BL-SD-045: The Fragility Family on Sealed Held-Out Data

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Bybit perpetuals Programme: II Verdict: FALSIFIED Axis of death: Mechanism, inverted out of sample Transferability: UNIVERSAL

1. Identification

This is the record of the only hypothesis in the archive to die on mechanism inversion, and the reason the sealed held-out era exists as a standing feature of the programme rather than an occasional formality.

2. Economic thesis

The mechanism is the one stated in BL-SD-043: forced flow localised to a single symbol, most acute when the broader market is fragile, drives an idiosyncratic move that reverts once the local coercion is exhausted. H-X4's refinement is in how each leg of the dollar-neutral pair is sized. H-X2 weighted names equally within the fragile bucket. A characterisation pass across four candidate information channels, open-interest change, path efficiency, turnover-surge, and funding-rate crowding, found that weighting each leg by its trailing change in open interest, oriented so both legs overweight names where open interest has moved least toward the crowd, produced the strongest research-era estimate of any construction examined: net +8.784 basis points at t=2.121 on the 569 fragile rebalances, against the equal-weighted construction's +6.850 at t=1.715. The reasoning was that a name whose forced deleveraging has genuinely occurred should carry more of the reversion signal than one merely ranked into the bucket by residual size alone.

Who is forced to lose. Unchanged from BL-SD-043: leveraged holders liquidated in a specific symbol, at the moment the market is least able to absorb it. The information weighting is intended to concentrate exposure on the names where that forced liquidation was largest, not to change who is forced to lose.

The channel search that produced this weighting looked at the research era's own outcomes before selecting oic over the other three candidates. That is disclosed here rather than smoothed over: it is the exact reason a sealed era, never touched by the channel search, was the correct next instrument rather than an optional check. A result chosen because it performed best on the data used to choose it carries no information about whether it will perform on data it has not seen. Section 7 is the test of that question.

3. Falsifiable claim

Four acceptance gates were fixed in advance of opening held-out B, before one combined-era figure was computed:

All four gates were pre-registered together and evaluated once, per the single-shot rule R-11 governing the sealed era.

4. Mathematical specification

Notation per BL-D-011. Base residual signal, fragility gate, and funding
correction identical to BL-SD-043 (H-X2) section 4.
 
Information weight, channel oic, discovered by the pre-primary
characterisation pass and locked into this pre-registration before held-
out data was opened:
 
  dOI_i(t)      = trailing change in open interest, symbol i, at t
  w_i(t)        = g( dOI_i(t) ), symmetric-low orientation: both legs
                  overweight names with the smaller (nearer-zero or more
                  negative) trailing change, where forced deleveraging is
                  most concentrated
 
  position_i(t) = sign(residual_i(t)) * w_i(t) / sum_j |w_j(t)|,
                  dollar-neutral, gross-1, restricted to rebalances where
                  the causal trailing-percentile fragility gate of
                  BL-SD-043 section 4 is satisfied
 
Funding: PnL_long  += -funding_rate accrued over (ts_entry, ts_exit]
         PnL_short += +funding_rate accrued over the same window,
         unchanged from BL-SD-043 Change 1
 
Exit: open[t+1+H], H = 24h, unchanged
Cost: c_rt = 8.85 bps nominal round trip, scaled by measured turnover
      (0.949 conditioned); sensitivity reported at 11.0 and 12.17 bps

Causality holds for the fragility gate, the residual ranking, the funding accrual, and w_i(t), each computed from information available at or before t, per BL-SD-043 section 4. The weighting scheme itself is the one term here that is not causal with respect to the research question: it was chosen after observing which of four channels performed best on the research era, a form of look-ahead into that era's own outcome. This is stated rather than concealed, and it is precisely why the acceptance gates in section 3 were required to run against data the channel search never touched.

5. Operating envelope

6. Instrument

The held-out gate test and the information-channel map, sharing the pinned cross-sectional module with BL-SD-042 through BL-SD-044, extended for the information-weighted leg construction. The post-mortem separation diagnostic produced the separation decomposition in section 7. The rebalance-anchor robustness re-run described in section 13 used the instrument built for BL-IN-006.

7. Results

The figure that justified opening the seal, research era only, the information-weighted specification against its own discovery data:

This is the number the channel search produced and the number that made H-X4 look worth the seal. It is not the number that decides the hypothesis; the four gates below are.

Per-era decomposition:

Cost sensitivity, combined eras:

The equal-weighted comparison, same fragility gate, no information weighting, combined eras: net +2.891, CI95 [-3.581, +9.116], t=0.893. Information weighting improves the point estimate but neither construction clears P1.

The post-mortem decomposes the reversal by era, equal-weighted against information-weighted:

Both constructions reverse sign in held-out B, at almost the same magnitude. The information weighting is not the source of the reversal; it is present under equal weighting as well.

Held-out B, decomposed into three subperiods: sub1, n=63, net -16.739, t=-1.865; sub2, n=66, net +4.194, t=+0.602; sub3, n=60, net -15.425, t=-1.407. Two of three subperiods inside the held-out era are strongly negative; only the middle subperiod is positive, and weakly so.

The separation the mechanism claims to depend on, fragile-regime gross return against calm-regime gross return, by era:

Fragility amplification measured 8.8 times in the research era. In held-out B the fragile-regime gross return is negative while the calm-regime return remains positive, and the separation itself changes sign.

8. Verdict, axis of death, and transferability

Verdict: FALSIFIED. Axis: mechanism, inverted out of sample. Transferability: UNIVERSAL.

Gates P2 and P3 pass. The calm-market control is not significantly positive, ruling out a construction that would have worked regardless of regime, and the liquidity ratio of 1.11x rules out an edge concentrated in illiquid names. Neither is where this hypothesis dies.

P1 fails on its own terms: the combined-era interval spans zero. On power grounds alone this might have warranted NOT PROVEN, the classification BL-SD-043 itself received. It does not receive that classification here because P4 settles the question P1 leaves open. The research era produced +8.242 bps at t=2.106, ordinarily read as confirmation. The held-out era produced -8.992 bps at t=-1.655, the opposite sign at comparable magnitude. Fragility amplification, measured at 8.8 times in the research era, reversed to -0.32 in held-out B. That is not a weaker echo of the research-era finding. It is the mechanism running backward on data the discovery process never saw.

Transferability is UNIVERSAL. A cost-axis or power-axis death names a parameter, round-trip cost or sample size, whose value in a different envelope would change the outcome. A mechanism that inverts sign has no such parameter: the held-out era's realised prices are what they are, and no execution assumption, capital scale, or latency band changes what the market actually did. Section 10 states this in full.

9. Cost and capacity analysis

Gross, primary combined eras           +12.129 bps
Cost, turnover-scaled (0.949)           -7.847 bps
Net                                     +4.282 bps, CI95 [-2.16, +10.72]
Liquidity ratio (P3)                     1.109x, PASS
Net at 11.0 bps                         +2.376 bps
Net at 12.17 bps                        +1.338 bps

The capacity gate passes cleanly: the ratio of turnover on the low-dOI leg to the high-dOI leg is 1.109, well above the 0.8 floor, which rules out the common failure mode of an edge concentrated in a handful of thin names. Capacity is not the axis of death and no further capacity work is warranted for this construction.

10. Envelope sensitivity and reopening conditions

Binding parameter: none. This is stated plainly rather than left implicit, because the template requires a reversal computation and none exists here.

Reversal condition: not computable, and not computable in principle.

A cost-axis death names the round-trip cost below which the verdict flips, and a power-axis death names the sample size at which the interval would narrow enough to decide. Both are properties of the measurement instrument, and both can be changed by choosing a different envelope. A mechanism inversion is not a property of the instrument. The held-out fragile-versus-calm separation is -0.32, against a research-era separation of +4.17. No round-trip cost, no capital scale, no latency band, and no execution channel changes the sign of a return series that has already been realised on recorded prices. Reversing this verdict would require the held-out era to contain different market outcomes than it does, which is not an envelope parameter this or any programme controls.

What would reopen it: a materially different substrate, a different venue, or a forward shadow period long enough to determine whether the inversion in held-out B was itself a regime-specific artifact of that particular window, tested against fresh data this programme has not yet recorded and pre-registered before being opened.

What would not reopen it: any reanalysis of the eras already examined. Both held-out A and held-out B are now spent on this specification, and the research era has been searched once already to find the weighting that produced the pre-seal gate figure. Re-running either against a new weighting scheme would repeat the exact look-ahead this dossier's section 4 already discloses and would not constitute a new test.

11. Reference implementation

The information-channel machinery is published as rank_weights and double_sort in 05_INSTRUMENTS/strategies/python/src/fragility_conditioning.py, with the rest of the conditioning layer. The anchor-robustness instrument specified in BL-IS-004 is published as part of 05_INSTRUMENTS/src/inference.py, and the portfolio arithmetic as 05_INSTRUMENTS/src/cross_sectional.py. The held-out gate driver and the post-mortem separation diagnostic remain scheduled.

One control in the published layer is worth reading closely, because without it this dossier's comparison would carry no weight. double_sort pre-selects the most extreme 40 percent of residuals and then keeps the best half of those by the conditioning channel, arriving at 20 percent of the universe. The unconditioned baseline takes the extreme 20 percent directly. Both legs therefore hold the same number of names, so a difference in result cannot be attributed to diversification or to name count. Without that match, a conditioned arm that outperforms has three candidate explanations and the measurement can separate none of them. Tests pin the size match, the disjointness of the legs, and the full-size fallback when too few names carry a channel reading.

Weights inside a leg are proportional to rank on the channel rather than to its value, because the four channels this family sorts on have unrelated scales and tails and a raw-value weighting would let one outlier own the leg. Publishing that corrected a documentation error carried in the source: the construction was described as giving a two-to-one weight range between the best and worst name in a leg, and it gives exactly three to one at every leg size. The arithmetic is unchanged and no measurement moves, but the description was wrong in the direction that matters, since the reason to state a bound at all is to say how much concentration the weighting permits. A test pins the ratio now.

12. Evidence register

13. Lessons extracted

Every gate this hypothesis faced before the seal was opened had already passed. Fragility conditioning passed, the permutation null passed at 5.83 standard deviations in BL-SD-043, the subperiod check passed, the liquidity control passed, and the information weighting found in characterisation improved the research-era estimate to significance at t=2.121. A programme measuring this hypothesis without a sealed held-out era would have had no remaining gate to fail. It would have carried H-X4 forward as validated, on the strength of a research-era result that the held-out era shows was not there.

The subsequent anchor audit, recorded in full in BL-IN-006, re-ran this exact specification across all twenty-four possible hourly rebalance phases, rather than the single 00:00 UTC phase the programme had used without recognising it as a choice. Zero of twenty-four phases were significant. The median net across phases was -3.54 basis points. The phase actually used, +4.28 basis points, ranked third of twenty-four: only two anchors would have produced a stronger result, and twenty-one would have rejected the hypothesis more decisively than the one actually used. The kill did not merely hold under the audit. It strengthened, because the phase this programme happened to use was already among the more favourable ones available, and the audit confirms the rejection is not an artifact of that accident.

This is recorded as the design working, not as a near-loss. The sealed era exists to catch exactly this failure mode, a specification that clears every gate available before it and fails only on data constructed never to inform the specification's own discovery. It caught one, on the first hypothesis to reach it with a genuinely strong prior record, and it caught the strongest possible version of the failure: not a weaker echo, but a sign reversal.

14. Cross-references

BL-SD-043 for the H-X2 construction and mechanism this hypothesis refines. BL-SD-044 for the variance-reduction sweep that cleared no better weighting existed among twelve equal-weighted variants before this information-weighted specification was pre-registered separately. BL-P2-002 section 7 for the programme-level account of the seal opening. BL-IN-006 for the full anchor-sensitivity audit. BL-D-002 on transferability classification. BL-SD-000 ledger row H-X4.

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