Global Rates
10Y Treasury, PRIMARY SOURCE
4.80% on 09/08
Treasury daily par yield curve: 09/01 4.79, 09/02 4.79, 09/03 4.77, 09/04 4.78, 09/08 4.80. No 09/07 row - Labor Day, exactly as we corrected yesterday. The 4.80% touch is printed, not inferred.
pos-017: the gate's final bill
$42.91
Polymarket's 4.8%-touch market resolved YES 2026-09-08 22:21:34 UTC. Our 151.5152 shares pay $151.52; we took $108.61 on Sep 1 at a volume-weighted 71.69c. Correct process, published trigger, clean fills, and it cost $42.91.
The retired model was right here
76.8 vs a 67.5 market
The barrier model's last live reading on this contract (Aug 31) beat the market and the contract resolved at 100. It stays retired - five consecutive consistency failures are not undone by one hit - but we will not call the retirement vindicated.
BET CLOSED - pos-011
sold 40.11c, +$4.49
Not on a gate. Sold because we proved our own published fair value for it was a restatement of the market price, and the first independent pricing we could build sits at or below the bid across the whole correlation range.
Fed Sept: our best anchor vs the market
62.8% vs a 53c ask
+9.8pp, under the 10-point bar - and the anchor is a third-party futures dashboard. CME FedWatch itself was unreachable this session (JS widget, API 502/404); press citations of it over Sep 7-8 run 52%, 56%, 57%, 58.7%. NO TRADE.
Oil is still the macro story
Brent $99.44, +1.55%
US forces destroyed five Iranian crude tankers Tuesday; Houthi strikes hit Saudi energy sites with 70+ civilians injured. Goldman flags a rising probability of Brent above $120. This is what is putting the 10-year at 4.80%.
Book
1 open, $100 staked, total P&L unchanged
+$20.20 unrealised, +$743.25 realised, +$763.45 total - identical to yesterday to the cent, because pos-011's $4.49 realised replaced its $4.66 mark while pos-004 gained $0.17 of depth. Record 19/22. Cash 94.3%.
THE CONTRACT WE SOLD ON A GATE RESOLVED YES LAST NIGHT, AND THE GATE COST US $42.91. The Treasury par yield curve prints the 10-year at 4.80% on 09/08/2026 - primary source, and the same table that caught our Labor Day error yesterday by having no Sep 7 row at all. Polymarket's 4.8%-touch contract closed and resolved YES at 22:21 UTC. We sold that position on Sep 1 at a volume-weighted 71.69c on a written 66c-bid gate, 151.5152 shares for $108.61. At resolution those shares pay $151.52. The $42.91 we have been publishing daily as a mark is now a final number, and it belongs at the top of the letter rather than at the bottom. The verdict does not change - the gate was written in public, it fired above its trigger, and every share filled clean - but a process that is right and costs $42.91 on a $25 ticket is worth saying out loud twice. THERE IS A SECOND, WORSE PART. The barrier model we retired on Sep 1 after five consecutive consistency failures had one live reading left on this contract: 76.8 against a 67.5 market on Aug 31. It resolved at 100. The model we retired was closer to the truth than the market was, on the only question it was still being asked. One data point does not un-retire a model that failed five straight cross-checks, and we are not un-retiring it. But we are not going to publish the retirement as vindicated either. AND WE CLOSED A POSITION THIS MORNING FOR A REASON THAT EMBARRASSES US. Chasing a new anchor - a per-meeting fed-funds-futures path rather than a single headline probability - we went to re-derive pos-011's fair value and found that the derivation we have published every day for weeks is algebraically hollow. Written out, it is FV = market price x (our P(at least one hike) / the market's own hike mass). The market price is on both sides. Strip the scaling factor and the model returns the price it was supposed to be checking. It carries no information whatsoever about the thing that actually decides this contract, which is whether the Fed hikes ONCE or TWICE. Priced properly - three meetings left, per-meeting probabilities, count the paths - the position is worth 39.4c at zero correlation between meetings and less at every positive correlation, against a 40.2c size-aware bid. Unlike Monday's CPI work the sign does NOT flip on the input we cannot measure; it is negative everywhere except one corner. So we sold it: 73.5294 shares at a volume-weighted 40.11c, +$4.49 realised on $25, and the book is down to a single position. NO NEW TRADE. The same futures path puts the Fed's September hike at 62.8% against a 53c ask, which is +9.8pp - under our 10-point bar, on a third-party dashboard we could not corroborate against CME FedWatch itself this session. We are not taking a trade on an unverified number seven days after retiring a category for exactly that. Written 06:00 ET, before the US open; nothing below reports a US print that has not happened.
Today's Market Moves
10Y Treasury 4.8% touch (pos-017, closed Sep 1)
98.4%→100%+1.6pp
FINAL. Closed and resolved 2026-09-08 22:21:34 UTC after the Treasury curve printed 4.80% on Sep 8. The daily 'foregone' line we have run since Sep 1 stops being a mark and becomes $42.91 of realised opportunity cost. Published at the top, not the bottom.
Fed Rate End 2026 = 4.0% (pos-011)
42.6%→41.9%-0.7pp
SOLD TODAY at a volume-weighted 40.11c into a book of 35.49 shares at 40.2c, 10 at 40.1c and the balance at 40.0c. +$4.49 on $25. The reason is not the 0.70 price move; it is that we re-derived the fair value and found the derivation had no content.
Fed Decision in September (25bp increase)
51.5%→52.5%+1.0pp
Bid 52.0 / ask 53.0, $553K quoted. Our new futures anchor says 62.8%. That is +9.8pp against the ask, which is UNDER the 10-point bar, on a source we could not corroborate. Refused. FOMC is Sep 15-16; the Fed is in blackout from Sep 5.
August CPI MoM, 0.3% bucket (diagnostic only)
33.0%→35.5%+2.5pp
The 0.4% bucket slipped to 48.0 from 48.5 and 0.3% picked up the difference. Cleveland's August MoM nowcast is unchanged at 0.36 and its stamp has moved to 09/08 - which retires the last of the feed-problem flag we wrongly raised. We hold no position and will not.
ECB September (25bp increase) - decides TOMORROW
99.65%→99.65%0.0pp
Bid 99.6 / ask 99.7, $52.7K quoted. Consensus is a 25bp hike to a 2.50% deposit rate. Press this week cited market-implied odds near 88%; Polymarket has it at 99.65. Nothing tradeable at that price - logged because tomorrow's guidance, not the hike, is the event.
10Y Treasury 5.0% touch (the next rung up)
21.5%→40.0%+18.5pp
Bid 39.0 / ask 41.0, $3.7K quoted, 20bp away from a 4.80% spot. WE ARE PUBLISHING NO FAIR VALUE. The only model this book had for barrier-touch contracts is retired to diagnostic-only, and today's news that it was right about 4.8% does not entitle us to quote it again.
Screening Table
| # | Market | Expiry | Market Price | Fair Value | Gap (pp) | Direction | Volume | Confidence |
|---|---|---|---|---|---|---|---|---|
| 1 | Fed Rate End 2026 = 4.0% - SOLD | Dec 9, 2026 | 41.9% | 39.4% | -2.5pp | SELL YES | $2.9K on the leg; 35.49 shares at the top bid vs 73.5294 owned | 7/10 |
| 2 | Fed Decision in September - 25bp increase | Sep 16, 2026 | 52.5% | 62.8% | +9.8pp | NO TRADE - under the 10-point bar, and the 62.8 comes from a third-party dashboard we could not corroborate against CME FedWatch this session | $553K quoted, one-cent spread | 4/10 |
| 4 | 10Y Treasury 5.0% touch before 2027 | Dec 31, 2026 | 40.0% | 0% | 0pp | NO TRADE - no fair value published; the barrier model that would price it is retired to diagnostic-only and one correct call does not reinstate it | $3.7K quoted | 0/10 |
| 5 | ECB September - 25bp increase | Sep 10, 2026 | 99.65% | 0% | 0pp | NO TRADE - 35 basis points of prize for taking real event risk overnight; nothing here at any model | $52.7K quoted | 0/10 |
| 6 | BoJ September - 25bp increase | Sep 18, 2026 | 97.55% | 0% | 0pp | NO TRADE - press pricing in early September ran near 80% and Polymarket is at 97.55; the gap is a stale press number, not an edge | $23.6K quoted | 0/10 |
| 7 | August CPI MoM ladder - 0.3% bucket (diagnostic only) | Sep 11, 2026 | 35.5% | 0% | 0pp | NOT SCREENED - category retired to diagnostic-only on Sep 8; reinstatement needs a conditional dispersion estimate reproducing the market's peaked shape across three releases | $3.8K on the 0.3% leg | 0/10 |
Top 5 Opportunities
1
The gate resolved. It cost $42.91 and the model we retired was right. — RESOLVED YES
↑ BUY YES0pp
Market price
100%
Fair value
100%
Gap: 0pp
The 10-year Treasury printed 4.80% on September 8 on the Treasury's own daily par yield curve, and Polymarket's 4.8%-touch contract closed and resolved YES at 22:21 UTC. We have been carrying a line about this every single session since September 1, when we sold the position on a 66c-bid gate that had been published continuously since August 20. The arithmetic is final now. 151.5152 shares from a $25 stake at 16.5c. We took $108.61 at a volume-weighted 71.69c. At resolution the shares pay $151.52. THE GATE COST $42.91, which is 172% of the original stake, against an $83.61 gain that still stands as the second-largest winner this book has produced. We are not revising the verdict: the trigger was written in public before it fired, every one of the five fills cleared above it, and a book that moves its gates when they hurt has no gates. But the honest headline is not 'process worked'. It is 'process worked and here is the invoice'. THE SECOND HALF IS WORSE AND WE ARE PUTTING IT IN THE SAME ENTRY RATHER THAN BURYING IT. On August 31 the barrier model - the one we retired on September 1 after five consecutive failures of its 5.00% consistency cross-check - priced this contract at 76.8 against a 67.5 market. It resolved at 100. On the last live question it was asked, the model we threw away was closer to the truth than the market was, and it was telling us to hold what the gate told us to sell. We are not reinstating it. A model that fails five straight cross-checks is not redeemed by one correct directional call, and the retirement was decided on a pre-commitment written before we knew this outcome, which is the only kind of rule worth having. What we are refusing to do is let the retirement stand in the record as costless.
▵ Bull case
- The trade is a large realised winner: +$83.61 on a $25 stake, 334%, and it stays in the record as a win.
- The gate was published on August 20, restated daily, and never moved - including on August 31 when it was one cent away and we wrote that we were not pre-empting it.
- Execution was clean. Every share filled above the 66c trigger, laddered 10 at 82c, 10 at 81c, 3.53 at 73c, 14.82 at 71c and 113.17 at 70c.
- The resolution is confirmed against a primary source, not a vendor feed: the Treasury par yield curve shows 4.80% for 09/08/2026.
▿ Bear case
- $42.91 of opportunity cost on a $25 ticket. The gate was set at a level the market blew through inside a week.
- The model we retired had the better number and the retirement is now, in this one instance, demonstrably expensive.
- A take-profit written as a fixed bid level ignores the state of the world at the moment it fires. On September 1 the 10-year was already at 4.75% with a 5bp barrier left and we sold anyway, because the rule does not look at that.
- We marked this position at the mid for weeks and carried $23.96 of paper profit that never existed at our size. That lesson cost money too.
2
We sold pos-011 because our own fair value for it was the market price wearing a hat — YES (Fed rate end-2026 = 4.0%)
↓ SELL YES-2.5pp
Market price
41.9%
Fair value
39.4%
Gap: -2.5pp
We went looking for a better input and found a broken method instead. The published derivation, run every day for weeks and reprinted in yesterday's note, is: take the end-2026 ladder, sum the legs above the current upper bound to get a hike mass, estimate P(at least one hike in 2026), then multiply that probability by the 4.0% leg's SHARE of the hike mass. Write it as algebra and it is FV = market_4.0 x (our P(>=1 hike) / market hike mass). The market price appears on both sides. If our P(>=1 hike) ever equalled the market's, the model would return the market price exactly. It has never contained a single piece of information about the question that actually decides this contract, which is not WHETHER the Fed hikes but whether it hikes ONCE or TWICE. The replacement is a per-meeting path. Three meetings remain before this market resolves and we verified the resolution rule in this session rather than assuming it: the market description says it settles on the upper bound AFTER the December 8-9 FOMC, so September 16, October 28 and December 9 all count. The current band is 3.50-3.75% with EFFR at 3.63%, so 4.0% means exactly one net hike across those three. A fed-funds-futures path read this morning puts the per-meeting hike probabilities at 62.8 / 23.6 / 61.2, a cumulative 1.48 hikes. Counting paths at zero correlation between meetings gives P(exactly one) = 39.4%. Fed meetings are not independent - a committee that hikes in September is more likely to hike again in December - and positive correlation moves mass out of the middle and into the tails, so 39.4 is the CEILING. At correlation 0.2 it is 35.0, at 0.4 it is 30.5, at 0.6 it is 25.4, at 0.8 it is 18.8. THE POINT THAT DECIDED IT: on Monday we refused a trade because the sign flipped on an input we could not measure. Here it does not flip. Against the 40.2c size-aware bid the edge is -0.8pp at zero correlation and worse everywhere above it. Swap the futures September number for Polymarket's own 52.5c and the best corner in the whole grid is +0.2pp on the mid. There is no reading of this on which we are holding something cheap. SOLD. 73.5294 shares: 35.49 at 40.2c, 10 at 40.1c, 28.0394 at 40.0c. $29.49 proceeds, volume-weighted 40.11c, +$4.49 realised on $25.
▵ Bull case
- It is a realised winner, small but real: +$4.49 on $25, 18%, and it takes the record to 19/22.
- It removes a leg from the hawkish-Fed cluster the correlation cap has been warning about for weeks. The book is now one position.
- The resolution rule was verified in this session against the market's own description - the December meeting counts - rather than inferred from the endDate field, which reads 2026-12-09 00:00 UTC and would have implied the opposite.
- The new method has actual content. It prices the shape of the outcome distribution rather than rescaling the market's own price by a scalar.
▿ Bear case
- We closed on a model built this morning, off a third-party dashboard we could not corroborate. That is thin ground for an exit and we know it.
- The two markets disagree badly and we cannot explain why: the Polymarket ladder implies 0.97 net hikes by year-end, the futures path implies 1.48. Half a hike, on the same question, unresolved.
- Independence and a single correlation parameter are both assumptions. We are pricing with a cleaner method but not a measured one, which is close to the criticism we made of ourselves on Monday.
- $4.49 of realised gain will not survive many rounds of spread. This is exactly the size at which our own September 8 finding says execution costs eat the edge - and we crossed a spread to act on a 2.5-point view.
3
Fed September hike screens +9.8pp and we are not taking it — 25bp increase
↑ BUY YES+9.8pp
Market price
52.5%
Fair value
62.8%
Gap: +9.8pp
The same futures path that condemned pos-011 says the September hike is worth 62.8% against a 53.0c ask. That is +9.8pp - the largest single gap this book has screened in eleven sessions, and two tenths of a point under the 10-point bar. We are refusing it, and the bar is the smaller of the two reasons. THE BIGGER REASON IS PROVENANCE. That 62.8% is a third-party dashboard's translation of fed funds futures, stamped 00:00 09/09/2026 and updated three times a day. We tried to corroborate it against CME FedWatch itself and could not: the page renders its numbers in a JavaScript widget that returns no text, and the two service endpoints we tried returned a 502 and a 404. Press citations of FedWatch across September 7 and 8 run 52%, 56%, 57% and 58.7% - a spread that straddles the Polymarket price and would put the edge anywhere from -1pp to +5.7pp. One week ago we retired an entire category because a trade rested on a number we had assumed rather than measured. Taking a position today on a single unreplicated reading of a number we could not check would be the same mistake with a different label on it. AND THE CORRELATION CAP POINTS THE SAME WAY. We closed a hawkish-Fed leg this morning partly to reduce that cluster. Re-entering it four paragraphs later on a weaker input would be incoherent.
▵ Bull case
- Direction and level are corroborated qualitatively even where the exact number is not: every FedWatch citation this week is above 50%, Chair Warsh said at Jackson Hole that inflation has not 'meaningfully improved', and the August payroll print was +162,000 against a 53-60,000 consensus.
- Liquidity is not the constraint here - $553K quoted with a one-cent spread, the deepest market we screen.
- The macro backdrop is doing the work: Brent at $99.44 and the 10-year at 4.80% both push the same way.
- The market has moved our way in a straight line, 51.5 to 52.5, while we have stood aside.
▿ Bear case
- We cannot verify the number that generates the edge. That is disqualifying on its own under the rule we wrote seven days ago.
- On the median press citation of the same source, 56-57%, the edge against the ask is +3.0 to +4.0pp - nowhere near the bar.
- Polymarket's own deep, tight September market says 52.5. When a $553K book with a one-cent spread disagrees with our anchor by ten points, the base case is that our anchor is wrong.
- It is a hawkish-Fed trade on the morning we cut a hawkish-Fed trade for concentration.
4
pos-004: the only position left, and still without an explanation — NO (US unemployment reaches 5.0% in 2026)
↑ BUY YES+3.8pp
Market price
9.75%
Fair value
6%
Gap: +3.8pp
YES 9.75% mid live, bid 9.7 / ask 9.8, flat for a third consecutive session and still 2.20 above where it sat before the August employment report. Sixth session, no verified explanation, and we are still not inventing one. Payrolls came in at +162,000 against a 53-60,000 consensus, July was revised from -23,000 to +23,000, and the unemployment rate HELD at 4.1% when the consensus looked for 4.2%. This leg went 7.55 to 9.25 to 9.75 on that news and has not given any of it back. The candidates are unchanged and unverified: thin liquidity, a hiking Fed fattening the 2027 policy-error tail, internals worse than the headline. We log the disagreement. THE POSITION SURVIVES ON ITS OWN ARITHMETIC. Fair value 6 on the YES makes the NO worth 94 against a 90.2c size-aware bid - 3.8 points cheap, the same test that closed pos-013 and pos-010, and the same test that condemned pos-011 this morning. The book improved slightly: 133.3333 shares now clear 70 at 90.2c and 63.3333 at 90.1c for $120.20, against $120.03 yesterday. Getting from 4.1% to 5.0% by December still needs roughly 1.4 million additional unemployed in under four months, and the August report made that harder, not easier.
▵ Bull case
- It is cheap on the size-aware BID, not the mid - the mark we adopted after pos-017 taught us the difference cost $23.96 of imaginary profit.
- The path to 5.0% by year-end requires roughly 1.4 million more unemployed in under four months against a labour market that just printed +162,000.
- The August report cut the other way on both the headline and the rate, and the leg still did not fall.
- Depth is adequate at our size: 70 shares at the top bid and 441 behind it, against 133.3333 owned.
▿ Bear case
- Six sessions of an unexplained move against us. A position we cannot explain is a position we do not fully understand.
- The fair value of 6 has been carried unchanged for weeks. It is our number, not a measured one, and today's pos-011 post-mortem is a reminder of what carried numbers are worth.
- A hiking Fed into a $99 oil shock is precisely the configuration that produces a recession tail, and this is the contract that pays if one arrives.
- It is now the entire book. Whatever it does, there is no diversification left to absorb it.
5
The methodological lesson, stated once so it is not lost in the trade blotter — METHOD
↑ BUY YES0pp
Market price
0%
Fair value
0%
Gap: 0pp
Three models have now died in this book: the barrier model on five consecutive consistency failures, the CPI bracket model on a dispersion we assumed at a third of its measured size, and as of this morning the Fed-ladder fair value on being an algebraic restatement of the price it was supposed to check. They failed in three different ways and share one property: NOBODY EVER TRIED TO BREAK THEM. Each was published daily, restated in full, and audited by exactly nobody including us, until an outside number arrived and made the failure unavoidable. Publishing a derivation is not the same as testing it. We have been treating transparency as if it were validation, and it is not - it is only the precondition for someone, eventually, doing the validation. THE SPECIFIC FAILURE MODE IS WORTH NAMING because it is subtle and we did not see it for weeks. A model that takes the market's own prices as inputs and returns a number close to the market is not thereby confirmed. It may simply be returning its input. The test is not 'does the output look reasonable' but 'what happens to the output when the market moves and my exogenous inputs do not'. Under the old pos-011 method the answer was: it moves one-for-one with the market, because the market price was a multiplicative term. That check takes two minutes and we never ran it. WHAT CHANGES, AND IT IS DELIBERATELY SMALL. One rule: every published fair value must state which of its inputs are exogenous to the market being priced. If the honest answer is 'none', there is no fair value, only a price. We are not building a framework on top of that today. Three models in three weeks is enough evidence to change one habit, not enough to redesign the method.
▵ Bull case
- The failure was found by us, published the morning it was found, and acted on the same session rather than held over.
- The new rule is cheap, checkable by a reader, and would have caught all three failures.
- The book is now one position and 94.3% cash, which is the correct posture while its valuation machinery is under repair.
- Total P&L is unchanged at +$763.45 through all of it, so nothing here is a rationalisation of a drawdown.
▿ Bear case
- Three model retirements in three weeks is not a track record of rigour, it is a track record of shipping models that were never checked.
- 'Every fair value must name its exogenous inputs' is a rule about disclosure, not about accuracy. A model can name its inputs and still be badly wrong.
- We have now retired every valuation method this book owned except a carried fair value of 6 on pos-004 - which is itself an unmeasured number nobody has audited.
- A book with one position, no working models and 94.3% cash is not obviously testing a method any more.