Where protection binds, an always-on keeper is worth 200.1 bps of equity, and the gap opens when nobody is awake.

The pooled figure is the study’s confirmatory result and it is not the headline sentence, because the distribution of paired gaps has no middle: 68 paths are exact ties where both arms did the same thing, and 121 more differ by less than a keeper fee. The pooled median falls between two populations rather than inside either one, 44.7 bps below the median of the paths where protection bound, and 580 times the largest gap on any path where it did not. Figure 2 is that distribution.

What ran

16 cells, shock (4) × clock (2) × books (2), fully crossed, at 50 paths each. 800 paths, 4,000 arm-runs, 8,000 book-level ledgers. One price path is replayed across all five arms, which is why the tests are paired.

Evidence label

simulation on everything generated here; the human arms are modelled person. The venue is one Aave-V3-shaped rule tuple, not three integrations, and the liquidator is testnet-shaped, mainnet MEV is faster than it, which is a pre-registered way for the agent to lose that this study does not run.

Corrections

Five defects were found by four independent passes over this study. 3 of them flattered the agent, one took something back from it, and one moved the result in no direction at all. All five are in §0, with what was published beside what is published now.

Data

• T1-runs.jsonl (23.8 MB), one JSON record per arm-path, with its full event ledger · not in the repository, see below
• T1-paths.json (29.0 MB), the 800 exogenous price paths, before any arm touched them · not in the repository, see below
• T1-cells.json (59 KB), per-cell summaries, one object per cell
• T1-confirmatory.json (565 KB), every published statistic, machine-readable
• T1-run-manifest.json (37 KB), seeds, frozen constants and the run's own checks

The two large ones are not committed, and saying so is the difference between a link and a data-availability statement: at 23.8 MB and 29.0 MB they are excluded by the repository’s own .gitignore and are regenerated rather than stored. What is committed is everything needed to regenerate them byte for byte, the per-cell seeds in T1-cells.json and the frozen constants and run checks in T1-run-manifest.json, plus every published statistic in T1-confirmatory.json. The two ledgers are available on request as a release asset.

Schema termix-sim/T1-mvp/v1, generated 2026-09-07T18:12:23.600Z.

The night does not make the human worse at the shocks that were always going to beat them. It takes away the two that had been leaving them time. The realistic human keeps 74.0 of every hundred positions on the slow grind at 09:30 and 3.5 at 01:00, and 23.5 against 0.0 on the crash. On the two wicks the number is identical at both clocks, 0.0 either way on wick-persist and 12.5 either way on wick-1block, because three seconds is not enough for a person at any hour. That is also why this is a claim about availability rather than about skill: across the same axis the agent moves by 1.4 positions per hundred and the attentive human by 1.2, both upward, and neither of those two movements is a difference this study can resolve. The agent’s95% interval runs -3.4 to 6.1 points and the attentive human’s -4.7 to 7.3; both cover zero, and passive moves by exactly nothing on an interval of -1.7 to 1.8. Neither of those two policies has a clock in it at all, so what the figure shows there is variation in the exogenous paths and the liquidator draws. The realistic human’s -23.5 points, on an interval of -28.3 to -18.8, is on a different scale entirely, and that separation is the whole finding rather than the ranking of the five bars.

One oddity in the figure is better named than left to be found. H3 is H2 with sleep forced over 23:00–07:00, so it is strictly less available by construction and cannot really be better, and it measures very slightly better at both clocks, 28.5 against 27.5 in working hours and 5.6 against 4.0 at night. The two arms share the market and share the liquidator, but they draw availability from separate streams, keyed by arm name on purpose so that adding or removing an arm cannot perturb another arm’s draw. This one contrast is therefore paired on the exogenous world and unpaired on the very quantity it tests, which is why the study reports it as a null this design cannot resolve rather than as evidence that a sleep schedule does not matter.

0What four passes over this study found wrong with it

An independent replay recomputed every headline in this study from the stored ledgers, in a second language. Every one matched, including the primary. What it found instead were three defects in the analysis layer; a later independent verifier, reading the corrected document, found two more. Four of the five are the same mistake, a statistic computed over a sample that mixes the study’s one-book and three-book portfolios, on a scale the book count itself sets.

This section is at the front because one of the five was a fourfold inflation of a primary, and because three of them moved the published result in the direction the study is selling. An error that flatters the thing being measured is the one a reader is entitled to assume was left in on purpose.

1. moved toward the agent · the second primary

   C1b's tail was selected on terminal equity in absolute numeraire over a sample that mixes one-book and three-book portfolios. The two clusters do not overlap, so no three-book path could enter the tail: what was printed as CVaR(5%) of 800 paths was arithmetically CVaR(10%) of the 400 one-book paths, and the effect was inflated about fourfold.

   127.3 bpscorrected to29.9 bps

2. moved toward the agent · a renderer, and the test that guarded it

   The report rendered the engine-gaps table without its direction column, dropping, among others, the statement that one engine choice is worth the whole of a pre-registered loss in the agent's favour. The unit test that claimed to assert the disclosure only checked that the field was non-empty in the data structure.

   not printedcorrected toprinted, with a direction column

3. moved toward the agent · the cost secondaries

   The gas table was computed over book records and labelled per arm-path, so the largest transaction count contradicted the mean transaction count printed two rows above it.

   max 1 tx per arm-pathcorrected tomax 3 tx per arm-path

4. moved against the agent · the latency secondaries

   The agent's realised delay was pooled over three round-robin book offsets and printed against book 0's model band alone, which read as the agent missing its own latency model by six seconds. It does not: every book index is inside the band that book's own offset gives it.

   9.000–20.985 s against one bandcorrected toevery book index inside its own band

5. moved neither way · the workload secondaries

   Protects and transactions per run were averaged over two levels whose support differs, 0 or 1 on a one-book run, 0 to 3 on a three-book one. The pooled mean is set by how many cells hold three books, not by anything an arm did, and it describes neither level.

   1.584 protects per run, pooledcorrected to0.865 at 1 book, 2.303 at 3

The primary is unchanged, and that is the check that any of this is trustworthy. C1’s gaps are formed per path and divided by that path’s own opening equity, so a one-book and a three-book pair enter on one scale and none of the five defects can touch it. If it had moved, something else would be broken.

And the correction to the second primary makes this study’s own point harder rather than easier. The tail gap fell from 127.3 bps to 29.9 bps, which puts it much further below the median gap than the superseded figure was, the worst 5% of paths are exactly where the shock outruns the agent and it is liquidated too. The estimand was fixed because it was wrong. Which way it happened to point is not the reason.

Four separate readings each found one more instance of one class and each was patched, so what the study carries now is a mechanism rather than a fifth patch: every summary the report prints is rebuilt three times, pooled, one-book only, three-book only, and every numeric leaf compared. 251 statistics swept, 37 moved by the book-count axis and therefore required to declare themselves, 0 undeclared. An undeclared leaf stops the run before a report is produced.

1How much, and how sure

Two primaries were pre-registered and both were run, Holm-corrected over the family of two that this study actually contains rather than the family of four the plan lists for three tasks. Both tests are two-sided, which is the conservative choice against a pre-registered direction.

C1b’s p is the bootstrap’s own floor, not a measurement: 10,000 resamples were drawn and none crossed zero, so the honest statement is p < 1/(resamples + 1). Both gaps are in basis points of each path’s own opening equity, the endpoint is scale-free, so a one-book and a three-book portfolio enter on one scale.

The two numbers disagree by a factor of five and the disagreement is the finding. C1 says what protection is worth on a typical path where it binds; C1b says what it is worth in the worst 5% of outcomes, and there it is worth far less, because the paths in that tail are the ones where the shock outran the agent and it was liquidated too. An insurance product that is weakest exactly in the tail is a thing a buyer should be told, and it is why both were pre-registered.

Both are medians, and a median is only a summary of a population that has a middle. This one does not. Laid out path by path, the 800 paired gaps fall into two populations with an empty span between them: on 189 of them the two arms finish within 10 bps of each other, and on the rest the agent reaches the position and the modelled operator does not, or not in time. Nothing in this study lands between 0.3 bps and 32.2 bps, a two-order-of-magnitude hole, which is why the study’s 10 bps threshold does no work: any cut inside that hole gives the same three counts.

The 22 paths on which the agent finishes behind the modelled human are in here too, as the short bar left of zero, at most 1.9 bps each. Every one of them comes from wick-1block, the shock that drops and then recovers, which is also the shock behind every gap above 1,000 bps, all 144 of them. The largest wins in this study come from the same shock seen from two sides, and the full study takes that up.

2In which conditions

The pooled effect is an average over sixteen deliberately different worlds, and it hides almost everything. Split by cell, the agent is worth nothing at all on the slow grind during working hours, a 28% drift over eight hours leaves a human three hours of slack between crossing the trigger and health factor 1.0, and three hours is enough for anyone. On a 35% wick that prints for three seconds and stays down, it is worth four hundred basis points. On the wick that recovers, where being liquidated is the most expensive thing that happens anywhere in this study, it is worth fifty times that.

Read the two widest intervals rather than the two largest dots. On both wick-1block one-book cells the 95% interval reaches all the way down to zero, because a third of the paths in each are exact ties, the trough is long enough for the agent to act, but also long enough for a liquidator to reach the human, and when both are seized in the same state they finish identical. The largest effects in this study are also its least certain, and a forest plot is the form that says so without being asked.

3How it was measured

A simulation, and the label is on every figure. One exogenous price path is generated first, a shock shape plus idiosyncratic volatility of 2.4% a day, measured on BNB Chain mainnet, and then replayed across all five arms. On a three-book cell the three liquidators are replayed too. The market and the searchers are therefore shared, only an arm’s own actions move its own state, and every test on this page is paired for that reason. An unpaired test would throw away the variance reduction the whole design exists to get.

3.1The five arms, and which two the confirmatory test is between

A · confirmatory

simulation the keeper. Unattended, evented, one protect per block round-robin across books. Its decision-to-execution latency is a measured input and is never zero.

H1 · exploratory

modelled person an attentive operator. Always at the screen, no operational error, threshold rules, and a total event-to-broadcast delay FASTER than the agent's. Sequential across books.

H2 · confirmatory

modelled person a realistically available operator. A published availability schedule by hour, phone notification at night, sequential books. This is the arm the confirmatory test compares against.

H3 · exploratory

modelled person H2 with sleep forced over 23:00–07:00 CST regardless of the schedule, which separates always on from smart.

P · exploratory

simulation never repays, never closes. It measures whether intervening helps at all, and it is the column left to a reader who rejects a modelled human.

Three of the five are models of a person and none of them is a measurement of one. Every parameter of H1, H2 and H3 is published and none is fitted to this study’s data, which is the most that can be said for a modelled comparator. The arm that would have settled it, a runnable rule-based comparator a reviewer could execute themselves, was removed by the operator’s decision and has no substitute here. A reader who declines to accept a modelled baseline is left with the agent against passive, and that column is on this page rather than in a footnote. Each arm below ran all 800 paths.

The two survival columns are two different questions and only the first is comparable across this design. Book survival is positions kept over positions held, so the Books axis divides out of it. Portfolio survival is whether every book came through, which is a three-of-three event on half the design and a one-of-one on the other: the agent’s falls from 86.5% at one book to 56.8% at three while its book rate falls only from 86.5% to 76.8%, and the difference between those two falls is the axis rather than the policy. That is why every survival figure on this page is book-level and why the portfolio column appears here, beside its own denominator, and nowhere else. The intervals on all of them, 185 rates and 55 differences, are in exploratory.survivalUncertainty of T1-confirmatory.json.

Every arm finishes down about half. The market itself falls on every path in this study, so the return column is a property of the shocks and not of any policy, the agent ends at -47.4% and passive at -49.4%, and the whole finding lives in the difference between two arms on the same path rather than anywhere in that column. The two latency columns are also two different quantities and are never summed: at one book it is the time to protect the only book, at three the time to protect the last of three under the round-robin. Pooling them was the fifth defect in §0.

3.2The position, and the venue rule tuple that scored it

A venue is a tuple read from published parameters, not a label. Swapping the label without swapping the rule is not a variable, so the tuple is printed rather than named, and two of its rows are marked PROVISIONAL because they are Aave V3’s canonical rule adopted, not a parameter read from a named venue’s published configuration. Every arm starts from the identical position on every path, per book.

At that opening health factor the position crosses the trigger on a 14.8% fall and health factor 1.0 on a 25.9% fall. Every shock in the design is larger than the second of those, which is what makes passive die by construction, otherwise there would be no errand to run and the task would be vacuous. That is a design choice and it is stated as one: this study measures what protection is worth in worlds where the position was going to be liquidated, and it does not sample the world’s own frequency of those.

The venue record, as the study wrote it: LT 7000 bps from DeployTestnetV2.s.sol listReserve(wbnb, 700e8, 7000, 7500). Close factor and bonus PROVISIONAL, see comment; resolve at freeze into outputs/sim/venue-profiles.json.

3.3The race, which is the only mechanism in the study

Everything the agent is worth here comes out of one comparison of delays. The agent observes, decides and gets mined in 9 to 15 seconds, a 6 s floor measured from a real testnet dip-to-repay transaction, up to 6 s of jitter above it, and a 3 s block, and it is charged 3 s for each extra book, one protect per block, round-robin. The human is charged 45 s per extra book, sequential hands, which is the plan’s own figure and not this study’s. Against both of them stands a liquidator drawn LogNormal with a median of 15 s, labelled CALIBRATION-TARGET, because it is a target rather than a measurement, and because mainnet MEV is faster than it.

The agent’s per-book cost is deliberately an upper bound. A keeper batching consecutive nonces lands all three books in one block, which would make the marginal cost zero, and zero would decide the parallelism question by assumption rather than by measurement. So the agent is charged a whole block per book and every figure on this page understates it. The same rule runs the other way on the liquidator: the three books draw three independent searchers, so some of them draw a slow one a slow hand can still beat. Perfectly correlated liquidators would kill every book at once and hand the agent the axis for free.

All 1,267 uncensored book records the agent produced sit inside the band that book’s own offset gives it, so the latency model is not violated anywhere in the study. Printing this per book rather than pooled is §0’s fourth defect: pooled and set beside book 0’s band alone, the same records read as the agent missing its own model by six seconds, and that was the one correction in this study that took something back from the agent.

3.4What one protect costs, and what one seizure costs

A repayment retires debt rather than destroying it, so a protect costs the keeper fee plus one transaction of gas and nothing else. Almost the entire gap against a human is therefore an avoided seizure, and what a seizure is worth is not the liquidator’s discount as the ledger memos it, which is marked at the seizure price. What it takes out of the primary is marked at the terminal price, and the two coincide only when the price does not move afterwards.

That ratio crosses one, and the direction it crosses in is the reason the study quotes the harder number. On paths where the price ticks up after a seizure the memo understates what was lost, by as much as ×12.7; on paths where it keeps falling the memo overstates it. Quoting the memo would have been the flattering choice on most paths and the wrong one on all of them, so this study prints the terminal-equity cost and puts the memo beside it. Gas is charged in numeraire rather than in the chain’s own token, a disclosed compromise, and the bound on the error it introduces is computed from the run rather than assumed: 0.028 bps at worst, against effects measured in hundreds.

3.5The tests, and the correction applied over them

Wilcoxon signed-rank on the paired endpoint for C1; a paths-resampled bootstrap for C1b, with the pairing preserved inside every replicate, both arms are recomputed on the same resampled set, because resampling the arms independently would destroy the correlation the design exists to exploit. Both are two-sided against a pre-registered direction, which is the conservative choice and roughly doubles each p.

The family is two, and the plan pre-registered four. A p corrected over two is not comparable with a p corrected over four, so the page says which was applied rather than leaving a reader to assume: the other two tests belong to tasks that are not in this study, and correcting over tests that were never run would be a courtesy to nobody.

Separately from the two tests, the plan wrote 80 survival predictions before any path existed, and 14 of them are falsified at α = 0.05. Which ones is the finding.

Every falsification lands on a number the plan asserted without printing its arithmetic, and not one lands on a number re-derived from the model’s own frozen constants. That is the strongest internal-consistency evidence in the study, the simulator does what its published inputs say it should, and where it disagrees with the pre-registration it is the pre-registration’s prose that was wrong. The two kinds are separate columns in the artefact so the claim can be checked rather than taken, and where a target was re-derived the planned figure it replaced is printed beside it. Re-deriving a prediction from published inputs is not the same act as tuning an input until a prediction comes true.

3.6Seeds, and how to run it yourself

Seed block {1000…1799}, fifty contiguous seeds per cell, bootstrap seed 90,000 at 10,000 resamples. The exogenous streams, the price path and each book’s liquidator, exclude the arm from their key, which is what makes every arm face the same market and the same searchers; the arm-internal streams include the arm name, so adding or removing an arm perturbs no other arm’s draw. Determinism is proved by regenerating the whole population from the same seeds and comparing digests, and again by re-running the engine on a sample of paths and comparing serialised results byte for byte, which covers the arm-internal streams a path digest cannot see.

The corrections in §0 were applied with the study’s --rescore step, which recomputes the manifest’s own digests over the files as they sit on disk and refuses to proceed if either has moved. A defect in an estimator, a renderer or a label is not a defect in the runs, and re-running the runs to fix one would have discarded the byte-identical reproduction an independent replay had already established over them.

4What this does not show

None of the following is a hedge added at the end. Each one is a decision that was taken before the run and is on the page because a reader is entitled to attack it.

1. The human is a model, not a person. Every parameter of the three human arms is published and none is fitted, but a reader who declines a modelled baseline should reject those three columns and read the agent against passive, 164.6 bps. The runnable non-agent comparator that would have settled it was removed and has no substitute.
2. One venue rule tuple, Aave-V3-shaped, and its close factor is that rule’s canonical form adopted rather than read from a named venue’s published configuration. Not Venus, not Lista, not three integrations.
3. The liquidator is testnet-shaped at a median of 15 s. Mainnet MEV is faster, and a faster liquidator is a pre-registered way for the agent to lose that this study does not run.
4. The books axis has two levels, one and three. The interpolation check at two books was not run and nothing here supports an extrapolation to a large book count. What is established is the sign and the mechanism at three.
5. Only one of the four shocks rewards holding. The mixture Figure 2 shows is one the design chose, not one the world produced. Converting these basis points into an expectation needs the real-world weight between the fall persists and the market repairs itself, and this study does not estimate it.
6. The pre-registration was not hashed. What survives of it is the discipline: predictions written into the code before the run, printed beside what was observed, and reported as failed when they fail. A reader who wants the stronger guarantee should note that this is not it.
7. Interest is off, because the mock pool accrues none. Any venue where it is not zero moves every arm here, since all of them carry debt for the whole path.
8. One contrast is unpaired on the thing it tests. The two realistic humans draw independent availability schedules, so the comparison that isolates a sleep window has far less power than its sample size suggests. It is reported as a null this design cannot resolve rather than as evidence that sleep does not matter.
9. Testnet units are not dollars and are never converted to any. Nothing on this page is a claim about money made or lost on BNB Chain mainnet. Only chains 56 and 97 appear as sources anywhere in the study.

Every figure on this page is drawn from a module compiled at build time out of T1-confirmatory.json and the run ledgers, and rendered on the server; scrolling changes which marks are lit and never what they say, and the numbers in the markup are the numbers in the artefact.