than the pipeline means planting an effect the instrument
should certainly find and confirming it finds it, then planting nothing at
all and confirming it finds nothing. Only the second pair of probes says
anything about whether the first result meant something.
Reading the library
Every item published with its own text declares two things at its foot.
Stands on points down
into the reading list, at the shelf the item drew
from, so a claim about prior art can be checked against the shelf, not
taken on trust. Cited by points across to the other items in this
library that build on it, naming each by its title. Between them, the
citations are what make this a body of work
rather than a pile of documents, and they are the reason a wrong item is
corrected by a new item, not edited: an edit silently changes what
every citation pointed at.
These cross-references are hand-maintained. There is no build step on this
site and therefore no generated backlinks and no build-time citation check; a
link check is run before publish and that is the whole of the machinery.
Claiming more than that, on the page that argues for fail-loud engineering,
is the one self-indulgence a reader would be right to catch.
Redaction policy
The protectable property in this shop is the calibration and the edges, not
the methods. The methods are published science and we publish them. These
classes of number do not appear anywhere on this site, in any item, in any
figure caption:
- Calibration values of any kind: score floors, cut-offs, tuned
thresholds, and the weights inside our own composites.
- Detection floors in absolute units.
- Discrimination values for our own instruments. The ordering is
published; the value is not.
- Per-book returns, Sharpe ratios, or drawdowns.
- Capacity in dollars.
- Leverage settings.
- Search-width configuration counts.
- The ranking of which statistics predict survival.
- Named live edges, instruments, pairs or venues we currently trade, and
the market lists we screen. A refuted result is safe to publish; the set of
markets it was screened on is not, because across a page a market list is
itself a disclosure of where we look.
The substitution rule is what we publish instead: the ordering, the policy,
the n, and the direction. Then the withheld value is declared as withheld, in
the same sentence, every time, so a reader can grep this site for the phrase
and audit that every proprietary claim announced itself.
A declared redaction is stronger than an omission. An omission tells a
reader nothing. A declared redaction tells them the number exists, that we
hold it, and that we chose not to print it, which is a claim they can hold us
to later.
Underneath all of it, a journal
The library sits on an append-only research journal. Substantive questions
and findings get a dated entry, chronologically. A prior entry is never edited
to fix it; a superseding entry is written instead, and the trail stays
reconstructable by someone who was not there.
Each entry records the surprises, the assumptions that turned out to be
wrong, and the things a reviewer should sanity-check: leakage risks, slices
too small to carry the claim, framing choices that would flatter us if
unexamined. That convention is the reason the discrepancy in the retraction's
companion probe is visible at all: the journal entry that was wrong is still
there, still wrong, with the entry that corrects it underneath.