BL-SD-000: The Kill Ledger

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Every hypothesis the programme tested, with the measurement that settled it.

Every row resolves to evidence. Eighteen resolve to a machine-readable artifact in BL-A-001; the remaining thirty-three resolve to a dated document in BL-A-006. The two are cited with different prefixes because they are not equally strong, and a reader is entitled to see which one a figure stands on.

Both sets have been read back against their sources: ninety-eight claims checked, ninety-two confirmed, six corrected or flagged. Section 7 records all six.

---

How to read this ledger

Fifty-one entries, covering roughly thirty distinct strategy families. Several families were tested more than once, either because a first instrument proved unfit or because a failed variant suggested a specific successor worth pre-registering. Those appear as separate rows, because each one consumed a pre-registration and produced its own verdict. Four entries are not hypotheses at all: three record a fact that was confirmed and never monetised, and one records the single component that reached production.

Verdict uses the fixed vocabulary from BL-FM-003 section 7. FALSIFIED means a pre-registered kill criterion was met with the statistical power to support it. CLOSED means the mechanism measured real and a declared envelope parameter blocked it. NOT PROVEN means the interval spans the threshold and nothing was concluded, which is not the same thing as a kill and is never recorded as one.

Axis names what actually did the killing, so that a falsification on one axis does not leak into axes that were never tested.

Transferability applies to FALSIFIED and CLOSED rows only. UNIVERSAL means no envelope substitution rescues the result. ENVELOPE-BOUND means the binding term was a parameter of our operating envelope rather than a property of the market, and a reader running different parameters should expect a different answer. Rows marked NOT PROVEN or OPEN carry a dash, because an unresolved hypothesis has nothing to transfer yet.

If you are reading this ledger to decide what is worth re-testing on your own infrastructure, the ENVELOPE-BOUND rows are the entire answer. Section 5 lists them on their own.

All returns are in basis points, net of the round-trip cost stated in the dossier unless the row says gross. The operating envelope those costs come from is declared in BL-FM-002.

---

1. Programme I, Bybit

---

2. Programme I, Hyperliquid

---

3. Programme II

---

4. What the ledger adds up to

Thirty-seven hypotheses reached a settled verdict. Twenty-seven were killed on their merits with the statistical power to support the kill. Ten measured real and were closed by a declared envelope parameter.

Five returned intervals spanning their threshold and are recorded as unresolved rather than dead. That distinction is enforced throughout the record, because labelling an underpowered result a falsification poisons every decision that later inherits the label, and the programme did exactly that once before the rule was written down.

By axis, across the thirty-seven settled rows:

Eighteen deaths on the idea itself, and twenty once the two data-defect cases are re-tested and die on the same axis, is the number worth sitting with. The dominant failure mode across four months was not execution, not cost, and not infrastructure. It was that the mechanism proposed did not exist in the data.

Cost accounts for seven. That is the second-largest group and it is the one an outside reader should look at hardest, because cost is an operator parameter rather than a market fact.

Only one hypothesis in the entire programme died on Mechanism inverted out of sample, and it came last. H-X4 is the only case where a mechanism held through research, held under conditioning, held against a permutation null at 5.83 sigma, and then reversed sign on sealed held-out data. That is the failure mode every other gate in the programme exists to catch earlier, and it is documented in full because it got through.

Twenty-four of the settled verdicts carry to any operator. Thirteen do not, and those thirteen are where an outside reader should spend their attention.

---

5. The envelope-bound rows

Thirteen hypotheses, grouped into eleven rows below, measured a real effect. What stopped each one was a parameter of our operating envelope rather than a property of the market. An operator with different parameters should expect different answers here, and these are the rows worth re-testing.

Three of these deserve a specific note.

IB1 is the cleanest case in the record. The signal is real and well powered: aggregated informed-flow bursts produce +3.0 basis points of drift at 30 seconds with a sign-permutation p-value of zero on 486 events. It failed because a 9.0 basis point taker round trip consumes it three times over. The arithmetic is not subtle, and neither is the implication for an operator quoting passively at 3.0.

WC1 looks like a falsification and is not one. Whale state changes carried no measurable drift at a 300-second horizon, but our detection latency was 92.8 seconds at the median. The honest statement is that we could not observe the event early enough to know whether the information was there, and that is a cadence result rather than a market result.

H-R1 was closed on cost at an assumed 22.0 basis point round trip. Tested at 11.0 the mean turns positive at +2.3 with an interval that still spans zero. That is not an edge, but it is not the same finding as the headline, and the dossier reports both.

---

6. What remains open

B1 and BR1 are the same mechanism found independently on two venues: reversion after forced liquidation flow exhausts, entered passively rather than by crossing. Neither reached the power to settle. Both are held at RESTRICTED classification and appear here by identifier and status only.

---

7. Corrections against source

This ledger was checked in two passes. The first read every row that resolves to a machine-readable artifact back against that artifact. The second read the remaining thirty-three rows back against the dated document that states each figure, ninety-eight individual claims in all.

Ninety-two claims confirmed. Six did not, and all six are recorded here rather than quietly replaced, because the discipline is worth more than the appearance of having got it right the first time.

7.1 Against artifacts

Neither correction moves the verdict. D3 was closed on adverse selection at the second gate, not on the spread census at the first, and a median of 5.26 against a maker round trip still leaves a majority of the universe nominally positive before adverse selection is applied.

One further caveat belongs on the D3 row and was not in the narrative at all. The artifact carries the field maker_fee_status marked as unverified at the time of measurement, with the fee reported as a grid rather than a confirmed schedule. The spread census therefore establishes what the universe quotes, not what a specific operator would net against a confirmed fee. The dossier states this in its envelope section.

The same pass confirmed a claim in the earlier internal record that was wrong in the other direction. Three Hyperliquid raw verification artifacts were described as empty. Two exist, both parse, and both carry data. The record follows the artifacts.

7.2 Against documents

Four further corrections came out of reading the prose-backed rows against their sources.

None moves a verdict. Three are transcription drift of the kind section 2 of BL-A-006 predicts, and the fourth is a citation pointing one document to the left of the one that carries the number.

The OI-ACC correction is the one worth pausing on. A ledger row asserted a short trade in a strategy whose design document says, in as many words, that there is no short side in the first version. Nothing measured that short, and the record should not have implied one existed.

Two claims could not be located in the source assigned to them and are recorded as such rather than repaired. The OI-ACC sample of 107 appears elsewhere in the archive but not in either document assigned to that row. The D5b figures are correct and were cited to the wrong report.

Two more are ambiguous in the source rather than wrong. The M3B sample is described as the same event set as its predecessor, and the realised counts differ by one, most likely an event accruing between two same-day runs. The record leaves both statements as the sources give them.

7.3 Against artifacts again, during the Volume V publication pass

Publishing the strategy sources meant reading each cell of a results table back against the artifact behind it, one cell at a time, rather than reading the row. That found four defects in a single table in BL-SD-039, the sizing-basis comparison, and none anywhere else.

The verdict does not move. Every corrected interval still spans zero at the baseline cost, and H-R1 died on cost rather than on any of these cells.

What is worth carrying forward is where the defects clustered. That table is the one BL-SD-039 describes as having changed how every later dossier reports a PnL table, and it was assembled by hand from three artifacts produced by two instruments on two different samples. Every other table in that dossier came from one artifact and every one of them was clean. The failure mode is not carelessness with numbers; it is a table whose columns quietly span more than one measurement, which is exactly the shape R-25 was written to catch and the shape a row-level read does not catch.

Two of the four were empty cells filled in rather than left empty. A table with a value in every cell reads as more complete than the measurement behind it, and that is a presentation defect with a statistical consequence.

7.4 Test counts, and a coverage gap the counts revealed

Later batches of the same pass read the test-count claims in each dossier back against the suite each named. Two claims had no source and nine pinned cases had gone missing from published suites without being recorded.

The third row is the one that led somewhere. Checking it showed the source suite does hold 19 cases and the published suite held 17, of which one was a new guard. Sixteen of the nineteen had survived. The same check on the stop primitive found 22 published against 26 in the source, four of them new guards.

Nine source cases had therefore been dropped across two published suites with nothing recorded, against a porting standard that requires a dropped case to be named and justified. All nine are restored. All nine pass against the published implementations unchanged, so every one of them closed a coverage gap rather than a defect.

That distinction is worth holding onto rather than filing as a near miss. The implementations were correct, so no number in this record moves. What was missing was the evidence that they were correct, and on this record those are not the same thing: R-25 makes a verdict valid only when a tested instrument stands behind it, and a case that is quietly absent is indistinguishable from a case that never existed.

Two of the nine are worth naming individually because nothing else covered them. A portfolio return with a negative short-leg mean, where every other case in that file used two positive legs and a sign error on the short side would have been invisible. And a bracket stop that gaps through its level, where the fill has to land at the worse open and still pay slippage on top.

The pattern across all four subsections is one thing said four ways. Numbers in prose drift, and a count of tests is a number in prose. It reads as bookkeeping rather than as a result, which is exactly why nobody checks it.

---

8. Evidence map

Rows resolving to a machine-readable artifact in BL-A-001.

Supporting artifacts not tied to a single row: E-0006 and E-0039 carry the information-decay maps behind the rule that a coefficient may characterise but never issue a verdict. E-0038 is the substrate inventory cited in BL-FM-002. E-0040, E-0041, E-0049, and E-0050 carry the rebalance-anchor and reconciliation work. E-0042 and E-0043 carry the eligibility overlay. E-0052 carries out-of-sample stability across channels. E-0055 through E-0057 are the venue verification pulls that established the Hyperliquid substrate.

The remaining thirty-three rows resolve to dated documents rather than to emitted JSON, and are registered as D-001 through D-030 in BL-A-006, which maps each document to the rows it carries.

Prose is where numbers drift, so those thirty-three were read back claim by claim against the documents that state them. The pass found four corrections and two claims it could not locate in the source assigned to them, against ninety-two confirmations. Section 7.2 records them.

That rate is worth reading in both directions. Six exceptions in ninety-eight claims is a low error rate for figures transcribed by hand into prose over four months. It is also six more than an artifact-backed row would have produced, and it is the reason the two evidence classes are kept visibly apart rather than merged into one register.

---

Copyright (c) 2026 Bijective Labs. All rights reserved.