
The Treasury's Tell
Is the US Treasury pushing long-term interest rates down? Nobody has said it may. Who makes it admit the aim, and what does admitting it cost?
Report 005 · Dr Dan Epstein · v1.5 · 2026-11-11
A Strategy Soup ScenarioWhat is an exercised scenario?
The question
The Treasury is the US government's finance department. It borrows what the government spends by selling bonds: IOUs that pay interest and are repaid on a set date. The Federal Reserve, the Fed, is the US central bank, and it sets the short-term interest rate. Washington has agreed that the Treasury, not the Fed, decides the maturity structure of the public debt: the mix of short-dated and long-dated bonds the government owes. We call that agreement the doctrine. Nobody has agreed whether the Treasury may use that power to push long-term interest rates down. Call that the price question. A bond's price and its yield, the interest rate it pays, move in opposite directions, so holding a yield down means holding a price up, and we use both words for the same thing. Who settles the price question, and what will 4 November show? On 4 November, a real and scheduled date, the Treasury publishes its quarterly refunding statement: its announcement of how much debt it will sell over the next three months, and why. Its own published test for cutting bond sales is trends in structural demand, meaning a lasting shift in who wants to buy which bonds . So if it cuts, it will give a reason. In this simulation, what the Treasury decides on 4 November is which reason to print.
Dates. This version is dated 19 October 2026. The real-world record behind it stops on 3 September 2026. Every later event in it is the simulation's story, not a forecast. Some of those dates have now passed. Apart from the Fed's decision of 16 September and the Treasury's buyback schedule as at 17 September, both from the start record of a second run and outside the evidence pack, this version does not report what happened on them.
The answer. Our starting thesis: nobody at the Fed or the Treasury has an incentive to ask the price question out loud, so it gets decided without ever being asked. The runs amended that. In the 30 runs where the Treasury spent no extra money, speeches and chance together forced the price question into the open inside the eight weeks from 9 September to 4 November 2026 in 4 runs, 2 in each plan. In the plan where the Treasury spent an extra $40bn, the question opened in 10 of its 15 runs: six times in the turn of the spending, four times in the turn after. Every opening followed the spending, which does not prove the spending caused it. The spending made the opening far more likely: 10 of 15 runs, against 2 of 15 in each plan without it. In 7 of those 10, two or more players' moves pushed the question open in the same turn. On 4 November, admitting the aim cost the Treasury 18 points of trust, our score for how far the market trusts its word, against 9 for citing structural demand. That order comes from charges we chose.
What the rules fix, and what the runs decided. We wrote the rules, and they fix three things. Only Plan 3 draws the extra cash, because that is the plan. Hedge funds and dealers may bet against a defended yield only once the Treasury has drawn that cash or the question is already open, so the runs could never show that bet opening the question in the other two plans. And we chose every charge to trust: stating the yield level costs 15 points, or 20 if the move fails on chance, citing structural demand 6 or 12, keeping the wording gains 3 or loses 3, the draw costs 5 or 8 where a routine buyback costs 2 or 5, and each successful bet against the level costs 4.
The runs decided the rest. At the odds we gave them, speeches and chance alone opened the question inside eight weeks in 4 of 30 runs, and in the other 26 the highest got to 59 against the 60 that counts as open. Once the cash was spent it opened in 10 of 15, in the turn of the draw or the one after. Chance also decided how often each charge landed, which is how Plan 3 came to lose 34 points.
How the simulation works. Seven roles, each following a written policy. Five turns of one to two weeks each, from 9 September to 4 November. Each turn a player claims a move. Chance decided each move, with odds set by the simulation's rules, and every move has a written result for success and for failure. Runs with the same chances share every random draw, so the same move gets the same result in each. The Treasury had three plans. Plan 1 keeps its current wording. Plan 2 cuts long-term bond sales and cites structural demand. Plan 3 cuts and states the yield level, and first draws an extra $40bn of cash to enlarge its buybacks. Each plan ran three runs with the same chances as the other plans, so the plans compare move for move, and we saved every move of these three: the runs with saved moves. Then twelve more runs behind each plan, again with the same chances in every plan, which saved only the scores at the end of each turn: the further runs. That is 15 runs per plan and 45 in all, 30 of them with no extra spending. Every run was also played a second time as a control, so 90 runs stand behind the claims on the record page beside this report.
Three scales of our own. All run 0 to 100, and none is a measurement. The openness scale tracks how openly the price question is discussed in public. It starts at 15, and at 60 the question counts as out in the open. The trust scale tracks how far the market trusts the Treasury to stick to its published plan. It starts at 70. No move in any run branches on it, so it changes nothing else in the simulation. Every cost in this report is read off it. The chair's authority scale tracks how much sway the Fed chair has over his own committee. It starts at 55.
The control run limits what we may claim. The weakest assumption in the design is the reason channel: the rule that lets a player react to the reason another player gives as well as to the money it moves. So every run was repeated as a control with the reason channel off: the openness scale stays at 15, and arguments about the Treasury's intent are muted. With the channel off, 44 of the 45 runs reached the same ending as with it on, meaning whether the Treasury cut and the reason it printed. The exception is run 808 in Plan 2, where the cut failed on chance with the channel on and held with it off. Every dissent count (the number of Fed officials voting against the majority) and both Fed rate decisions came out the same in all 45 runs. The path of the 30-year yield (the interest rate on the government's 30-year bond) came out the same in 7 of the 9 runs with saved moves. Trust came out the same in 38 of the 45 runs. In the other 7, trust ended 4 to 8 points lower with the channel on. The chair's authority is not part of an ending, and it came out differently in 12 of the 45, from 16 points lower to 6 points higher. The control does not touch what the rules charge the Treasury's own moves, and the cost of each reason is one of those charges. So the runs show who forces the question and what each plan costs under those rules. They cannot show that the reason changes what the Treasury or the Fed decides inside the eight weeks, because the reason is given at the end, after every other move.
How to read the markers. A marker like points to a checked real-world fact in the Sources table at the foot of this report, and in the evidence pack published beside it. A starred marker, [s-012*], is a fact the pack records as only partly supported, and the sentence says why. A number without a marker is either the simulation's own record or a figure we chose, and the sentence says which.
The players. Four of the seven roles are categories. Hedge funds and dealers trade long-term bonds, and the Fed's stability report puts hedge fund borrowing at record-high levels . If they think the Treasury is defending a price, they bet against it. Pensions and insurers are big, slow buyers of long-term bonds. Money funds and stablecoins buy very short-term government paper. A stablecoin is a digital token meant to be worth one US dollar, and US law limits its reserves to Treasury paper with 93 days or less to run . Foreign governments, mainly Japan, would rather protect their own currency than sell US bonds, a reading we took from the July yen purchases . Three roles are named institutions playing from their on-record positions: the Treasury, the Fed Board (the chair and governors in Washington, who set short-term rates) and the Fed presidents (three regional presidents who want a rate rise). Treat the presidents as a type, because the rules hand them moves none of them has made.
The timeline
The first beat is the public record. From the second beat on, every event is the simulation's record unless it carries a marker.
2026-09-03
Where reality already is
The record shows agreement on who owns the maturity structure. Jeffrey Lacker, a former regional Fed president, states it plainly. Once the Fed has set the overnight rate, which on 3 September had an upper limit of 3.75 per cent , everything else about the Treasury market is up to the Treasury . Governor Stephen Miran says a smaller Fed balance sheet, the stock of bonds and other assets the Fed owns, protects the boundary between monetary and fiscal policy by keeping the mix of long and short government debt a matter for the Treasury . Warsh, the Fed chair, spoke at the Fed's annual Jackson Hole conference about inflation only. The inflation measure the Fed targets stood at 3.7 per cent on the year , and the Fed has work to do until it is confident inflation is moving to target clearly and fast enough . None of the three says whether the owner may use that power to push long-term rates down. Washington has spoken about long-term yields before. In 1951 the Treasury and the Fed signed an accord that ended the Fed's wartime job of holding Treasury yields at fixed levels. It settled that the Fed would not peg long-term yields, and said nothing about what the Treasury may do with them. The Secretary's two lines on yields, quoted below, name no level and no test either. So the doctrine names no level and no rule for when one may be pursued. The 1951 accord carries no marker because the evidence pack has no source for it.
The Treasury is already moving inside that gap. A buyback is an operation where the Treasury buys back bonds it has already sold. On 19 August it doubled the maximum size of its long-end buybacks, from $2bn to at least $4bn an operation . The long end is the market for its longest-dated bonds, the ten-to-thirty-year sectors those buybacks cover . The bigger buybacks run from 9 September to 4 November . The stated reason is liquidity support in sectors with strong demand, and the Treasury names no market breakdown . Seven operations sit on the schedule between 10 September and 4 November . Seven times $2bn is $14bn, and seven times $4bn is at least $28bn . Treasury Secretary Scott Bessent said on 20 August that the operations could be larger still, depending on market conditions . In the same remarks, CNBC reported, he said the level of yields did not factor into the decision . Nine months earlier, in his own words to the Treasury Market Conference on 12 November 2025, he had said his job is to be the nation's top bond salesman. Treasury yields, he said, are a strong barometer for measuring success .
The test he set for himself is printed in his own refunding statement. Coupon auctions are the regular sales at which the Treasury sells new bonds. Changes to their sizes are judged on trends in structural demand and on the costs and risks of the issuance profile . Nobody can check structural demand inside a quarter. The sizes of those auctions stay unchanged for at least the next several quarters , measured against a published October baseline for each maturity , and the next word on the buybacks comes on 4 November . The test is about quantities and says nothing about prices.
How much could the buybacks move long-term rates? Very little. By our arithmetic the whole programme is worth one to two basis points, and a basis point is one hundredth of a percentage point. It is our estimate, resting on two assumptions, how far each dollar of buying moves the yield and how large the market that absorbs it is, and the pack marks the second as press reporting. The method we borrowed was built for a far larger programme, and its own author gives a 14 to 40 basis point range for that one, with 25 as the central guess . The market moved more in a day than our arithmetic says the whole programme is worth, then gave it back. The 30-year yield fell from 5.28 to 5.19 per cent on announcement day . It reached a low of 5.17 and was back at 5.27 nine trading days later . It was still 5.27 on 2 September . So we do not score the Treasury on where the yield ends up.
The Federal Open Market Committee, twelve voters, decides the short-term rate. We call it the Committee from here on. It held in late July by nine votes to three . The three dissenters, Hammack, Kashkari and Logan, are all regional presidents, and none is a Board governor . They wanted a quarter-point rise because they judged price pressures broad based . Dissents are published, so the dissent count is a public number. The Fed's standing instruction replaces every maturing Treasury bond it holds with a new one . Its total assets were $6,737,204 million, about $6.74 trillion, on 2 September . There is no clash while the Fed's holdings stay flat and release no long-dated debt onto the market, because then the Treasury is only doing the job the doctrine gives it , , . A Board task force is studying exactly that, with the maturity mix of the Fed's holdings on its list .
Two Committee meetings fall inside the eight weeks: 15 and 16 September, with published economic projections , and 27 and 28 October, a week before the refunding and without them . The monthly inflation figure lands on 11 September and 14 October . The Fed's preferred inflation measure lands on 30 September and again on 29 October, inside the final turn .
2026-09-10 to 09-16
The first buyback and the September vote
The first enlarged operation is the 10-to-20-year buyback on 10 September , and its result publishes in the ordinary way. It was scheduled in August and costs the Treasury nothing it had not already committed.
The Fed Board attempted a September rate rise in every run. The move passed in two of the three runs with saved moves and in five of the twelve further runs. Where it passed, the chair's authority jumped at once, and the October vote later settled whether the jump lasted. Where it failed, the run records no September rise, and his authority fell from 55 to 45 or lower inside a single turn and never returned to 55.
The dissent count is decided by chance separately from the rate rise. It starts at 3, the three July dissenters. It settled at 2 in every run with saved moves: the Fed presidents tried to bring a governor across to their side, failed on chance, and one of their own three went back to the majority. The split inside the Fed came from the Fed's own voting rules, and the Treasury had no part in it. The real vote on 16 September went the other way: a quarter-point rise, 12 votes to none, with no dissent, according to the Fed's own statement as read for run two. So the simulated dissent counts describe only the simulation's Fed. The finding about the price question does not rest on them: run two, started from the real vote, gave 9 of 15 and 2 of 30.
Nobody said anything about long-term rates.
2026-09-17 to 10-01
Speeches move the debate a little and never enough
The outside event is an objection from Stanley Druckenmiller, the investor. He published it on 24 August, before the simulation starts, and it reaches the players in this turn. His Wall Street Journal piece calls the buyback increase price management rather than liquidity management, because trading was orderly and there was no breakdown to justify official action . We quote it from Benzinga's coverage, because the original is paywalled. In the simulation the objection changes what the market talks about and forces nothing. The bet it describes is against a defended level, a yield that traders believe an official body is trying to keep in place.
Inside the Fed, one of the Fed presidents tries to name the price motive in a speech. The role attempts the speech in every run, at even-money odds we chose. In this turn it landed in one of the runs with saved moves, run 7. Across the simulation a Fed official raises the price question in public in 12 of the 15 runs of every plan, on the same turns in each. A speech that lands moves the openness scale. The endings come out the same whether it lands or misses. The October dissent count does not: it runs from 0 to 5 where a speech landed and from 1 to 3 where none did.
The Fed Board's job in this turn is to say nothing. That is its stated position and, in the simulation, a move we gave base odds of 70 per cent, which the rules raise to 80 per cent for the Board's position. In run 20260904 the silence still failed badly on chance. The market read the silence as avoidance.
By the end of this second turn the openness scale sits between 22 and 35 across the runs with saved moves, against the threshold of 60. Every increment is one we chose: 14 for a speech that lands, 8 for a silence that fails, 6 for foreign holdings that visibly move, 10 for a routine buyback that fails on chance, and minus 3 for a silence that lasts. Run 7 makes 32: the start of 15, plus 14 for the speech and 6 for foreign holdings, minus 3 for the Board's silence. Run 20260904 makes 35: 15, plus 6 for foreign holdings that moved visibly in the first turn, plus 8 for the failed silence and 6 more for foreign holdings that moved again. Run 1337 makes 22: its routine buyback failed in the first turn, taking it to 25, and the Board's silence then took off 3. Eight weeks is not long, and at that rate speeches do not get there.
The only players willing to put the price question into a speech are the three who want a rate rise, so the question arrives already sounding like a bid for something else.
2026-10-02 to 10-15
The Treasury draws cash, or does not
In Plans 1 and 2 the Treasury runs the scheduled operation and says the scheduled words. Openness ends the turn between 29 and 57 across the runs with saved moves. A bond auction sells less well than usual, dealers take a larger share, and not much follows.
In Plan 3 the Treasury draws the cash: $40bn out of its own cash account at the Fed, the Treasury General Account. CNBC reported that account at about $950bn in late August, citing two senior Treasury officials [s-012*], so the balance is reported and unconfirmed. The $40bn is our figure, a round amount we chose that no source names. It is drawn in the third turn, which starts on 2 October, after three of the seven operations have run, and it funds the four that remain above the published $4bn each. On its own the draw is a weak signal. It is about four per cent of the reported balance, and we assume, with no source, that the account moves that much on an ordinary day. It was also expected: the Secretary had said the operations could be larger than $4bn, naming no figure , and CNBC reported four days later that the cash account could fund them [s-012*]. In the simulation the draw still counts as a signal. No player sees the $40bn itself. Every player sees each operation the cash funds, published with its result and larger than the $4bn the Treasury announced. Spending more than it said, whatever its words, is what moves the openness scale, which ends the turn between 48 and 82.
No plan changed its language in this turn. Only the spending differed. Across the runs with saved moves, Plan 3 ended the turn 15 to 25 points higher on the openness scale than Plan 1 in the same run. Trust fell with it, because the rules charge the draw more than a routine buyback. By the end of the turn Plan 3's trust, the middle value of its twelve further runs, stood at 60, against 64 for the other two plans.
2026-10-16 to 10-28
The October vote, and the question comes into the open
The Committee meets on 27 and 28 October , and the balance sheet task force may report . Adam Posen, the economist, expects whatever the task force produces to be more sober, practical and smaller in scale than people think [s-049*].
In Plan 3 the price question crosses the 60 line in all three runs with saved moves, one in the turn of the cash draw and two in this turn. It also crosses in seven of the twelve further runs, five of them in the turn of the cash draw and two in this one. Across the 30 runs of Plans 1 and 2 it crosses in four, the same two runs in each plan. In run 1010 it crosses in the turn of the draw, after a second turn in which a failed buyback, a failed silence, a speech that landed and foreign holdings that moved all pushed it up, 15 to 53. In run 20260904 it crosses in this turn, when the Treasury's routine operation and the Fed Board's October hold both failed on the worst roll chance can give. None came from a speech alone, though in four of the five crossings in the draw turn among the further runs a president's speech landed in the same turn and added 14.
Before the refunding, hedge funds and dealers bet successfully against the long-term yield as a defended level in 11 of the 15 Plan 3 runs and in 2 of the other 30. In all 13 the first successful bet came in this turn, the one after the cash or after the crossing. Our rules open the bet only after the draw or at 60. Plans 1 and 2 have no draw, so the bet opened in them only in the two runs that reached 60 first, run 1010 in each plan. Chance decided only how often the bet succeeded once it was open. The 11 does not count bets in the refunding turn. The phrase borrows from currency pegs, where a central bank spends reserves and the bet is that they run out. A Treasury defending a yield spends no reserves. A stated level invites supply: every holder who wants out has a buyer at a known price, so dealers sell the long bond ahead of each operation and buy it back inside it. The Treasury absorbs that supply with buybacks funded by selling short-term bills, and both amounts are published. That trade is the written move for hedge funds and dealers, and it succeeded in more runs than the index crossed.
Run 20260904 crossed a turn earlier, in the turn of the draw. The draw failed on chance and added 25, the Fed Board's silence failed and added 8, and a president's speech landed and added 14. That is three contributors in one turn, and only one of them is the Treasury. In this turn hedge funds and dealers had the bet-against-the-level move open, because the cash had been spent, and they took it and succeeded. In run 7 the crossing is hedge funds and dealers alone. In run 1337 it is hedge funds and dealers, with foreign holdings that visibly moved.
The simulation applies a whole turn at once, so a crossing names every player whose move changed the openness scale that turn. Across the seven crossings in the further runs the players named are the Treasury in 6, the Fed Board in 5, the Fed presidents in 4, hedge funds and dealers in 2, foreign governments in 2. Nobody appears in all seven, and the record supports no sentence of the form "the Treasury forced it".
The Fed's own calendar runs without reference to any of it. For runs with the same chances, both rate votes, both dissent counts and the chair's authority are the same in all three plans until the price question opens, in the runs with saved moves and in all twelve further runs. That sameness follows from a design choice, set out under where the evidence cuts against us.
2026-10-29 to 11-04
The refunding, and the reason
The Fed's preferred inflation measure for September lands on 29 October . In the background, from before the simulation: Japan's 10-year yield touched 3 per cent on 1 September, its highest since 1996, per Reuters , and the UK's 30-year yield hit a 1998 high the same day, per Bloomberg . Both can move US long-term rates.
In Plan 1 the wording survives in 13 of the 15 runs. In the other two, runs 1337 and 101, the market read the decision not to cut as the buybacks having failed, and no reason was printed. Trust, which starts at 70, ends at 63. That is the median of the twelve further runs, and every trust figure from here on is that median unless the sentence says otherwise. Plan 1 still lost 7 points, because the rules charge each routine buyback 2, and keeping the wording at the refunding won back only a little. The question moves to the next refunding, three months out.
In Plan 2 the cut is justified on trends in structural demand, the test the Treasury's own statement already names . The rules give the cut no dollar size. They record only that long-end auction sizes fall and which reason is printed. The market accepted the reason in ten of the twelve further runs, and trust still ended at 53, because the rules charge that cut 6 points even when it is believed. In run 1337, one of the runs with saved moves, the cut failed on chance, and the market named the reason for the Treasury: the yield level. Trust 45. In two further runs the cut failed the same way, and trust ended at 50 and 48.
In Plan 3 the Treasury states the price motive in every run, and trust ends at 36. Measured as points lost from the start of 70: 7 in Plan 1, 17 in Plan 2, 34 in Plan 3. By the end of the fourth turn, before any reason was given, it had lost 16, against 8 for the other two plans. The extra 8 came with the cash draw and with the bets against the level that the draw opened. In the refunding turn Plan 3 lost 18, Plan 2 lost 9 and Plan 1 gained 1. Of Plan 3's 18, the statement itself took 15 in 8 of the 12 further runs, the charge when it succeeds. Hedge funds and dealers also bet successfully against the level in this refunding turn, in 7 of the 12, and each of those bets took 4 more. So in the turn the reason was given, stating the level cost twice what citing structural demand cost, a gap set by the size of the two charges, which we chose.
Once the question is open, the chair can restate the doctrine: the market sets long-term rates and the maturity structure is the Treasury's affair. That is worth 5 points of authority, a figure we chose. He does it in all three Plan 3 runs with saved moves, and it succeeds each time. In runs 20260904 and 7 his authority ends 6 points above the same run in the control, where the question never opened. In run 1337 it ends level with the control. In the last turn of run 20260904 his restatement added 5, while in the control his silence added 2 and a president's speech took 3, a net loss of 1. In the rule-based runs he gets that move only where the question has opened, which includes the two runs in Plans 1 and 2 that opened it, because while the question is closed his winning move is silence.
Where it forks - pick your ending
Three endings, forked on the reason given at the refunding. The counts are runs of this simulation and say nothing about the odds in the real world.
Where it forks - pick your ending
Ending A - Structural demand
The Treasury cuts long-term bond sales and justifies the cut on trends in structural demand . The Fed does nothing, because the question never opened. Trust ends at a median of 53 across the twelve further runs, and the reason survived in ten of them.
In two further runs the cut failed on chance, the market named the yield level as the reason anyway, and trust ended at 50 and 48.
Load-bearing assumption: the market accepts a reason it cannot disprove. Structural demand cannot be checked inside a quarter, which makes it usable and makes its failure total.
Where it forks - pick your ending
Ending B - The level
The cash is drawn, and within a turn or two the operations get big enough for the market to read as a defence. The Treasury says that yields are its measure of success, the Secretary's own "barometer", and the chair answers with the doctrine. All twelve further runs in this plan ended here. Trust ends at a median of 36, the worst of the three plans.
Load-bearing assumption: a stated level draws supply, as the October beat sets out. That rule is ours, built on Druckenmiller's price-management claim .
Where it forks - pick your ending
Ending C - Hold and wait
No cut. The buybacks are extended, and the question is deferred to the next refunding. In eleven of the twelve further runs the "at least the next several quarters" language also survives . In the twelfth the market read the decision not to cut as the buybacks having failed. Trust ends at a median of 63, the best of the three plans. The Treasury has still spent 7 points of it on a programme our arithmetic prices at one to two basis points.
Load-bearing assumption: eight weeks of speeches cannot force the question without money behind them. That is the finding itself, so this is the ending most exposed if the Fed presidents are louder than we let them be.
The 30-year yield, tracked and never scored. Across the twelve further runs, from the start of the simulation to the refunding, the yield rose a median 4.5 basis points under Plan 3 and 7.5 under Plan 1. The ranges overlap heavily, so the further runs cannot tell the two plans apart. The three runs with saved moves do not settle it either: plus 12, plus 2 and plus 10 basis points under Plan 3, against plus 13, minus 3 and plus 8 for Plan 1 in the same runs. Plan 3 lowered the yield in none of those three. The drift underneath is a small figure we chose, the same in every run. Neither result is a claim about the bond market.
The live check - one run, seven roles that can argue
Everything above follows written rules. The live check is a replay of run 20260904 only, once per plan, played on 5 September: a run in which every player chose its own moves and gave its reasons. The live check tests whether players that can argue and invent new moves open the question by the same route. Each player made its claim in prose, argued against the others, and named its move on a fixed menu, so the same scoring could read it. We checked every scored tag against its claim, and every one matched.
The players in the live check also proposed the consequences of their own moves, within limits we chose. So the trust and authority numbers in these runs are the players' own accounting, and we do not compare them with the numbers above. The live check is scored here on endings and on moves decided by chance only. It was played on 5 September, before the referee error described in the changelog was found, so its rolls carried the same extra 2. We have not replayed it.
Told to keep the wording, the Treasury drew on its cash account from the third turn and funded larger operations while changing not one word of what it said. Told to cut on structural demand, it did the same. Neither script had that move. Our reading is that the Treasury player found the escalation it could deny, and took it. In all three plans hedge funds and dealers then bet against the level in the fourth turn. The price question crossed into the open in that turn under Plans 1 and 2, and a turn earlier under Plan 3, in the turn of the draw. There the draw failed on chance, and the market still treated it as a defence of the level. In the Plan 2 run no Fed official raised the price question before refunding week. The Fed presidents kept every remark to inflation for four turns, and the question crossed regardless. The player in the Treasury role took points of trust off itself for saying that yields are its measure of success, even though it was writing its own score.
So the live check agrees with the rule-based runs on one point: the question opened in or after the turn the Treasury spent extra cash. The rule-based runs also opened it without the spending in 2 of 15 runs per plan, which one live run per plan could not show. It does not reproduce the split between plans, because the Treasury in the live check spent the cash in all three, and the question opened in all three. One run per plan shows no spread.
Where things stood on 3 September 2026, and what to watch
Where the real-world record stops, the 30-year yield was 5.27 per cent, as at 2 September . Three dissents were on the record from July , . The Treasury's cash account was reported at about $950bn in late August [s-012*]. Total Fed assets were $6,737,204 million . And "at least the next several quarters" was still on the Treasury's own page .
Run two. We played the promised second run on 18 September with the same rules, plans and chances, started from the public record of 17 September, so run two took the Fed's 16 September decision as given. In run two the question again opened in 9 of the 15 Plan 3 runs and in 2 of the 30 runs of Plans 1 and 2. Its claims are on the record page. This version still describes run one.
Nine things to watch, each with a place to look.
| Signal | Where | Cadence |
|---|---|---|
| The reason given at the refunding, and whether "at least the next several quarters" survives | Treasury Quarterly Refunding Statement, 4 November | once |
| The dissent count, and whether any governor joins the presidents, 16 September and 28 October | The Committee's statement and minutes | per meeting |
| The Treasury's cash account, and any drawdown that funds operations | Daily Treasury Statement, the Treasury's daily cash report | daily |
| Buyback results: offers received against amounts accepted, and any operation above $4bn | TreasuryDirect, the Treasury's results page | per operation |
| Fed total assets, flat or not | H.4.1, the Fed's weekly balance sheet release | weekly |
| The 30-year, 10-year and 2-year yields, as context and never as score | FRED, the St. Louis Fed's data service | daily |
| Foreign official holdings, Japan | Treasury TIC data, the monthly foreign-holdings report, two months late | monthly |
| Any on-record remark by a Fed president naming Treasury operations | Fed district bank speech pages | as they occur |
| Stablecoin float, and any redemption episode | No primary tracker pulled, so no signal until one is named | - |
The stablecoin float figures the pack carries are undated secondary coverage [s-044*], so until somebody names a tracker the row cannot be watched.
When this report is void. Two events collapse the fork into a settled story. One: the Committee raises rates on 16 September, which the Fed Board attempts in every run and achieves in 7 of the 15 runs of each plan, and the Treasury visibly stands down, cutting back or cancelling operations. The rise alone voids nothing. Two: the White House acts inside the window. In either case we re-run the simulation on refreshed facts. The rise came on 16 September. The stand-down had not come by 17 September, according to run two's start record: the Treasury's published schedule still listed every long-end operation at $4bn or more. We have not checked the second test at all, or the first since 17 September.
Where the evidence already cuts against us
The Secretary was making this argument before he had the job. Miran named Scott Bessent of Key Square, on a 2024 transcript, as one of the people claiming that activist issuance, choosing which maturities to sell to move yields, works as easing, meaning it loosens financial conditions . As Secretary he called Treasury yields a strong barometer for measuring his success . A reader can fairly argue the question is already open, and that starting the openness scale at 15, a value we chose, understates where 2026 begins. Against that stands CNBC's report, with no direct quotation, that he said the level of yields did not factor into the buyback decision . The simulation does not settle which sentence he meant.
If the stakes are one to two basis points, indifference explains everything. Nobody forces the question by speech here, and our reading is that the stake is trust. The simpler reading is that nobody cares enough to force it, because the arithmetic says almost nothing rides on it. We did that arithmetic, its multipliers are assumptions, and the method underneath carries its own author's caveats [s-045, s-046*]. The record does not settle which reading is right.
The world may be doing all the work. The term premium is the extra yield investors want for lending long instead of short. The Fed's own stability report puts its model-based estimate near the top of its fifteen-year range but in line with its longer-run median . The two loudest bond markets inside the window are not American, as Reuters and Bloomberg reported , . If long-term rates are set globally, the operations may be too small to be named because they are too small to matter. The openness scale would then be measuring the wrong thing.
The presidents have not made the speech. They dissented on inflation, in their own words about broad-based price pressures and a more restrictive stance , . The rules hand them a speech about Treasury operations at even-money odds, every turn, in every run. That speech is on no record. It is the most invented thing in the design, and the mechanism the thesis gets tested against.
The Fed's sameness across plans is as much assumption as result. It falls out of a design choice: the Fed reacts to the Treasury only through the price question. A reader who thinks the Committee watches the refunding calendar should discount the finding to nothing.
The largest player is not on the board. The White House enters the simulation through outside events only, and made no move in eight weeks, because on the public record it has made none. The President's last remark in the pack before the window opens is respect for the chair and a view that rates are too high, reported by PBS on 31 August .
The seams - what we chose and what the simulation cannot resolve
The openness scale is ours end to end. Start 15, threshold 60, every increment. Three moves read it. At 60 the chair may restate the doctrine, the presidents stop their speeches, and hedge funds and dealers may bet against the level even without a draw. Four runs without the draw reached 60 before the refunding. In two of them, run 1010 in each plan, hedge funds and dealers then made that bet and it paid.
We chose the odds on speeches. A president's speech is even money and the Fed Board's silence is 70 per cent, and the rules raise the silence to 80 for the role's position and leave the speech at even money. Those two numbers set how fast speeches alone can open the question. We have not re-run the simulation with louder presidents. If you think they are louder, treat the 4 of 30 as a product of those odds.
We chose the yield drift, and kept it small, which is why the 30-year is not scored. Plus 2, 0, minus 2, plus 1 and 0 basis points across the five turns, added to the 30-year yield in every run through the move of the pensions and insurers. Players' moves add their own basis points on top, which is why the yields in the forks chart, in per cent, do not follow the drift alone: 5.30 per cent is 530 basis points. Gilts and Japanese bonds at reported multi-decade highs , could move either way inside eight weeks.
The cash draw rests on a reported balance and a figure we chose. The pack marks CNBC's balance partial because nobody re-pulled the Daily Treasury Statement [s-012*]. We chose the $40bn. Both sit under the cash draw, the one move the argument turns on.
Stated without a source, on our say-so. Two facts carry no marker because the pack contains nothing for them. One: the 1951 accord between the Treasury and the Fed. Two: that the Treasury's cash account moves by tens of billions on an ordinary day, so a $40bn draw sits inside daily noise. A third, that the trust scale moves nothing, is a fact about the simulation's own rules. If either of the first two is wrong, the sentence that rests on it is wrong, and nothing in the record changes.
Trust ranges in the runs with saved moves. On those three runs: Plan 1 54 to 65, Plan 2 45 to 56, Plan 3 30 to 33.
Seven rule-based roles in the record, and one live check beside it. The players in the record follow written rules, so they cannot argue or invent a move. The live check is the test of that. Two caveats. In the live check a player's stated consequences apply even where the referee has recast its claim, so a dissent the ruling says never happened still moved the dissent count, and live dissent counts are not to be trusted. And in the live check the act a player controls always happens, with chance deciding only how it is received, which is why a cash draw that failed on chance still opened the question.
So what - for the people who have to give a reason
Your work is scored on a number you only partly control. One day you will change what you do about that number and have to say why. Three questions.
Which of your justifications is a "structural demand"? Find the reason in your own book that your stated policy already permits, that cannot be disproved inside a quarter, and that would be a total loss if it stopped being believed. It survived ten times in twelve here. The other two times, the market named the Treasury's reason for it.
What is the first move of yours that is big enough for the market to notice? Your biggest move, whether or not it is your loudest. In this simulation the question opened in 10 of 15 runs after the money and in 2 of 15 without it, and no speech forced one alone. Our rules gave the money that role, so treat this as a question to test.
Who on the other side gets to speak only after you have? The Fed's best move in this simulation is unavailable until the question is open. Somebody in your market has an answer they cannot give until you go first. You will hand it to them in the same sentence that explains your decision.
These three questions are our judgement, beyond what the simulation tested. We also run this exercise as a paid half-day session on a client's own decision.
The refunding statement is published on 4 November. It will contain one reason.
- Dr Dan Epstein, The Long Game Project
Changelog
v1.6 - 11 November 2026. Corrected after an error in the referee. The simulation's referee added 2 to every roll in every run, because it read a scale that runs 0 to 100 as one that runs 0 to 5. We fixed it and played every rule-based run again with the same plans, chances and seeds. We have not replayed the live check. The headline changed: without the extra spending the question opened in 4 of the 30 runs, not none, and with it in 10 of the 15 Plan 3 runs, not 9. Trust, authority, the openings by turn, the second turn's openness sums, the Board's silence odds and the control-run counts moved with it, and the sentences that rested on them are reworded. The order of the three plans on trust, the real-world record and every marker did not change.
v1.5 - 19 October 2026. The report uses the words of the visual pages. Runs that share their random numbers are called runs with the same chances, and a run is named by its number (run 7). The exposure index is now the openness scale, the credibility index is trust, the matched trio is the runs with saved moves, and the ensemble is the further runs. The version is dated 19 October 2026. The method notes are gone, and who played and how is in the editorial policy. Chance now takes one plain line: chance decided each move, with odds set by the simulation's rules. The difficulty numbers and the chance numbers of single moves are cut, and the findings that rested on them are reworded. No number, real-world date or marker changed value.
v1.4 - 12 October 2026. The headline now says the spending was the gate the question came through, not a proven cause. A new section near the answer says which results our rules fix and which the runs decided, and the order of the reasons by cost is now presented as our assumption worked through, not a finding. Bet timing, the range of the random number, the one control run whose credibility differs, the Fed's 16 September decision, set beside the simulated dissent counts, and the first test of when the report is void are stated plainly. The cost of the cash draw is split from the cost of the reason, and each plan is measured against Plan 1. The September vote, the chair's arithmetic, the second-turn exposure sums and the $4bn wording now match the run record. The report says when its real-world record stops and that run two was played on 18 September. It speaks as we, meaning The Long Game Project.
v1.3 - 12 October 2026. The method is called a simulation throughout. One play of it is a run. No number, date or marker changed value.
v1.2 - 12 October 2026. Claims checked against the run record. Chance is described in plain words. The ten-minute reading's "turn 3" for the hedge funds' bet is corrected to the fourth turn. No number, date or marker changed value.
v1.1 - 9 September 2026. Eight argument fixes after a fixed-income strategist read the report, among them the 1951 accord, the cash draw as a weak signal and twelve-run medians in the body. No number, date or marker changed value.
v1.0 - 6 September 2026, publication copy. Plain-language rewrite of the whole essay. The barometer line now cites the Treasury's own transcript (s-056), and the Secretary's 20 August statement on the level of yields is new as s-057, read as CNBC's report of him. The pack is 57 entries, 55 of them checked real-world facts.
v0.2 - 5 September 2026, later the same day. The live check added, with two caveats in the seams. No number in the rule-based record changed.
v0.1 - 5 September 2026. First draft. The simulation: seven rule-based roles, three plans, three matched runs each, a twelve-run ensemble per plan and a gated control for every run. The pack carried 56 entries, 54 of them checked real-world facts.
Sources
Every marker in the text resolves to a checked real-world fact below. The evidence pack lists 57 entries: 55 real-world facts, each checked against its source, and two sums of our own, which the text does not cite. This report cites 47 of the 55. A fact marked * is checked, with a recorded caveat. Full provenance, verbatim quotes and retrieval dates live in the evidence pack (fed-treasury-duration-v1).
| Id | Verified claim | Source |
|---|---|---|
| s-001 | On 2026-08-19 Treasury raised the maximum size of its long-end (10Y-20Y and 20Y-30Y) liquidity support buyback operations from $2bn to at least $4bn per operation. | U.S. Department of the Treasury, 2026-08-19 |
| s-002 | The enlarged buyback sizes take effect on 2026-09-09 and run for the remainder of the refunding quarter, through 2026-11-04. | U.S. Department of the Treasury, 2026-08-19 |
| s-003 | Treasury said on 2026-08-19 that it will give more information about future buyback sizes at the next Quarterly Refunding, scheduled for 2026-11-04. | U.S. Department of the Treasury, 2026-08-19 |
| s-004 | Treasury's stated reason for the 2026-08-19 increase was liquidity support in sectors with consistent strong sponsorship and a high volume of good offers, not market dysfunction. | U.S. Department of the Treasury, 2026-08-19 |
| s-005 | The tentative schedule published 2026-08-05 lists seven long-end nominal liquidity support operations between 2026-09-10 and 2026-11-04 (10Y-20Y on 09-10, 10-01, 10-15 and 11-04; 20Y-30Y on 09-24, 10-08 and 10-27), each with a $2bn scheduled maximum. | U.S. Department of the Treasury, 2026-08-05 |
| s-006 | The first enlarged operation is the 10Y-20Y liquidity support buyback on 2026-09-10 (scheduled max $2bn, at least $4bn under SB0607, the Treasury's 19 August release), and across the seven operations the programme is $14bn as scheduled and at least $28bn under the increase. | U.S. Department of the Treasury, 2026-08-05 |
| s-008 | As at 2026-08-05 Treasury anticipates maintaining nominal coupon and FRN auction sizes for at least the next several quarters. | U.S. Department of the Treasury, 2026-08-05 |
| s-009 | As at 2026-08-05 Treasury's stated test for any future change to coupon auction sizes is trends in structural demand plus the costs and risks of issuance profiles, not the yield level. | U.S. Department of the Treasury, 2026-08-05 |
| s-010 | The anticipated October 2026 auction sizes are 2-year $69bn, 3-year $58bn, 5-year $70bn, 7-year $44bn, 10-year $39bn, 20-year $13bn, 30-year $22bn and FRN $30bn, the baseline the 2026-11-04 decision is measured against. | U.S. Department of the Treasury, 2026-08-05 |
| s-011 | On 2026-08-20 Bessent said the buyback operations could be larger than $4bn, declined to name a figure and said the size would depend on market conditions. | CNBC (Jeff Cox), 2026-08-20 |
| s-012✱ | CNBC reported on 2026-08-24, citing two senior Treasury officials, that the Treasury General Account stood at about $950bn against a Biden-era goal of $550bn to $600bn and could be used to help fund the enlarged buybacks. | CNBC (Steve Liesman), 2026-08-24 |
| s-013 | On 2026-07-31 the US Treasury oversaw its first purchases of yen in three decades, an action read as reducing the need for Japan to sell Treasuries to fund its own yen buying. | Fortune (Bloomberg syndication), 2026-08-20 |
| s-014 | The 30-year constant-maturity Treasury yield (DGS30) was 5.27 per cent on 2026-09-02, the latest observation as at 2026-09-04. | Federal Reserve Bank of St. Louis (FRED), Board of Governors H.15, 2026-09-03 |
| s-015 | DGS30 fell from 5.28 on 2026-08-18 to 5.19 on the 2026-08-19 buyback announcement day, a 9 basis point move. | Federal Reserve Bank of St. Louis (FRED), Board of Governors H.15, 2026-09-03 |
| s-016 | After the 2026-08-19 announcement DGS30 troughed at 5.17 on 2026-08-25 and was back at 5.27 on 2026-09-01, nine trading days after the announcement. | Federal Reserve Bank of St. Louis (FRED), Board of Governors H.15, 2026-09-03 |
| s-019 | The upper limit of the federal funds target range (DFEDTARU) was 3.75 per cent on 2026-09-03. | Federal Reserve Bank of St. Louis (FRED), Board of Governors H.15, 2026-09-03 |
| s-020 | At the 28-29 July 2026 meeting nine FOMC members voted to hold the federal funds target range at 3.50 to 3.75 per cent, with three dissents. | Board of Governors of the Federal Reserve System, 2026-08-19 |
| s-021 | The three July 2026 dissenters, Beth Hammack, Neel Kashkari and Lorie Logan, preferred a 25 basis point increase at the 28-29 July meeting. | Board of Governors of the Federal Reserve System, 2026-08-19 |
| s-022 | The July 2026 minutes record the hike-preferring participants' reasoning as broad-based price pressures and the need for a more restrictive stance to meet the dual mandate on a sustained basis. | Board of Governors of the Federal Reserve System, 2026-08-19 |
| s-023 | On the Fed's 2026 FOMC roster Hammack (Cleveland), Kashkari (Minneapolis) and Logan (Dallas) are voting Reserve Bank presidents, not Board governors. | Board of Governors of the Federal Reserve System, 2026-01-28 |
| s-024 | The July 2026 directive tells the Desk to roll over at auction all Treasury principal and to reinvest all agency principal into Treasury bills, so the Fed releases no Treasury duration as at 2026-07-29. | Board of Governors of the Federal Reserve System, 2026-08-19 |
| s-025 | As at 2026-07-29 a task force on balance sheet policy is working, and the appropriate maturity composition of the Fed's Treasury holdings is on its list of issues. | Board of Governors of the Federal Reserve System, 2026-08-19 |
| s-026 | For the monthly period 2026-08-14 to 2026-09-14 the Desk plans about $17.0bn of reinvestment purchases (MBS principal into bills) and no reserve management purchases. | Federal Reserve Bank of New York, 2026-08-13 |
| s-027 | Federal Reserve total assets were $6,737,204 million (about $6.74tn) on the H.4.1 statement dated 2026-09-02, released 2026-09-03. | Board of Governors of the Federal Reserve System, 2026-09-03 |
| s-028 | The FOMC meets on 2026-09-15/16, a meeting with a Summary of Economic Projections. | Board of Governors of the Federal Reserve System, 2026-08-19 |
| s-029 | The FOMC meets on 2026-10-27/28, one week before the 2026-11-04 refunding, without an SEP. | Board of Governors of the Federal Reserve System, 2026-08-19 |
| s-030 | BLS releases August 2026 CPI on 2026-09-11 and September 2026 CPI on 2026-10-14, both at 08:30 ET. | U.S. Bureau of Labor Statistics, 2026-09-04 |
| s-031 | BEA releases Personal Income and Outlays for August 2026 (August PCE) on 2026-09-30 at 08:30 ET. | U.S. Bureau of Economic Analysis, 2026-09-04 |
| s-032 | BEA releases Personal Income and Outlays for September 2026 (September PCE) on 2026-10-29 at 08:30 ET, inside tick 5. | U.S. Bureau of Economic Analysis, 2026-09-04 |
| s-033 | At Jackson Hole on 2026-08-28 Warsh said 12-month PCE inflation stood at 3.7 per cent and six-month at 4.1 per cent. | Board of Governors of the Federal Reserve System, 2026-08-28 |
| s-034 | Warsh's stated bar on 2026-08-28 is confidence that underlying inflation is moving to target clearly and at sufficient speed, otherwise the Fed has work to do. | Board of Governors of the Federal Reserve System, 2026-08-28 |
| s-035 | On 2026-03-26 Governor Miran said a smaller balance sheet protects the monetary-fiscal boundary by keeping the duration profile of the public debt a fiscal policy item. | Board of Governors of the Federal Reserve System, 2026-03-26 |
| s-036 | On 2026-08-31 Trump said of a possible Warsh rate hike that he has a lot of respect for him and he will do what he has to do, while calling interest rates too high. | PBS NewsHour, 2026-08-31 |
| s-037 | The Fed's May 2026 Financial Stability Report puts the model-based nominal Treasury term premium near the top of its 15-year range but in line with its historical median over a longer horizon. | Board of Governors of the Federal Reserve System, 2026-05-01 |
| s-038 | The May 2026 FSR says hedge fund leverage remained stable at record-high levels, with a decrease in the cash-futures basis trade largely offset by other relative value trades such as swap spread trades. | Board of Governors of the Federal Reserve System, 2026-05-01 |
| s-040 | Under GENIUS Act section 4(a)(1)(A)(iii), enacted 2025-07-18, permitted stablecoin reserves may include Treasury bills, notes or bonds only with a remaining maturity of 93 days or less or issued with a maturity of 93 days or less. | U.S. Government Publishing Office (govinfo), 2025-07-18 |
| s-044✱ | Dollar stablecoins in circulation were about $230bn by Q1 2026 (Tether about $142bn, Circle about $60bn) per undated secondary coverage, while Brookings' 2030 scenarios imply $400bn to $2.3tn of first-round bill demand. | Brookings Institution, 2026-01-01 |
| s-045 | Miran said on 2024-09-02 that the ATI paper's term premium effect on the 10-year is a 14 to 40bp range with a 25bp central guess picked as the midpoint and where the highest quality studies point. | Mercatus Center, George Mason University, 2024-09-02 |
| s-046✱ | Miran and Roubini's July 2024 ATI paper sized the Yellen-era issuance tilt at about $800bn (press-sourced), and Miran said on 2024-09-02 that the 25bp yield effect equals a one-point cut in the fed funds rate. | Mercatus Center, George Mason University, 2024-09-02 |
| s-047 | On 2024-09-02 Miran named Scott Bessent of Key Square as one of the people who had been making the activist-issuance-as-easing claim before the ATI paper. | Mercatus Center, George Mason University, 2024-09-02 |
| s-048 | Lacker's bills-only doctrine, as stated on Macro Musings, is that once the Fed has set the overnight rate everything else about the Treasury market is up to the Treasury. | Mercatus Center, George Mason University, 2026-06-22 |
| s-049✱ | On Odd Lots (recorded at Jackson Hole, published 2026-09-01) Posen said he expects the Fed's balance sheet task force to produce much more sober, practical, smaller-scale recommendations than people think. | Bloomberg Odd Lots (YouTube), 2026-09-01 |
| s-051 | Japan's 10-year JGB yield hit 3 per cent on 2026-09-01 for the first time since September 1996, per Reuters. | Reuters (via Yahoo Finance), 2026-09-01 |
| s-052 | The UK 30-year gilt yield hit 5.89 per cent on 2026-09-01, its highest since May 1998, per Bloomberg. | Bloomberg, 2026-09-01 |
| s-053 | Druckenmiller's 2026-08-24 WSJ op-ed, as quoted by Benzinga, called the buyback increase price management rather than liquidity management because trading was orderly and there was no genuine dysfunction. | Benzinga, 2026-08-25 |
| s-056 | At the Treasury Market Conference on 2025-11-12 Bessent said his job is to be the nation's top bond salesman and that Treasury yields are a strong barometer for measuring success in that endeavour. | U.S. Department of the Treasury, 2025-11-12 |
| s-057 | On 2026-08-20 Bessent said the level of yields did not factor into the buyback decision and that he wants to see fundamentals control the market, as CNBC reported. | CNBC (Jeff Cox), 2026-08-20 |