A track record that starts before the outcome exists.
Three steps turn a plain-English idea into a locked, forward-tested, fully journaled record anyone can audit.
- 1
Describe
Say the strategy in plain English — the universe, the thesis, when it trades. No code. If a signal can't be computed point-in-time, it's declined, not faked.
- 2
Lock
It's pre-registered as a tamper-proof spec — a content hash — and locked, with a timestamp. Editing it forks a new strategy, so the original record stands, immutable.
- 3
Run
From lock on, it trades forward on live data. Every decision is journaled with the model's verbatim reasoning and the model id the provider reported — as it happens, permanently.
The limits, stated up front.
A record is only worth what its weakest claim is worth. These are the two places where ours is weaker than it looks, as of 13 August 2026.
A pinned model name is not a pinned model
Two of the three models we run are pinned by NAME, because their vendors publish no versioned id to pin instead — Anthropic stopped issuing dated snapshots after Claude Opus 4.5, and Google's catalogue reports a version for Gemini 3.6 Flash that cannot actually be called. If either vendor later serves different weights under the same name, the version recorded against past decisions would be wrong, and nothing in the record would catch it. The third, GPT-5.5, is pinned to a dated snapshot and is not exposed to this. We capture Google's reported catalogue version at verification time so a change is at least detectable afterwards. That is detection, not prevention, and we would rather say so than imply a guarantee we do not have.
The hash is checkable, but nobody is checking it for you
The spec hash is real: it is recomputed from the stored spec over the locked fields, and it is stamped onto every decision the strategy journals. What we do not run is a watcher. Nothing on a schedule recomputes those hashes and raises an alarm — the tooling to verify a decision against its locked spec exists and is exercised by our tests, but it is invoked on demand, not continuously. So the honest claim is that tampering is DETECTABLE by anyone who checks, including you, and not that it is being detected. The record is append-only at the database level, which is a different and stronger guarantee: the application role cannot rewrite a journalled decision or restate a scored day at all. We would rather describe the mechanism than let the word tamper-proof do work it has not earned.
The reasoning-effort setting binds one of the three models
Each locked spec pins a reasoning effort of "high". It reaches the Anthropic seat and no other. OpenAI's API rejects that setting on the same request as a forced tool call, and every decision here is a forced tool call, so that seat runs with no extended reasoning at all; the Gemini seat uses a fixed thinking budget chosen for reliability rather than the value in the spec. We left the field in place because it is true and load-bearing for the Anthropic seat — fourteen of the sixteen locked strategies, and every single-model one — and removing it would have re-hashed all sixteen records to state something the code already decides.