Skip to the evidence
AI Trading SignalsDefinition / Population / Proof

AI Trading Claim Workbench

Separate a model score, probability claim, proposed order and observed account result. AI terminology is not a substitute for source data or reserved evaluation.

Your evidence states are not a provider score, audit opinion or performance finding.

0 documented10 unresolved0 not applicable

RECORD 01

Output definition

What does the model actually emit: score, label, probability or prose? A numeric value does not become calibrated confidence merely because it is between zero and one.

Evidence: Output type and target definition

RECORD 02

Training population

Which eligible observations and source versions produced the model? An impressive demo can conceal selected instruments or incomplete historical membership.

Evidence: Training universe and exclusions

RECORD 03

Feature timing

When was each input available at the intended decision? Revised or future-completed values can create a seemingly intelligent but noncausal predictor.

Evidence: Availability contract and revisions

RECORD 04

Model selection

What alternatives and tuning results influenced the final release? Reporting only the strongest candidate hides the search behind the AI claim.

Evidence: Candidate family and selection record

RECORD 05

Calibration claim

What held-out evidence supports a probability interpretation? A confidence label alone does not establish frequency calibration or an individual chance of success.

Evidence: Calibration design and source results

RECORD 06

Reserved evaluation

Which observations were not used for tuning or explanation? Repeatedly inspecting a holdout changes its role even when the model architecture stays fixed.

Evidence: Evaluation identity and access history

RECORD 07

Action translation

How does the output become an entry, exit or no-action decision? Generated commentary and executable instructions require different validation records.

Evidence: Frozen output-to-action policy

RECORD 08

Cost dependence

What data, computation and execution costs enter the proposed implementation? Higher forecast accuracy need not imply positive net account value.

Evidence: Operational and trading cost assumptions

RECORD 09

Human override

Who can change or stop the model action and under what rule? A discretionary intervention should remain visible rather than inherit the original automated label.

Evidence: Override authority and action log

RECORD 10

Marketing boundary

Which result is independently reproduced and which is supplied? Do not convert AI branding into a guarantee, regulator approval or invented subscriber-return statistic.

Evidence: Claim source and evidence limit

Original source, user evidence record and bounded conclusion
Different evidence objects, not a numerical quality ranking.

Primary reference documents

References support source definitions and mechanisms, not a completed review, measured return or endorsement. The questions are this publication's editorial record design.

An AI Trading Score Is Not Automatically a Win Probability

A completed record does not create a model score

A fictional reviewer documents architecture and target, but leaves causal inputs and evaluation population unresolved. The workbench reports those evidence states rather than an AI quality percentage. The references still need independent authenticity and their own source versions. A filled field is not a commissioned audit opinion or proof that a private model was operated.

Imports retain the original ten question identities. Invalid JSON, another publication or duplicate field IDs cannot silently overwrite prior notes. CSV export protects literal values from formula interpretation. The record is a reader aid, not a remote vendor submission or a retained private workspace; actual account identifiers and credentials should not be included.