BL-IS-014: Substrate Inventory

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-IS-014: Substrate Inventory

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

1. What it measures

What the programme physically holds on disk: file counts, byte counts, date coverage and column names, per stream, read from the files themselves rather than from any manifest that claims to describe them.

A manifest is a claim. A byte count is a fact. Sixteen streams are scanned this way, from the Bybit historical klines through the Hyperliquid position recorder to the lead-lag event capture, and the instrument reports what it can verify from a sample of files rather than an assertion about what was supposed to have been written.

2. Why it exists

The immediate cause was a measurement that had already been made. A cross-sectional carry construction was evaluated as an apparently unopened question and returned roughly +14 percent annualised at a Sharpe near 1.0, a result that looked like a new finding. It was not. Cash-and-carry had already been evaluated and retired months earlier with the structural verdict that its yield is linear rather than compounding [BL-SD-016]. The two measurements differed mechanically, one on a perpetual-versus-perpetual construction and one on the classic spot-versus-perpetual form, but at the level of class they were the same question, and the second measurement's headline confirmed the first verdict rather than opening anything new [BL-P2-003 section 2.3].

The deeper cause was that the programme's own knowledge was spread across roughly nineteen pre-registrations, roughly eleven verdict reports, and dozens of decision documents, with no map connecting them. Below a certain size, remembering what has already been tested is cheaper than re-testing it. Past that size the reverse is true, and re-testing becomes the path of least resistance simply because running a measurement is easier than searching a document tree for whether someone already ran it. This instrument is the first artifact in that repair: a fact base a researcher can query in seconds before committing to a new measurement, independent of whether anyone remembers the prior one.

Classification of what a stream implies for a hypothesis class lives in a separate document, deliberately, so that the byte counts here are never contaminated by an opinion about what they mean.

3. Interface

scan(pattern) -> stream_report | None
  pattern   a glob, recursive where the stream is Hive-partitioned by date
            and coin
  probe     three files sampled per stream, first, middle and last by sorted
            path, rather than every file, so the scan completes in seconds
            and can be rerun at any time without cost
 
returns:
  n_files, bytes, mb
  date_lo, date_hi   from a recognised timestamp column where one exists,
                     else from the date= partition segment of the file path
  rows_in_probe      row count of the sampled files only, not the stream total
  columns            column names of the first successfully read file

Byte and file counts are exact, computed over every file in the stream. Row counts and date ranges are estimates drawn from the probe and are reported as such; an exact row count would require reading every file in a corpus this scan exists to summarise in seconds.

4. What the scan found

Total substrate measured at 8,195.8 MB across the sixteen streams scanned, spanning venue histories from 2021-10-12 on the Bybit side to a three-week Hyperliquid capture window [E-0038]. The two largest single streams are the Hyperliquid position recorder at 2,921.4 MB and the lead-lag event capture at 817.1 MB.

The research panel actually consumed by the Programme II measurements is a narrower slice of that total: 12,970,563 rows across 154 instruments and four streams, occupying 314 MB. That figure and the retry count below it are reported by the harvest step itself rather than recomputed by this scan, and the two are read together in [BL-TR-002 section 1].

5. The retry count is the load-bearing number

The harvest that built the research panel completed with zero retries.

That figure belongs beside the byte count rather than in a build log read once and discarded. A retry means the venue rate-limited a request, and rate limiting is not distributed evenly through time: it concentrates during high market activity, which is precisely the regime most hypotheses in this record care about. A substrate assembled with retries therefore has holes whose placement correlates with the phenomenon under study, and no instrument downstream of the harvest, including this one, can detect that correlation after the fact. Zero retries is a claim that the sample was not silently thinned in the direction that would matter most.

6. What it does not do

It does not verify correctness of what a file contains beyond column names and a coarse row sample. A column can hold a systematically wrong value while a scan built to summarise the corpus in seconds still reports it as present, unremarkable, and structurally sound.

It does not compute an exact row count for any stream beyond the three-file probe. The research panel's row count is exact because the harvest step that built it counted every row it wrote; this scan's row figures are estimates by design and are labelled as such rather than presented at the same precision as the byte counts.

It does not decide whether a stream is worth using for a given hypothesis. That judgment belongs in the substrate map this inventory feeds, kept as a separate document so that a fact and an opinion about the fact never share a cell in the same table.

7. Cross-references

BL-TR-002 for the inventory table and the retry account read together. BL-SD-016 for the retired carry verdict. BL-P2-003 section 2.3 for the class-level carry closure the near-repeat confirmed. The substrate inventory tool. Evidence E-0038.