Research status: REJECTED. This is a methodological postmortem, not a deployable algorithm, trading signal, or claim that language models are generally useless in finance.

Research question

Can a constrained language-model veto improve a frozen mechanical residual-spread strategy after costs when success is defined by a complete precommitted gate rather than one selected metric?

The model was deliberately narrow. It received an already formed, past-only mechanical situation and could return only approve or reject. It could not change the pair, direction, hedge weights, entry, holding period, exit, or economics.

The research hierarchy matters:

  1. V63 was mechanical development.
  2. V64 was an unchanged temporal replication on earlier windows.
  3. Their combined 85-trade history was explicitly a post-selection diagnostic, not independent confirmation.
  4. V65 tested whether an LLM veto added value on those 85 diagnostic situations.
  5. Only V66 was the one-shot locked family final, and it tested the unchanged mechanical control without GPT.

Frozen comparison logic

The public pseudocode is intentionally incomplete but shows the control structure:

mechanical_signals = frozen_rule(past_only_snapshots)

for signal in mechanical_signals:
    decision = constrained_llm_veto(signal.snapshot)
    gpt_arm keeps signal only if decision == approve

compare gpt_arm with unchanged mechanical_arm
pass only if every frozen economic and statistical gate passes

Execution occurred no earlier than the next hourly open. The aggregate round-trip allowance was frozen at 0.14% under base economics and 0.21% under stress. The source does not support decomposing that aggregate allowance into separate public estimates of commissions and slippage.

V65: better averages, failed added value

All 85 structured requests completed: 67 approvals and 18 rejections. API cost was $0.6658775 under an authorized $2.01 ceiling.

Metric Mechanical control GPT veto Direction Trades 85 67 Lower frequency Stress expectancy +0.31768% +0.34341% Improved Positive windows 5/8 7/8 Improved Stress total return +27.0027% +23.0082% Worsened Maximum additive drawdown 10.7389% 12.5471% Worsened

The veto improved average stress expectancy and window breadth, but it removed profitable trades and retained the loss that deepened drawdown. The frozen gate required drawdown to be no worse than the mechanical control, so V65 failed.

Family-aware White Reality Check and Hansen SPA p-values were 0.0639 and 0.0675. These diagnostics do not override a failed precommitted economic gate. V65 was rejected before final; describing V66 as “GPT failing on holdout” would be incorrect.

V66: mechanical-only locked family final

V66 used no GPT and incurred no paid API calls. It evaluated the unchanged mechanical contract across four consecutive locked quarters from July 1, 2025 through July 1, 2026.

Metric Base Stress Trades 126 126 Wins / losses 67 / 59 67 / 59 Total additive return −3.9136% −12.7336% Expectancy per trade −0.03106% −0.10106% Profit factor 0.9470 0.8368 Maximum additive drawdown 13.9744% 19.8200%

Remove-best expectancy was −0.06394% per trade, and both family-aware p-values were 1.0000. Costs worsened the loss but did not create it because the base result was already negative.

Two middle quarters were positive, while the first and last were negative; the final quarter erased the accumulated gain. With 126 trades, this is better described as regime dependence than a simple shortage of events.

Integrity boundary and conclusion

The V66 outcome was computed after its precommit, and residual-spread outcomes for this exact contract had not previously been opened. However, the market interval was not globally pristine because other strategy families had used the underlying period. The valid conclusion is therefore a locked rejection inside the registered residual-spread family, not a theorem about every possible strategy.

The evidence rejects two different propositions:

  • the constrained GPT veto did not add acceptable value to the selected mechanical rule;
  • the stronger mechanical control did not survive its one-shot family final.

The family is closed. Retuning the prompt, thresholds, direction, or exclusions after seeing V65/V66 would convert the revealed samples into development. A valid successor requires a distinct economic hypothesis, a new precommit, family-level multiplicity accounting, and a new independent data stream.

Questions for methodological critique

  1. Is “no worse drawdown plus improved return/frequency” an appropriate added-value gate for a veto model, or would you precommit a different joint criterion?
  2. How would you account for multiplicity when model reasoning can create adaptive but difficult-to-count search paths?
  3. Is “locked within the registered family but not globally pristine” a sufficiently precise label, or would you use a stricter taxonomy?

The complete prompt, private project, batch identifiers, source endpoints, and reproducible trading parameters are intentionally omitted. A sanitized public evidence appendix may be prepared separately. This material is for research and education only and is not investment advice.

Author: SERGII / Evidence-Based Trading Lab