Public-data options bridge: final findings

# Closure: free data acquired; market-risk application not established

September17,2026. Final data cutoff15:05UTC; reporting checkpoint16:00UTC /
18:00CEST. All four fixed prospective collection slots and the development
episode are complete. No replacement slots or further model/proposer variants
were introduced. The collector and evaluator exited normally after the last slot.

## What was established

- A no-account public production-data source: Deribit linear BTC_USDC options.
  The frozen cohort has102 instruments,17 paired strikes at each of three expiries.
  Every retained view has complete pair coverage and passes the declared
  conditional normalization. Five captures contain62,518 application messages;
  their20 overlapping/correlated views are not20 independent sessions.
- A tested conservative bridge from terminal-payoff operator calculations to
  thirty-minute averaged settlement. It uses conditional Jensen and martingale
  convex order under explicit assumptions; it is not a new theorem or exact
  average-payoff formula. All72 exact-tree and48 high-precision calendar checks
  pass. Fees, collateral, independently measured funding and execution remain
  outside the contract.
- A specific numerical advantage: on120 saved-field terminal reference cases,
  exact-cell evaluation uses38 spatial nodes and3,872 bytes of persistent arrays,
  taking0.602-1.389ms, with maximum-error upper bounds below4.69e-12. The finest
  prescribed FEM meshes use4,814-4,818 nodes and take5.000-8.322ms, with maximum
  error upper bounds2.75e-6 to1.30e-5. All360 FEM cases are retained. This is not an
  optimized/adaptive FEM or SOTA benchmark; array bytes are not whole-process
  memory. Comparator-process peak RSS ranges approximately86.1-89.2MB on the shared workstation.

## The application result is negative

No admitted model was found in any view under the frozen procedures. The closed
homogeneous study has12 failed proposals; the clock study has60 failed midpoint
proposals; the same-clock direct-feasibility audit has20 failed proposals. These
are bounded local searches, not a proof of infeasibility of the model class.

| Episode (UTC) | Views | Direct final slack range | Direct optimizer successes | Admitted |
| --- | --- | --- | --- | --- |
| Exposed development | 4 | 0.980-1.514 | 3/4 | 0 |
| 07:00 prospective | 4 | 1.390-1.672 | 1/4 | 0 |
| 10:00 prospective | 4 | 1.909-2.436 | 4/4 | 0 |
| 13:00 prospective | 4 | 0.937-1.825 | 3/4 | 0 |
| 15:00 prospective | 4 | 3.183-4.274 | 4/4 | 0 |

Positive slack is not accepted. All20 direct proposals satisfy the point hard
field constraints but miss quote bands. All120 prescribed risk directions are
explicitly not tested because independent admission never occurs. No market
risk-warning pair, narrow full-class range, predictive improvement, achievable
profit, weak-hardware advantage or practical/SOTA breakthrough is established.

The post-outcome exact-rational necessary-band audit finds no elementary
within-expiry contradiction in any of60 expiry/view blocks. This is not a complete
joint-curve or cross-expiry feasibility proof. It neither excuses the failed
application nor proves the local proposer/model/normalization responsible.

## Audit and preservation

Forty unit tests pass in the final run (TESTS_FINAL.json). The read-only final
replay verifies all7 downloaded source identities, rebuilds the cohort, reconstructs
all5 episodes' causal views and full contracts, and reproduces all92 saved fit
prechecks without rerunning optimization. See REPLAY_AUDIT_final.json.

Recorded HTTP-body and WebSocket-application bytes total20,771,504. Conservative
quota accounting charges504,510,446bytes, below the5,000,000,000-byte hard cap.
No data-source registration, authentication, checkout or paid service was used.
No trades, production changes, installations or reserved trading data were used.

Protocols, code, model/proposer results, failures, audits and the reviewed report
are backed up through scoped GitHub commits. Vendor raw records and derived quote
rows remain local because redistribution rights are not established. Their hashes
are recorded, but hashes cannot restore lost data: GitHub is not a complete backup
of the ephemeral feed. Exact raw-data replay requires preserving those local
inputs in an appropriate private backup. FINAL_MANIFEST.json separates first-party
artifacts from identities of local-only inputs and records software versions.

## Practical conclusion and continuation gate

Retain the verified data machinery, settlement bridge, exact-cell engine,
sensitivities, independent numerical verifier, and replay/accounting tests. Do
not present this version as a validated market-risk product or tune this campaign
further. The scientific outcome is a working numerical/data interface and a
failed external application gate under a carefully bounded procedure.

A future investigation should first establish a demonstrably market-compatible
observation/model contract against an appropriate established calibration baseline.
Only then is it meaningful to test whether independently verified risk warnings
improve an actual risk-review decision, including end-to-end cost on target
hardware. [Carr-Nadtochiy local variance gamma](https://arxiv.org/abs/1308.2326)
and [Le Floc'h's quadratic formulation](https://arxiv.org/html/2305.13791v2)
are relevant prior art, not implementations we have tested. A new model or source
would require a new named protocol, with its assumptions and comparison fixed
before outcomes. No such rescue search was started here.