The ruling · Register of 93 false claims · The paper · English map of the record · Limits · Cite

Note

Status: closed. The research stopped on 9 September 2026 at commit 93f3e35 of the private repository, and this copy was published on 22 September 2026. The research is frozen. Apart from one test fix recorded in MANIFEST.md, only the publication files listed there change.

I looked for a trading edge in crypto perpetual futures (futures contracts with no expiry date) between June and September 2026 and did not find one. The programme's headline estimate was +18.24 basis points (hundredths of a percent) per filled order, on the LIT perpetual of Lighter, a decentralised exchange. On 26 August 2026 it was invalidated as a confirmatory result. On 7 September it was withdrawn as evidence of an edge: a post hoc re-analysis also counted as filled the orders that vanished while trades took place both at their price and beyond it, and the estimate fell to +2.575 basis points. The programme ran to 457 commits. The withdrawal needed a rebuild of 79 days of order-book data from re-acquired files, and the rebuild matched the original run on all 632 recorded numbers.

This repository is the laboratory as it stood when it closed, without 171 of its files (see About this copy): the code that survived the final cleanup, 476 tests, six checks that run on every push, and the written record of what was tried and decided, mostly in Spanish. It contains no market data, so the result cannot be re-run from here; the tests and the checks can. Most of the code was written and audited by AI agents under a protocol I set up, and How it was built says who did what. The companion repository ninety-three-wrong-claims holds a paper on the register of false claims the programme kept.

The question

Can a small participant with no speed advantage earn a lasting edge in crypto perpetual futures?

The main lines of inquiry were these eight, listed with the verdict the record gives each; the ledger records others, among them a hypothesis on ATOM funding, falsified and closed on 22 June (LEDGER.md 374). The first six were on Binance, and only one of them, row 3, was tested on held-out data: days set aside and not looked at until the final test (LEDGER.md 30).

#	line of inquiry	data	how it ended	source
1	three service premia: provision, execution against order flow, funding carry	Binance recordings and downloaded history	three screens, all negative	LEDGER.md 31
2	cross-venue lead, Binance to Bybit	Binance, Bybit	a lead of about 50 ms: a speed game, archived	LEDGER.md 32
3	fading extreme order-flow imbalance	eight thin Binance perpetuals, held-out	−10.71 bps per event after costs, t −26	LEDGER.md 33
4	liquidation cascades	Binance perpetuals	inconclusive: the condition occurred in 0.57 % of cases, under the 1 % floor	LEDGER.md 36
5	cross-asset lead-lag	Binance perpetuals	null: liquid coins did not lead thin ones; the lead statistic D was −0.026, pointing the other way	LEDGER.md 37
6	the window after a new listing	new Binance listings	closed on a low prior: market makers were in the book from the first minute	LEDGER.md 39
7	passive quoting on Hyperliquid	one day of Hyperliquid books	failed its first screen: 6 candidate symbols against a floor of 30 fixed in advance	ARCO.md 40
8	patient resting orders on Lighter	Lighter order-level archive, 79 days	+18.24 bps, formally inconclusive; withdrawn on 7 September	LEDGER.md 1681

Row 8 covers several steps. On Lighter the programme first dropped a taker idea (trading against resting orders) when Lighter's order flow proved uninformative. It then took a maker idea (resting orders and waiting to be filled) through a screen, a simulated maker, a pilot on real resting orders, an atlas of the wallets behind them and the confirmation test that produced the +18.24. The attack letter, in which the programme set out on 5 July the strongest objections to its own result, counts about ten attempts in all and says no statistical correction spans them (CARTA_DE_ATAQUE.md 12).

How a fill was inferred

Figure 1. Five cases a resting order can go through between two snapshots, and which of them each rule counts as a fill. The July code put rows 3 and 4 in one class, through; the flag that splits them was added on 10 July. Drawn by hand.

The programme never placed an order. It followed other participants' orders through a commercial archive of Lighter's order book and matched them with the trades that took place between snapshots.

venue and instrument	Lighter LIT perpetual; standard accounts paid no trading fees at the time
order data	order-level snapshots about 2.9 minutes apart, from a commercial archive
trade data	trade files without order ids, so a fill had to be inferred
sample	79 days between 6 March and 29 June 2026; the other days of the window went to the pilot or were excluded
cohort	slow orders: seen in an earlier snapshot, so they had rested at least one full interval
regime	the most volatile third of the sample, by trailing 20-minute volatility
horizons	1, 5, 10, 25 and 60 s; 5 s was the primary horizon of the re-analysis, 25 s the headline cell
reference price	the nearest trade on Binance's LIT perpetual; the preregistration asked for the mid price

Each fill was scored by its markout: how far the price moved after the fill, in basis points, relative to the order's price p. A positive markout means the move favoured the patient trader; a negative one, that it went against them. In symbols it is side × (reference − p) / p × 10,000, with side +1 for a bid and −1 for an ask. The reference is taken one horizon after the fill time. The fill time is the first trade at p or, for a through event (a trade beyond p), the first trade beyond it: below p for a bid, above it for an ask.

Over the 79 days there were 132,876 slow-order events with a trade at or beyond the order's price; cancellations are not among them. The July rule counted 13,368 of them as fills (10.1 %) and the wider rule 115,302 (86.8 %).

Why the rule mattered. The July rule set aside the orders in the third row of Figure 1, called through-with-at below: an order that vanished in an interval where trades took place both at its price and beyond it. On 6 July, one day after the +18.24, AI agents re-auditing the statistical code found why that matters, and a passage naming the mechanism was added to the attack letter. A trade beyond an order's price goes with a later move against the order, so leaving those orders out raises the mean. The passage made a sensitivity test mandatory, and predicted that the affected orders would be a minority whose return would shrink the mean without reliably turning it negative (CARTA_DE_ATAQUE.md 100-115). A post hoc version of that test, on the 79 confirmation days, ran on 7 September. It added 101,934 orders, three quarters of the slow-order events rather than a minority, and the 25-second mean fell by 86 % to +2.575.

What happened to the positive result

Figure 2. Left: each day's mean markout at 25 seconds under the two rules, days in date order. 79 of the 116 calendar days entered the test; the others had gone to the pilot or were excluded, and the checklist review found the excluded days busier. 1 May, at +223, is off the scale; it is the day with the one zero-price reference at 25 seconds. Right: the mean at each horizon, each on its own days; at 5 seconds three invalid references pull the July-rule mean from +10.38 to +2.61. Post hoc diagnostic, run once on 7 September 2026. Data: d1_per_day_25s.csv, from the run whose hash is recorded in the ruling; drawn by make_figures.py and checked against its data on every push.

On 5 July 2026 the preregistered confirmation test (its thresholds were committed to git before the data was read) gave +18.24 bps at 25 seconds for slow orders in the most volatile third of the sample (t +6.69, 79 days). The same test required the 5-second horizon to agree, and it did not (+2.61, t +0.29), so its formal verdict was inconclusive (LEDGER.md 1681). The estimate was withdrawn in three steps.

1. On 24 August an adversarial checklist, frozen on 13 June when the finding did not yet exist, was run against it for the first time, by AI agents working one threat each. It returned DESTROYED, with all 8 of its threats applying (report, checklist).
2. On 26 August the markout pipeline was found to accept reference prices of zero. Dropping the one affected fill from the headline cell (slow orders, most volatile third, 25 seconds) moved the estimate from +18.24 to +15.60. It also cut the effective number of independent days from 79 to 14.7, because the outlier day had been hiding the correlation between consecutive days' means. The incident report invalidated the +18.24 as a confirmatory result and put no other figure in its place (incident).
3. On 7 September the 79 days were rebuilt, the July number came back exactly, and the re-analysis counted the through-with-at orders. Among slow-order events with a trade at or beyond the order's price, the share admitted as fills went from 10.1 % to 86.8 %, and the 25-second mean fell to +2.575.

horizon and reference	July rule (bps)	wider rule (bps)	paired difference (bps)	descriptive t
5 s, July reference	+2.6101	−1.0843	−3.6944	−0.5505
5 s, validated reference	+10.3795	−0.0976	−10.4772	−13.4025
25 s, July reference	+18.2368	+2.5750	−15.6618	−6.9043
25 s, validated reference	+15.6036	+1.4988	−14.1048	−9.8270

The validated reference drops events whose reference or fill price is not a finite positive number, or whose markout reaches 9,000 bps in absolute value, which is the correction for the 26 August defect. The t values are descriptive: no correction for multiple comparisons is applied.

The ruling withdrew the +18.24 as evidence of a tradable edge, kept the +2.575 on the record without calling it profit, and left in place the standing halt on production and trading. The re-analysis had fixed 5 seconds as its primary horizon. There the mean went from +2.61 to −1.08, but the paired difference has t −0.55, so the withdrawal rests on the 25-second rows. The re-analysis was specified after the +18.24 was known. Its reference price was the nearest Binance trade, where the preregistration had asked for the mid price. And fast orders also scored +11.78 bps under the July rule, so the effect was never specific to slow orders.

Timeline

Figure 3. Commits per week in the private repository, with the events below numbered on the time axis. Data: commits_per_week.csv and timeline.csv; drawn by make_figures.py.

#	date	event	source
1	12 Jun	Repository created; Binance collector starts recording	private history: first commit c6f1c9e; the collector start is logged in a private build diary
2	13 Jun	Adversarial checklist written and frozen, never edited again	quant-reviewer.md
3	1 Jul	Binance lines 1-6 closed; no edge for a slow trader	LEDGER.md 23
4	5 Jul	Lighter LIT: +18.24 bps at 25 s, formally inconclusive	PREREG_CONFIRMACION_LIT.md; LEDGER.md 1681
5	6 Jul	Attack letter names the through-with-at weakness	CARTA_DE_ATAQUE.md 100
6	6 Jul–6 Sep	Ledger silent for two months, across 248 commits	LEDGER.md 1820
7	17 Jul	A transactional log for results begins; 59 commits, deleted 9 Sep	private history: eco_libro_tx.py, 7ff4822 to d15cf94, deleted in ac165b9; ARCHIVO.md 107
8	13 Aug–22 Aug	Unpaid hosting bill: no capture from 13 Aug 08:57 to 22 Aug 20:15	AUDITORIA_DEL_METODO.md 140
9	24 Aug	June checklist run on the finding: destroyed, 8 of 8	AUDITORIA_QUANT_REVIEWER_2026-08-24.md 28
10	26 Aug	Zero-price references found in the markout pipeline	INCIDENTE_2026-08-26_REF_INVALIDA.md
11	7 Sep	79 days rebuilt from new data; 632 of 632 numbers match	LEDGER.md 2249
12	7 Sep	Wider fill rule: +18.24 falls to +2.575; withdrawn	DICTAMEN_MESA_2026-09-07_I5_D1_79.md 57
13	9 Sep	Code cut to 5,933 lines on 8-9 Sep; closing ruling	DICTAMEN_MESA_2026-09-09_V2_CIERRE.md 31

The checks

Figure 4. The six steps of ci.yml, each with the failure or review that added it. Drawn by hand. None of these steps looks at an estimate: the +18.24 was withdrawn by a checklist review, an incident report and a post hoc re-analysis, none of which runs in CI.

git clone https://github.com/Lostmanu/quant-system
cd quant-system/quant-system-ingesta/qs
pip install -r requirements-test.txt -c requirements-lock.txt

python -m pytest tests/ -q                    # 476 passed
python tools/mutacion_ref_valida.py           # 49 deliberate breaks, each caught by its named test
python tools/guardia_documental.py            # preregistration dates, spec against code
python tools/guardia_completitud.py           # numpy I/O only where it is wrapped
python tools/registro_sonda.py                # probe registry (empty)
python tools/recuento_auditoria.py --check    # register count is current

In GitHub Actions runs 35822218188 and 35789838535 (Ubuntu, Python 3.12.14) the whole job took 68 and 72 seconds: about 13 for installing, 13 for the 476 tests, 35 for the mutation harness and under a second for each of the four guards.

Each of the harness's 49 rows disables a guarded check, in the guard tools themselves or in analysis/ and ingestion/, and passes only if a named test fails with the failure message declared for it. A red test is not enough, because a test can fail for an unrelated reason. On its first run the harness found two defects in the shared code it was built on: the regular expression that parsed pytest failures missed parametrised test ids, and the function that judged the results counted them as collateral failures (ci.yml 34).

What the harness prints (CI run 35822218188, excerpt)

49 mutaciones · copia nueva para cada una · 67 ficheros vigilados
pre-vuelo sobre las 49 filas: OK
  linea base (sin mutar): 111 passed, 153 deselected in 1.19s
  linea base VERDE y valida: 111/55 tests, rc=0
  los 55 tests declarados EXISTEN en la coleccion
  ROJO ✓                                 R1 quitar el rechazo de NO FINITA (NaN / +-inf pasan a ser markouts)
                                         asigna_la_clase_correcta→assert
  ...
  ROJO ✓                                 CENSO: aceptar la compuerta dentro de una RAMA (un `except` que se la traga)
                                         compuerta_que_NO_domina→assert
  ...
49/49 comprobaciones MUERDEN con su marcador o fragmento declarado
  LIMITES, y son los mismos que los del arnes hermano:
   · acredita la SENSIBILIDAD de los tests, NO la CORRECCION del codigo.
   · cada mutacion corre con `-k` sobre SUS tests declarados: un test no seleccionado que
     la mutacion rompiera seria invisible aqui. No se afirma que no exista.

In English: 49 mutations, each on a fresh copy of the tree; the unmutated baseline passes; every row turns its named tests red with the declared message. The harness then states its own limits: it shows that the tests are sensitive to these changes, not that the code is correct, and a test outside a row's selection that the mutation would break is not seen.

One of the six steps cannot fail at present: the probe registry is empty, and an empty registry passes. Half of the documentation guard is in a similar position. Its chronology check exempts six of the seven preregistrations as older than its rule, and the seventh declares two artefacts that do not exist yet, so there is nothing to date. Its other half checks that four rules declared as implemented still name a function or constant in ref_valida.py.

The code, and what was removed

Figure 5. Lines of Python (wc -l) under analysis/ with tools/, and under tests/, at each commit on the first-parent line of the private repository (455 of its 457 commits; the other two came in through its one merge). Data: code_history.csv; drawn by make_figures.py.

The cleanup of 8 and 9 September retired the modules that only closed studies used. analysis/ and tools/ went from 36,946 to 5,933 lines, and the CI suite from 2,023 tests to 476 (closing ruling). Each removed file is listed in ARCHIVO.md with the commit that recovers it.

What is left in analysis/ has three parts. The reference validator, with its artefact and lane layers, is the code the mutation harness guards. A set of tested libraries covers purged and embargoed cross-validation, the probabilistic and deflated Sharpe ratios (PSR, DSR), the probability of backtest overfitting (PBO), multiple-testing haircuts, bars, toxicity, regimes, cross-section and venue adapters; most of it is no longer called by any code outside the tests. Four download scripts have no tests.

What is where

.
├── README.md                    this page
├── ARCHIVO.md                   files retired on 8-9 Sep, with the commit that recovers each
├── EXCLUIDO.md                  the 171 original files not copied here, by category
├── MANIFEST.md                  SHA-256 of the 173 copied files
├── CITATION.cff
├── .claude/agents/
│   └── quant-reviewer.md        the adversarial checklist frozen on 13 June
├── .github/workflows/
│   ├── ci.yml                   the six checks
│   └── figures.yml              the charts against their data
├── assets/                      the banner and Figures 1-5; figures/ holds the script and CSVs
├── docs/                        the record: ledger, preregistrations, rulings, audits
│   ├── README.md                an English map of every document, with a glossary
│   ├── testigos/                attack letter, sanitisation act, June-July story, A11 audit
│   └── inventarios/             machine-readable inventories of the cleanup
├── infra/lighter_collector/
│   └── collector.py             the Lighter collector that ran on the server
└── quant-system-ingesta/qs/     the original layout, kept so hashes and links resolve
    ├── ingestion/               the Binance collector: feed, order book, writer, audit
    ├── analysis/                reference validation, artefacts, lanes, libraries
    ├── tools/                   guards, mutation harness, pre-flight, register counter
    ├── tests/                   476 tests
    ├── config/                  instrument list
    └── docs/                    work orders and run receipts the record links to

ARCHIVO, EXCLUIDO and MANIFEST are in Spanish, like the record.

Reading the record

The record is in Spanish. docs/README.md maps every document in English and has a glossary of the terms that appear in file names. A reader with an hour could take them in this order:

1. RESEARCH_BRIEF.md, in English: what the programme set out to do.
2. testigos/ARCO.md: the story of June and July, chapter by chapter, with commits.
3. PREREG_CONFIRMACION_LIT.md: the test that produced the +18.24.
4. testigos/CARTA_DE_ATAQUE.md: the eleven strongest attacks on the programme, written by the programme; docs/README.md has an English summary.
5. AUDITORIA_QUANT_REVIEWER_2026-08-24.md: the checklist review.
6. DICTAMEN_MESA_2026-09-07_I5_D1_79.md: the re-analysis and the ruling.
7. AUDITORIA_DEL_METODO.md: the register of 93 false claims.

Lessons, with dates

1. The adversarial checklist was frozen on 13 June and first run against the finding on 24 August, ten weeks after it was written and seven after the finding. It returned DESTROYED (report).
2. The weakness that decided the withdrawal was written down on 6 July in the programme's own attack letter, and a post hoc version of the test it asked for ran on 7 September (CARTA_DE_ATAQUE.md 100).
3. New checks were once placed after a CI step that had been red for three days, which meant GitHub would have skipped them without reporting. Every verification step now runs unless the job is cancelled (ci.yml 21).
4. Data capture stopped from 13 to 22 August because a hosting bill went unpaid (register, D-09). Two of the five commits made in those ten days went into a results log that had already taken 56 commits since 17 July; the log was deleted in September.
5. The programme was started to build a bot that would add 100-200 € a month to my salary. The calculation of what that goal required, in notional, fills and share of the symbol's volume at each level of net edge, was written on 25 August, about eleven weeks in, in feasibility_objetivo.py; its docstring says it should have been the programme's first line of work.

About this copy

This is a snapshot without history of the original repository at commit 93f3e35, which stays private. It has 173 of that repository's 344 files; EXCLUIDO.md lists the other 171 by category, and MANIFEST.md gives the SHA-256 of every file as exported. Server addresses, account names, key locations and local paths were replaced by markers such as <VPS_IP> and <USER_HOME> in 35 files, following a sanitisation act written on 5 July 2026 (SANEAMIENTO.md). One link, to a custody folder on my machine, points outside the repository on purpose.

One of the 173 copied files has changed since the snapshot: two collector tests read the real disk and failed on any machine whose temporary folder was more than 85 % full, so a fixture now reports the disk to them as half full. Hashes and seals quoted inside the record refer to the private originals, so the 35 files with markers will not match them byte for byte.

Figures 2, 3 and 5 are drawn from CSV files by assets/figures/make_figures.py, and a second workflow fails if a published chart no longer matches its data. assets/figures/data/README.md says where each number comes from. The banner and Figures 1 and 4 are drawn by hand.

How it was built

I chose the questions, paid for the data and signed the decisions. The code was written by AI agents. Most commits carry the signature of Claude Code (Anthropic), 383 of 457. Codex (OpenAI) appears in the record from late July, and by September it implemented while Claude Code audited and ran the long jobs. A review role, called la mesa in the documents, issued the rulings. Another AI model filled it, Codex in the September documents, and I signed the result, so the rulings are not independent reviews. The agents account for part of the register of 93 false claims; the record does not attribute its entries to a named person or system, and it records who or what caught each one.

Limits

1. No edge was found, and nothing here supports a strategy. The repository documents how the search was run.
2. The trading result cannot be reproduced from this repository, which contains no market data.
3. The re-analysis that withdrew the estimate was post hoc, and the reference price in every branch of it departs from the preregistration.
4. There is one laboratory and one author, no second implementation and no control group.
5. No one outside the programme has reviewed the rulings.

Licence

Code under Apache 2.0, documents under CC BY 4.0. To cite it, use CITATION.cff.

Manuel Beardo Campo · LinkedIn · the paper