@flame0510/project-aether 1.6.1 → 1.6.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/docs/dev/GATEWAY.md +23 -0
- package/models.config.json +937 -45
- package/package.json +4 -2
- package/scripts/model-info-verify.mjs +194 -0
package/docs/dev/GATEWAY.md
CHANGED
|
@@ -206,6 +206,29 @@ block, and every number there carries its source:
|
|
|
206
206
|
No generated field ever replaces a curated one silently: prices are separate files, and the
|
|
207
207
|
modal shows the generated part's date (`asOf`).
|
|
208
208
|
|
|
209
|
+
### Curating an `info` block
|
|
210
|
+
|
|
211
|
+
Nothing here is estimated, and every number carries its source and date. The workflow has
|
|
212
|
+
three steps, one command each:
|
|
213
|
+
|
|
214
|
+
```bash
|
|
215
|
+
npm run info:suggest -- <model id> … # candidate facts from the model card (read-only)
|
|
216
|
+
# the maintainer, or a worker agent asked to follow the same contract, proposes the block
|
|
217
|
+
npm run info:verify -- proposals.json # the gate: refuses what the card does not support
|
|
218
|
+
```
|
|
219
|
+
|
|
220
|
+
- `scripts/model-info-suggest.mjs` prints the card's size sentences, its benchmark tables
|
|
221
|
+
(headers included) and its doc links, so the reading is a lookup rather than a hunt.
|
|
222
|
+
- `scripts/model-info-verify.mjs` checks a proposal against the card: `evidence` present,
|
|
223
|
+
every number in `params` in the card text, every benchmark value in a row named after it
|
|
224
|
+
**under the column that names this model** (never a sibling variant), and `docUrl` when
|
|
225
|
+
asked with `--check-links`. It exits non-zero on a rejected proposal.
|
|
226
|
+
- The proposal contract is what makes this safe: one JSON object per model with an
|
|
227
|
+
`evidence` entry per field, the verbatim line, the column used and the card's own date.
|
|
228
|
+
A worker that skips the evidence — or takes a value from a variant's column — is caught
|
|
229
|
+
here, not in the catalogue. Benchmarks that a card states only for a sibling
|
|
230
|
+
(`-Max`, `-Base`, `-HighSpeed`) are dropped.
|
|
231
|
+
|
|
209
232
|
---
|
|
210
233
|
|
|
211
234
|
## Per-agent model assignment
|