How it works

GQI exposes one managed input-to-number system. Its implementation is private; the public contract has three values:

1. A fit request containing task_name, labeled X, finite y, and optional configuration.
2. An asynchronous job with explicit terminal states.
3. An immutable ModelHandle returned only after fitted state is durable.

 labeled examples                         new inputs
   X + finite y                              X_new
        │                                      │
        ▼                                      │
      fit ───── immutable model handle ─────────┤
                                               ▼
                                            predict
                                               │
                                               ▼
                                median + IQR + finite samples

Prediction requires an explicit model choice: the complete fitted handle, or the literal "GQI" for zero-shot. There is no optional adapter field, omitted model default, or fallback from an invalid fitted handle.

Fitting

Automatic fitting measures the encoded dataset, selects the smallest supported context bucket, and manages the fitting and validation budget. The result reports the exact fit/validation rows, optimizer steps, completed epochs, loss, context, and truncation count.

No sign-based filter is applied to targets. Only non-finite values are rejected.

Prediction

Prediction defaults to 16 numeric samples and aggregates their median. Each row contains its finite sample count, IQR, token count, and truncation flag. Saved samples can be disabled without changing the aggregation.

Large requests are divided among warm workers and restored to caller order with row_ids.

Compatibility

GQI is the only public model name. A small set of retired names is accepted with a warning that GQI was used. Old fitted identifiers are not silently accepted as current state; callers must use a current immutable handle.