Live A/B Experiment

Two identical agents. One is verified. One is not.

Same model (Kimi K2.6), same live market, same starting capital of $1035 per arm. One agent trades unverified. The other must clear ThoughtProof Sentinel before every action. This measures reasoning integrity, not market timing.

Snapshot: 2026-08-30 · 5,424 total cycles · 2,998 gated · 2,101 divergent · Still running

Result

Arm	Equity	Return	Executed	Blocked
Verified (Sentinel gate)	$821.80	-20.6%	997	3316
Unverified (counterfactual)	$9.83	-99.1%	1,394	—
Delta	+$811.98	+78.5pp

In a down market the gate did not make the agent profitable. It kept the portfolio recoverable. Without the gate: -99.1%, near-total loss. With it: -20.6%, still trading. The 3316 blocked actions are the substance of the claim — the gate stops what it should.

Verdict distribution (2,998 gated cycles · 2,101 divergent excluded)

ALLOW_AFTER_REPLAN — first plan failed, agent produced a better one1408

UNCERTAIN — human-in-the-loop required, no autonomous execution457

BLOCK — reasoning did not hold, action stopped1124

ALLOW — plan sound on first pass9

N/A — portfolios diverged, trade not comparable2101

Methodology

Both agents run Kimi K2.6 against the same live USDC spot markets on a major US exchange (a rotating set of liquid pairs — ETH, SOL, NEAR, XLM and others). The verified agent's actions pass through the ThoughtProof Sentinel reasoning gate on every cycle. Sentinel runs on OpenServ's SERV Reasoning models and judges independently of the trading agent — it never sees the agent's own confidence, only the plan and the market. Each check costs under a cent, which is what makes verifying every action economical. Fail-closed: both BLOCK and UNCERTAIN stop execution. No verdicts are hand-set — every one is a live API call. (High-stakes actions can escalate to an adversarial RV panel; in this run the Sentinel gate resolved every cycle, so RV did not fire.)

Reproducibility

The gate that produced these verdicts is public and callable: sentinel.thoughtproof.ai. Anyone can send a plan and market context to the same endpoint this experiment uses and get the same kind of structured verdict back. The A/B harness that drives the two agents is internal (it holds live-capital credentials), and trade-level data is not published for capital privacy — but the part that does the judging, and the fail-closed policy (both BLOCK and UNCERTAIN stop execution), are fully inspectable. These numbers update daily from the running experiment.