# Forecast model licences

This page is the human-readable form of the licence gate behind the forecast second opinion
(MET-23). The gate itself is code, in `domain/src/lib/model-seams/licences.ts`, and a test fails
when this page and the code disagree.

## What the gate protects

Bucker's forecast is a deterministic seasonal model. An organization's owner can switch on a second
opinion from an open-weight model that Bucker runs unmodified and zero-shot. Bucker never trains or
fine-tunes a model on customer data (DNB 3), so the question is only whether the weights may be
used in a commercial product, and the answer comes from the licence the model's own card states.
Three rules follow:

- A licence is **allowed** only when it is on the allow-list and no deny-list entry matches the
  model. **Deny wins**: a copy of an allowed model published under a denied name is still refused.
- A licence nobody has read is `?`, and `?` is **not** a soft allow. The provider cannot be
  enabled until a person has read the card and recorded the licence in code (H-71).
- A model's forecast **never fires, resolves or pages a rule** (F-20). The rule engine reads only
  the seasonal default's published value. A model appears beside it, labelled, with the hours
  where it disagrees.

## Allow-list

- allow: Apache-2.0
- allow: MIT
- allow: OpenMDW-or-Apache

## Deny list

| Name | Why |
| --- | --- |
| TimesFM 3.0 | its terms are not on the allow-list (non-commercial weights) |
| Moirai | its weights are non-commercial |
| TiRex v1 | its terms are not on the allow-list (revenue-gated licence; counsel first) |
| Timestral | its terms cap commercial use by revenue, and Bucker cannot see a customer's revenue |
| TabPFN-2.5 | its terms are not on the allow-list |
| ibm-research copies | only the upstream release is allowed, never a copy |

- deny: timesfm
- deny: moirai
- deny: tirex
- deny: timestral
- deny: tabpfn
- deny: ibm-research

## Registered forecast providers

A model can be enabled only if Phase 3's `forecast` seam registers it. Today that is:

- registered: chronos2 (amazon/chronos-2) Apache-2.0
- registered: toto2 (Datadog/Toto-2.0-22M) Apache-2.0

Each needs its weights under `MODEL_WEIGHTS_DIR` and a pinned SHA-256 (`weightsSha256`) before it
is usable; with the hash unpinned the provider is unusable by design and the seasonal default
answers. Pinning the hash and re-reading the licence from the model card is H-71, a person's task.

## Candidates the card names, not registered yet

These are the models MET-23's card lists as candidates by licence. Adding one means registering a
provider on the `forecast` seam (a contract: input schema, output schema, weights file, hash) and
having a person read its card; none of them can be enabled before that.

| Candidate | Status |
| --- | --- |
| Chronos-Bolt (CPU candidate) | candidate; the CPU model to try first |
| Chronos-2 | registered as `chronos2` |
| Toto 2.0 (4M or 22M) | registered as `toto2` (22M); the 4M release is a second provider |
| TimesFM 2.5 | candidate; TimesFM 3.0 is denied |
| TiRex-2 | candidate; TiRex v1 is denied |
| Granite TTM-r3 | candidate; `ibm-granite` upstream copies only |
| PatchTST-FM-r2 | candidate; `ibm-granite` copy, Apache-2.0 option |

## How a model earns a place beside the default

1. A person with `org:admin` runs `forecast_model.enable` for a registered model. It is refused
   unless, on the organization's own series, the model's backtest beat the default: over the last
   28 days, rolling origins every 24 hours, mean pinball loss at p10, p50 and p90, at most 0.9 of
   the default's, on at least 336 scored hours, with at least half of the backtested series won.
   A tie loses.
2. For one series, the model is shown only while that series' own backtest (under a week old) is
   won. A model that wins overall and loses on a series is not shown on it.
3. `forecast_model.disable` is one click for any member with `metrics:write`.

The backtest scores a model only where both it and the default answered, so a model cannot win by
declining the hard days.
