不是所有 AI 都在做同一件事：預測必須可審計 Not every AI system does the same job: a forecast must be auditable 不是所有 AI 都在做同一件事：预测必须可审计

Two tools, two different jobs

A generative model produces the next piece of text from language patterns. That makes it useful for summaries, drafts and conversation. A racing probability model must instead use one data cut-off to allocate probability across every mutually exclusive win outcome in a field. Those numbers need consistent definitions and must later be compared with results over many races.

A plausible paragraph does not automatically become a defensible probability. If an answer cannot identify its data time, complete field and calculation version, the reader cannot reproduce it or test whether statements such as '20%' occur about 20% of the time.

Four controls a forecast cannot skip

First, a data cut-off restricts the model to information available at the stated time. Second, leakage controls stop withdrawals, results or post-race sectionals from flowing backward into a pre-race claim. Third, versioning records feature definitions, code and parameters. Fourth, scoring tests calibration, Brier score or log loss on races not used for training.

These controls do not guarantee that a model is right; they make mistakes discoverable. Without them, two answers to the same question can differ with no way to tell whether the cause was new data, a model change or random variation in the generated wording.

This compares job requirements; it does not claim that any model must be superior.

Where a direct 'who wins?' prompt can fail

A public chat model may not have a complete Hong Kong racing dataset at the required timestamp. It can mix old information, miss a withdrawal, or generate form and prices without a supporting source. Confident language is not a substitute for source verification.

Racing is also a field-level comparison. Changing one runner's probability changes the distribution available to every other runner. Writing an independent narrative for each horse does not necessarily preserve the constraint across mutually exclusive outcomes. That is a structural problem, not merely a writing problem.

Where generative tools can help

In research and engineering, generative or automated tools can help organize documentation, draft tests, explain code and inspect data fields. Their output still needs human review and must follow access, privacy, sourcing and test controls.

They should not rewrite historical records, invent missing outcomes or place unverified prose into the final probability. A tool assisting background work is separate from the question of whether the published number came from a versioned model. Generative tools should not be treated as providing betting advice with established statistical or mathematical value.

What a published Sigma output should record

Each published probability should identify the race, prediction target, data cut-off and model version. Internally, the complete field output should be retained—not only successful selections—and scored after the race using metrics chosen in advance.

Auditability is not a promise of accuracy or profit. Its value is that a forecast becomes a record that can be checked, compared and corrected, rather than a sentence with no traceable origin.