Section 01 / 03
Research and writing
26 pieces where the argument is the point. Every one of them started because I did not believe a claim, or could not find two jurisdictions held apart properly, and the fastest way to find out was to build the thing. 268,124 words, none of them assigned. The order is mine; the figures are the build's.
The independent shelf
The statement, filtered to the work nobody assigned. Each row carries the piece's own declared rule, verbatim from the record, and the figures lifted from four of them.
26 of 69-
01
Flagged in Hindsight
Five Canadian frauds, put back through the forensic screen ex ante
An interactive bench rebuilds each score from the filed figures and lets the reader drag the cut-off to watch sensitivity buy false positives.
5,00846 -
02
Crucible Cockpit
Adversarial claim auditing: decompose, attack, grade what survives
The backtest run replays a question whose answer is now known with evidence locked at the cutoff, and returns UNDETERMINED: the leak test the live runs stand on.
61202 -
03
Crucible Run 0: the Bank of Canada backtest
Evidence locked at January 2023; the engine says what was knowable
No declared rule on file.
18,59700 -
04
Crucible Run B: the maturity-wall thesis
The 2026 coupon step-up claim, audited and refuted
No declared rule on file.
60,15300 -
05
Crucible Run C: the goodwill-impairment thesis
The too-little-too-late critique, audited against TSX issuers
No declared rule on file.
53,06600 -
06
Epistemic Quality Directive
The evidentiary standard distilled from the claim-audit engine, in four modes
No declared rule on file.
2,54900 -
07
The Brittle Network
A pre-registered protocol on the Late Bronze Age collapse
Colour is used as a register throughout: each evidence stream (metals, network, textual, failure) carries its own colour, and nothing else does.
9,237416 -
08
The Price of Divergence
A pre-registered protocol on ESG-rating disagreement and portfolios
No declared rule on file.
11,9644924 -
09
Morphogenesis
A laboratory of emergent pattern, computed live
Pattern that nobody draws: every stripe, spiral and flock is computed from local rules as you watch.
48400 -
10
IFRS 15 Judgment Trainer
Sixty-two revenue scenarios, every answer anchored to the Handbook
The bank distinguishes what the standard settles from what it leaves to judgment, and scores only the former: a trainer that scores nothing teaches nothing, and one that scores judgment calls teaches something false.
3,94601 -
11
IFRS 15 Item Bank: Verification and Backfill
The verification memo behind the trainer
No declared rule on file.
2,98603 -
12
The Five Shall Paragraphs
Materiality and sampling under the CAS, requirements separated from habit
No declared rule on file.
19,112013 -
13
Zakaria's Putin, Audited
A 2022 documentary, claim by claim, against the 2026 record
No declared rule on file.
4,14500 -
14
Predictive History
The map, and the method that audits it
The instrument reports the result that undermines its own headline: panel-to-author agreement is 78 per cent, which is 22 per cent divergence against a 20 per cent withdrawal threshold the project set for itself before it knew the answer. The finding is published in its narrow form rather than its strong one, and 4 of 9 adversarial mutations are documented as still uncaught.
An interval that overlaps is not a difference
Predictive History
The obvious reading is that writing a verdict rubric lifted agreement from 53 per cent to 96. It did not. Only the second and third rows hold the record constant, and between those two the rubric is worth four points with the intervals overlapping. The other forty-three points are the record and the number of raters, which is a different claim entirely.
a condition that keeps a written record
Open the pieceThe numbers this figure draws
0% · 25% · 50% · 75% · 100% · 1 rater, no record, no rubric · 53% · 32–73 · n=19 · 2 raters, record, no rubric · 92% · 85–97 · n=80 · 3 raters, record, rubric · 96% · 90–99 · n=80 · Rubric effect, record held constant: +4 points, and the intervals overlap.
9,6211114 -
15
Predictive History: the withdrawn first edition
Kept uncorrected, with the divergence computed rather than described
A project about the distance between confident presentation and the record, showing its own. Nothing below the banner is corrected, and both columns of the comparison are computed at build time, the left by parsing this frozen file and the right from the current ledger.
11,6971719 -
16
The Trillion-Dollar Vintage
A study edition of the AI capital-cycle adjudication
The largest essay here (Crucible Run B is the largest piece on the site), and the one that shows the method at full size. Two rival readings of the same capital cycle are carried side by side to the end rather than averaged into one, and the page corrects its own arithmetic in public where a later attribution changed the base.
The fork is never closed
The Trillion-Dollar Vintage
Two ways of pricing the same vintage, 2.3 times apart, carried side by side to the end rather than averaged. The refused marker between them is the point: no instrument measured that value, so nothing is drawn there.
Anchor A, realised, from a filerAnchor B, constructed, from market prices
Open the pieceThe numbers this figure draws
weight 0.40 · weight 0.60 · designated base case · $8.99 M revenue per active megawatt · annualised mean of 9.25 / 8.73 / 8.99 across three quarters, · $50 M all-in build cost per megawatt · 18.0% gross revenue yield per year · $39.9 M all-in cost per megawatt · a $3.0 M GB200 NVL72 rack [S] drawing 132 kW gives · $22.7 M of accelerator per MW, over the 57% weight [I] · $16.7 M revenue per megawatt per year · $3.50 per GPU-hour [S] rented around the clock, · 8,760 hours a year [I] · 41.9% gross revenue yield per year · 0% · 10% · 20% · 30% · 40% · A · 18.0% · B · 41.9% · 2.3× apart
Figure lifted from the piece. One vintage on two bases: 1,082 billion dollars as first reported, 990 billion after an attribution correction the page makes in public. The fork stays open.
The numbers this figure draws
band $931 to $1,232 bn · $1,082 bn · band $822 to $1,159 bn · $990 bn · −$92 bn, −8.5%: the D-1 defect, found in the analysis’s own evidence pack. · Effect on the verdict: failure share 75.6% to 75.7%, probability unchanged at 0.743. · $303 to $433 bn · 28–40% of V · Solid: Alphabet $122.8 bn and Oracle $40.0 bn, the only two firms that disclose it. Open: modelled, not measured. · 0 · 250 · 500 · 750 · 1,000 · 1,250 · US$ billions of incurred capital, 2023 to mid-2026 · Owner classes run light to dark by unlevered hurdle, 8.13% to 16.97%.
18,6101615 -
17
Whose Losses Count
Corporate ethics and the micro/macro foundations of strategy
An accounting question asked as an ethics question. Colour means one thing on every figure: blue is inside the accounting boundary and priced, red is outside it. A result that is not distinguishable from zero is drawn as an open outline at full size rather than shrunk to nothing.
Blue is inside the number, red is outside
Whose Losses Count
Seven literatures making the same boundary decision, drawn on one vertical rule. Whatever falls to the right of it is real and appears on no ledger. Open outlines are results not distinguishable from zero, drawn at full size rather than shrunk away.
inside the numberoutside it, on no ledger
Open the pieceThe numbers this figure draws
0.8–5.6× value added · USD 190 a tonne · ~90% US firms and consumers · gross payout, 91% · 50% · counterparties, 7.5× · strangers, +0.24%,
6,96027 -
18
Not Significant
An essay on the accounting of harm
The whole essay turns on a negative result, and reports it as one. Solid marks are quantities that reach someone's books; open outlines at full size are quantities that are real and appear on no invoice.
Solid reaches a ledger, open does not
Not Significant
Deceive an investor and the market takes 7.53 times what the law does. Injure a stranger who does not trade with you and the share price moves 0.24 per cent, which is not significant. The essay is built on that gap.
counted: a loss that reaches a ledgerpriced by the market, on no ledgerharm that reaches no ledger and no price
Open the pieceThe numbers this figure draws
Misrepresentation · loss per $1 of value inflated · $0 · $1 · $2 · $3 · $0.36 · $2.71 · 7.53× · 0% · 1% · 2% · 3% · +0.24%
6,319101 -
19
Global Spending and Wealth
Scale, structure, concentration
A unit declared once and held for the whole document, so quantities stay comparable by eye across every figure. Two sources that disagree are shown side by side instead of averaged.
6,521137 -
20
Does chronic caffeine spend your heartbeat budget faster?
A question taken seriously
A claim I did not believe, checked properly, and reported at the strength the evidence actually supports.
1,43455 -
21
Exam Day Final Load
The last hour
A method piece: the rule for what gets cut, written down before the pressure arrives.
47670 -
22
Daily Learning Cockpit
Prediction log, spaced rotation, calibration
Turns the research on durable learning into something with a button on it.
25100 -
23
The Delayed Test
Learning, judgment, and seeing clearly
Every claim carries the evidence grade its source file gave it, unaltered. Where a source marked a figure second-hand or contested, that flag travels with the figure into this page rather than being quietly dropped.
5,37700 -
1,87606
-
1,34203
-
26
The tax base nobody discloses
A Dollarama valuation built on capital cost allowance
Two things are tested rather than asserted. The reconstructed tax base is run forward over a year it was not fitted to and asked to reproduce the current tax expense the issuer disclosed, and it lands 5.5 per cent short, which the page reports as a miss rather than tuning an input to close it. The beta is regressed from five years of weekly returns instead of assumed, correcting an earlier draft that had asserted a figure outside the interval the data supports. The valuation range is two anchors with no midpoint, because the disclosure that would justify a midpoint does not exist.
5,78154 - Independent, 26 pieces268,124143146