Deterministic vs Jev-gated spot strategies

The design in plain language

Both arms get the same screened coins, the same paper wallet of 10,000 USDT, the same fees and the same 90 days of live market data. The only difference is the gate.

• Arm A: deterministic

  Puts an equal slice into each of 10 screened coins once a week for 13 weeks, as scheduled. It never exits a position.

• Arm B: Jev-gated

  The same schedule. Before each of the first 12 weekly slices, a small decision-only model (TypeSafe Jev) is asked to approve it or skip it: 12 slices × 10 coins = 120 decisions.

  Whatever it holds back carries into that coin’s next week, and the 13th slice invests everything left with no model call, so both arms put in the same money. The model decides when, not whether, and it can only answer approve or skip.

  It can never invest outside the schedule or above the coin’s budget, choose a coin, move money between coins or close a position. If it still gives no valid answer after retries, arm B puts in the coin’s whole budget for that week, carried part included.

The model is TypeSafe Jev (jev-1.13.0), a hosted model that answers a fixed question with a choice and a confidence and cannot write free text. It is pinned with its question text and settings before the start. It is a gate only and never originates a position. It is a hosted service, and text it reads can influence its answer, so both are limits of this design.

Strategies

• DCA: the primary strategy

  The verdict that counts is about this strategy. DCA is the yardstick the validation gate measures other strategies against, so instead of the gate it must pass eight pre-registered mechanical checks (schedule, carry-forward, no exits, minimum stake, capital, replay, gate wiring and config) before the clock starts.

• TrendCash

  control (failed validation gate)

  TrendCash failed the validation gate. It runs only as a labelled, paper-only control, where the model gates each entry and each discretionary exit. It is left out of the headline and can never be promoted to real money. Its configured stake is 2,500 USDT, with capital headroom 0.50 and at most 2 open trades.

Baselines

Every arm is also compared against two passive baselines that pay the same fees. Beating the other arm while losing to a baseline is a valid, reportable result.

• Hold

  The same coins in equal weights, all put in at the start and left untouched.

• DCA baseline

  The same weekly plan computed from prices, with no bot. For the DCA strategy it is also a mechanical cross-check on arm A.

The pre-registered verdict

There is one verdict per strategy and never a pooled one. The headline is about DCA alone, the one counted strategy. Three tests are fixed in advance for it:

• H1-D, return: arm B’s net return minus arm A’s is at least +1.0 percentage point, and its 90% confidence interval lies above zero. It fails at −1.0 point or worse with the interval below zero.
• H2-D, risk: arm B’s maximum drawdown is no more than 2.0 points worse than arm A’s.
• H3-D, timing: a randomization test. The weeks in which B held money back are compared with B’s own decisions placed on randomly chosen weeks, on the same prices. It passes if B’s real weeks did better than 99.5% of the random placements, and fails if they did worse than 99.5% (α = 0.005 on each side).

• Agentic better: H1-D, H2-D and H3-D all pass.
• Agentic worse: any one of the three fails.
• Inconclusive: anything else, with its reason. This is the expected verdict.
• Void: the run is broken, for example a bot is touched by hand, more than 10% of the model’s decisions fail, or a pinned setting changes. A voided run stays public with its reason. If final data recorded by seven days after day 90 is still incomplete, the run is Void with the reason “readout data incomplete”.

Unavailable: if a bot’s restart and downtime record cannot be shown complete, no verdict except Void is published. The seven-day data deadline still applies.

This experiment is unblinded by design: live equity curves and model decisions are public. Formal statistics, p-values and bounds publish only at the frozen readout. After day 90, “Computing final readout” lasts until the final data is complete or the seven-day deadline is reached. The published readout is stored once; later data cannot change it. Integrity rests on frozen rules, the fixed end date and Void on intervention.

Power, stated honestly. In the experiment’s own power simulation, even a gate with a real timing edge of 7.5 percentage points reads “Agentic better” in only about 52% of runs, and it needs about 9 points to do so in three runs out of four. The coins move together, so the effective sample is much smaller than the 120 decisions.

Second finding: “LLM beyond simple rules”. It is shown beside the verdict as yes, no or “rule better”, and never replaces it. It compares B with one benchmark fixed in advance: skip a coin’s slice while its price is above its 28-day moving average. “Yes” needs the verdict “Agentic better” and a clear lead over that rule; “rule better” means the rule clearly did better. It is not reported on a Void or Unavailable run, or when the benchmark cannot be replayed. Even a skilled gate rarely reaches “yes” in 90 days (about 7% of simulated runs at a 7.5-point edge), so a “no” is not evidence against the model.

TrendCash gets its own verdict under the entry-gating rules and is left out of the headline. Every test, threshold and seed is in section 3 of the pre-registration.

What we already know, honestly

• In the owner’s prior backtests (sanitized evidence committed in the pack), six spot strategies lost money in 16 of 18 strategy-by-regime runs, including 5 of 6 during a +48% BTC rally.
• TrendCash failed its pre-registered validation gate, which is why it appears here only as a control. Historical gate attempt 3 (2026-10-08) returned −1.24% over 40 holdout trades, against Hold +27.45%, with zero capital skips. These are prior backtests, not this experiment’s results.
• One study found that a language-model crypto agent did not beat traditional rule-based strategies (CryptoTrade, EMNLP 2024). In US equities, claimed advantages of language-model strategies faded over longer periods and broader sets of stocks (FINSABER, arXiv 2505.07078). Both are cited in section 2 of the pre-registration.
• We set no return target. Ninety days of paper trading cannot establish one, and a result where the gate does nothing or does harm is a fully acceptable outcome.

Design limits: a paper wallet of 10,000 USDT per arm, in all four bots; a DCA basket of 10 coins at 75.38 USDT per coin per weekly slice, over 13 slices (the pre-registered basket rule gives 10 coins at today’s exchange minimums and can only lower that number); fees of 0.10% per side; spot only, long only, no leverage; a model budget of USD 4, or USD 2.00 per strategy, which the expected cost should not reach. The coin list is frozen at the start and labelled “Passed the Conservative screening profile as of” its date.

Live dashboard