BL-A-001: Evidence Register

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-A-001: Evidence Register

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Fifty-nine machine-readable artifacts, 2,524,772 bytes, published under _evidence/ as E-nnnn.json.

1. How to use this register

Every number in a released document carries an identifier of the form [E-nnnn]. Look it up here, then open _evidence/E-nnnn.json and read the value for yourself.

The register is deliberately blunt. It gives the subject of each artifact, the two or three values a verdict actually turned on, and the instrument that produced it. It does not summarise the artifact, because a summary is one more place for a number to drift.

Artifacts are published exactly as the instruments emitted them, with two exceptions recorded in section 4: pretty-printing, and redaction of third-party wallet identifiers.

Where an artifact and a narrative report disagree, the artifact wins. That happened twice during this build and both cases are recorded in BL-SD-000 section 7.

2. The register

Values are as they appear in the artifact. Basis points unless stated.

ID	Subject	Values a verdict turned on	Instrument
E-0001	Accumulation and distribution verdict by direction	long mean net -27.11, CI95 [-59.11, 2.33], n=152, REJECTED; short mean -51.57, n=5, INCONCLUSIVE	accdist verdict
E-0002	Adverse-selection markout by symbol against half-spread	BTCUSDT control markout at 60 s -0.5332, n=9,352,722; USUSDT at 10 s -3.1748, n=903,641; BTC half-spread proxy 0.1414	adverse selection
E-0003	Carry channel: dispersion, persistence, decomposition	research net mean 4.523, CI [0.577, 8.497], t=2.24, n=1,277; rank autocorrelation at 24 h 0.5144	carry characterisation
E-0004	Carry channel with horizon sweep	H=72 net 11.077, t=2.13, n=591; H=24 net 3.898, t=2.19, n=1,775	carry characterisation
E-0005	Carry channel with horizon and weighting sweep	equal weighting net 3.898, t=2.19, carry per sd 0.0827; H=24 annualised 1,423.0	carry characterisation
E-0006	Conditional information coefficient by signal, horizon, regime	63 rows; strongest cell move_1h_z at h=1 under high fragility, IC -0.0754, t=-32.58, 154 symbols	provenance unresolved
E-0007	Options cost floor: variance risk premium, BTC and ETH	BTC premium 8.1 vol points, t=22.0, n=1,277 days; ETH 3.98, t=7.7	door 5 cost floor
E-0008	Hedged option overlay path simulation	hedged mean 1.5123 pct, t=2.43, n=42; unhedged 0.3953 pct, t=0.25	door 5 simulation
E-0009	Hedged overlay, re-run of E-0008	identical payload apart from run timestamp	door 5 simulation
E-0010	Risk gate: Brier skill for volatility expansion	out-of-sample skill 0.012067 against a 0.02 threshold, n=2,356,200, base rate 0.2073	door 6 risk gate
E-0011	Risk gate, re-run of E-0010	identical payload apart from run timestamp	door 6 risk gate
E-0012	Risk gate under a revised, causal label threshold	out-of-sample skill 0.010534, n=2,356,200, base rate 0.2234	door 6 risk gate
E-0013	Effective execution cost: half-spread, markout, round trip	half-spread median 2.532; at 60 s markout -1.158 with adverse selection 3.689, n=169; taker round trip 16.063 against maker 2.709	execution cost, superseded
E-0014	Fill probability and effective round trip by rest-time budget	taker only 15.02; hybrid at 60 s 8.85; maker pure at full fill -0.025; 394 rows	execution fill
E-0015	Take-profit and stop bracket geometry against no-bracket baseline	baseline -8.735, CI [-27.485, 10.073], n=11,252 collapsed; cells beating baseline 0	exit geometry
E-0016	Fragility event configuration and 5-minute proxy validation	chosen configuration 896 events, 62 in band; pooled proxy correlation 0.1366, FAIL, 5 symbols	G2 characterisation
E-0017	Fragility event configuration, revised choice	chosen configuration 905 events, 62 in band	G2 characterisation
E-0018	Five-minute against one-hour fragility proxy	5-minute pooled correlation 0.7017 against 1-hour 0.0848, 25 symbols; capped at WEAK, gate passed	G2b proxy
E-0019	Crowded positioning skew and tail test at 24 h	verdict dead; skew delta 0.6754, CI [-0.938, 2.020], n=904 crowded, permutation p=0.034	G3 test
E-0020	Crowded positioning retest at 72 h and 24 h	verdict dead; 72 h mean 0.942, CI [-1.356, 3.326], n=18,129; 24 h mean 0.079, CI [-1.236, 1.362], n=28,788	H-P2 test
E-0021	Conditioned reversion net return at 24 h and 4 h	verdict dead; 24 h mean -8.735, CI95 [-27.002, 9.595], n=11,252 collapsed; 4 h -12.417, CI95 [-18.442, -6.306], bootstrap p=0.0	H-R1 test
E-0022	Conditioned reversion, symbol and PnL concentration diagnostic	raw mean 0.86 against bootstrap mean -8.74, CI95 [-28.07, 10.2], n=43,209; top ten symbol share 0.0799	H-R1 diagnostic
E-0023	Volatility stop overlay, stop-size sweep	no stop -8.735, CI [-27.485, 10.073], n=43,209; at 1.5 sigma -16.65, CI [-30.42, -3.167], bootstrap p=0.0168	H-R2 test
E-0024	Stop sweep with equal and capped notional sizing bases	no stop equal weight -8.735; capped notional -1.035, CI [-4.856, 2.458], bootstrap p=0.574	H-R2 test
E-0025	Cross-sectional residual reversion at 24 h and 72 h	verdict dead; 24 h net 1.659, CI [-3.165, 6.485], n=1,275 rebalances; 72 h -1.07, CI [-15.959, 12.925], n=424	H-X1 test
E-0026	Fragility-conditioned rebalance with permutation null	verdict dead; primary net 4.995, CI [-3.05, 12.667], n=569; null centred -9.597 against observed 4.995, permutation p=0.0	H-X2 test
E-0027	Decile map: price, funding, dispersion, fragility	price decile long-short 10.959, n=1,275 rebalances; funding-paid-by-long decile 1.974	H-X2 map
E-0028	Variance and liquidity quantile cells with beta-gap control	best cell net 5.962, t=0.697, n=569; beta gap median 49.34, share above 50 bps 0.7592	H-X3 variance map
E-0029	Rebalance-anchor robustness audit	kill robust true; net median across phases -3.54, net at the phase used 4.28 ranking 3 of 24; net positive 8 of 24. The stored `n_phases_significant` of 0 counts positive phases only; recomputed two-sided from the per-phase statistics the artifact carries, 7 of 24 are significantly negative, worst at t=-4.80. See `BL-IS-004` section 8	anchor audit
E-0030	Information channel map, baseline pass	best cell, residual only, net 6.85, t=1.715, n=569	H-X4 information map
E-0031	Information channel map with within-leg and subperiod cells	best cell, funding-weighted, net 7.527, t=1.871, n=569	H-X4 information map
E-0032	Information channel map, re-run of E-0031	identical cell results	H-X4 information map
E-0033	Information channel map with post-hoc filter and liquidity control	post-hoc filtered cell net 4.123, t=1.204; liquidity ratio 1.108	H-X4 information map
E-0034	Information channel map with unconditional comparison	unconditional baseline net 1.833, t=0.703, n=1,174	H-X4 information map
E-0035	Information channel map with calm-regime control	calm control baseline net -4.287, t=-1.416, n=605	H-X4 information map
E-0036	Post-mortem: research against held-out gross separation	research era separation 4.17, fragile gross 16.08 against calm 3.86; held-out B separation -0.32	H-X4 post-mortem
E-0037	Final gate test against four acceptance criteria	verdict failed, family dead; primary net 4.282, CI [-2.16, 10.718], t=1.303, n=826; significance gate false	H-X4 test
E-0038	Data lake inventory by stream	total 8,195.8 MB; Bybit hourly klines 215.0 MB across 154 files	substrate inventory
E-0039	Information decay: coefficient by prediction horizon	50 rows; strongest move_1h_z at h=1, IC -0.05228, t=-31.54, n=4,181,948	information map
E-0040	Rebalance hour seasonality, gross and net by UTC hour	best hour 00 UTC gross 6.506, t=2.598, net -0.339, n=1,774; hours net positive 3 of 24	rebalance hour
E-0041	Rebalance hour with session-block aggregation	Asia block 23:00 to 09:00, all eras, gross 5.57, t=6.98 across 11 hours	rebalance hour
E-0042	Layer 3 eligibility overlay against static weights	overlay mean -28.09, t=-2.59; incremental against best static -42.2, n=249 weeks	layer 3 eligibility
E-0043	Layer 3 eligibility, revised static baselines	overlay mean -0.89, t=-0.08; incremental -8.89, n=249 weeks	layer 3 eligibility
E-0044	Momentum net return sensitivity to anchor hour	24 rows; maximum at hour 10, net 14.11, t=2.49, n=249; minimum at hour 7, net 4.20, t=0.72	provenance unresolved
E-0045	Momentum cells by lookback and horizon	worst cross-sectional cell L=24 H=24 net -7.95, t=-3.47, n=1,775; best cell L=720 H=168 net 71.13, t=1.74	momentum map
E-0046	Momentum cells with beta neutralisation and yearly breakdown	L=720 H=168 neutralised 16.24, t=2.86; year 2021 total 103.33, neutralised 29.94	momentum map
E-0047	Momentum cells, refined neutralisation pass	L=720 H=168 neutralised 16.24, t=2.86	momentum map
E-0048	Momentum cells with capacity and breadth sensitivity	full universe neutralised net 14.11, t=2.49, n=249; liquid half 18.73, t=2.26, n=241	momentum map
E-0049	Rebalance-hour bisection between build and load paths	lowest-residual long cell without funding net 3.33, t=1.24, n=1,275; build hour 00 against load-grid hour 10	reversion bisection
E-0050	Reversion reconciliation: history threshold and yearly net	research era without history threshold net -6.36, t=-2.33, n=1,277; year 2022 net -15.59, t=-1.63, n=108	reversion reconciliation
E-0051	Perpetual universe quoted-spread census	median 5.26, mean 8.077, n=676 measured; 62.1 pct with two ticks of room; maker fee marked unverified at measurement	spread universe
E-0052	Out-of-sample stability: width, direction, amplification	research width difference 14.37, t=10.25, n=1,277 rebalances; research direction IC mean -0.02983, t=-5.10	OOS stability
E-0053	Settlement-window pre and post markout screen	verdict dead; pre-leg median -3.39, n=4,033, fail; post-leg median -1.44, win rate 0.48	B2 settlement screen
E-0054	Accumulation and distribution verdict, second copy	byte-identical to E-0001	accdist verdict
E-0055	Venue verification pull: clearinghouse state and matched fills	3 account states, 302 matched fills; wallet identifiers redacted	provenance unresolved
E-0056	Venue verification pull: price calibration and fill taxonomy	62 calibration rows, relative error at the 1e-14 scale; 10 direction categories; liquidation fills 0	provenance unresolved
E-0057	Vault clearinghouse state snapshot	account value 222,138,927.11, total margin used 477.87; vault identifier redacted	provenance unresolved
E-0058	Taker signal backtest matrix by signal, symbol, hour, horizon	4,800 cells at a 30-trigger minimum; extreme cell -265.26 mean realised, n=40	T0 backtest
E-0059	Symbol universe discovery and tiering	570 instruments, 192 passing the movement filter; top candidate composite score 7,738.57 at a 360.82 median move	universe discovery

3. Duplicate and superseded artifacts

Six artifacts are re-runs of another artifact in this register, retained because the instrument emitted them and removing them would misrepresent what was actually run.

ID	Duplicates	Difference
E-0009	E-0008	Run timestamp only
E-0011	E-0010	Run timestamp only
E-0032	E-0031	Identical cell results
E-0054	E-0001	Byte-identical, written to two locations

E-0013 is retained with its instrument marked superseded. The instrument that produced it was found to double-count adverse selection and withdrew its own output. E-0014 is the replacement. Both are published, because the pair is the evidence that the withdrawal happened.

Two pairs of artifacts record a genuine revision rather than a re-run. E-0012 supersedes E-0010 after the expansion label was made causal, and the risk gate lift figure moved from 1.52x to 1.49x as a result. E-0043 supersedes E-0042 after the static baselines were corrected. Both earlier artifacts stay in place.

4. Publication treatment

Artifacts are published as emitted, with two changes.

Formatting. JSON is re-serialised with indentation for readability. No key, value, or ordering is altered. Total published size is 2,524,772 bytes against 2,667,102 bytes at source; the difference is whitespace normalisation.

Redaction. Third-party wallet identifiers are replaced with a stable pseudonym of the form wallet_<8 hex>, derived by hashing the original. The pseudonym is stable, so a reader can still tell that two rows refer to the same participant, which is what the research depended on. The identity itself is not recoverable from the register.

ID	Fields redacted	Keys affected
E-0055	305	`addr`, `user_a`
E-0057	1	`hlp_addr`

No other artifact contained a wallet identifier, a host address, an email address, or a credential-shaped string. The published set was re-scanned after redaction and returned zero residual identifiers across all 59 files.

Redaction never touched a value carrying a result. Every number cited anywhere in this record is present in the published artifact exactly as the instrument wrote it.

5. Provenance

Fifty-three of the fifty-nine artifacts resolve to a named instrument in the research toolchain, and that instrument is published or specified in Volume V.

Six do not, and the register says so rather than guessing.

ID	Artifact	Status
E-0006	Conditional information coefficient map	No producing instrument found in the repository
E-0044	Momentum anchor sensitivity	No producing instrument found in the repository
E-0055 to E-0057	Venue verification pulls	Producing script did not survive into the repository

The rule that a verdict is only valid when it comes from an in-repository instrument with tests attached was written partway through the first programme, after an ad-hoc acceptance script reported a false result and was subsequently lost. These six artifacts predate that rule or sit outside it. Their contents are published and their numbers are usable, but a reader should treat them as observations rather than as instrument output, and no verdict in this record rests on any of them alone.

E-0044 is the one to watch. It carries the anchor sensitivity that revised a published momentum figure downward, and it has no traceable producer. The finding it supports is corroborated by E-0029, which does have one.