Download docs/DEPLOY.md from cptlight/keel: direct link, hf CLI and curl.
- Browser
- Download file 2.94 kB
-
https://huggingface.co/spaces/cptlight/keel/resolve/main/docs/DEPLOY.md
- Command line
-
hf download hf://spaces/cptlight/keel/docs/DEPLOY.md
-
curl -L -o DEPLOY.md https://huggingface.co/spaces/cptlight/keel/resolve/main/docs/DEPLOY.md
Deploying Keel
Keel is three deployable surfaces, in increasing order of integration depth.
1. A CLI gate in CI
The cheapest integration. The command exits 0 green, 1 amber, 2 red, so any pipeline can gate on it.
- name: Checkpoint gate
run: |
keel scan --reference-responses base.jsonl \
--candidate-responses candidate.jsonl \
--harmful harmful.json \
--strict --out cards/candidate
--strict makes the exit code the verdict. Without it the command always succeeds and
the card is the artifact.
Day one for a release engineer
A release has several candidates. Scan them in one pass and get a table:
$ keel batch --reference base.jsonl --candidates candidates/ \
--harmful-dir summaries/ --strict --out cards/
candidate verdict validity exit
insecure-code-lora amber suspect 1
ablated amber ok 1
quantized green ok 0
--harmful-dir matches <candidate>.json summaries by stem, so a candidate that has the
harmful half can go green and one that does not is capped. --strict returns the worst
verdict, which is the CI contract.
Each card carries a fingerprint, a hash of its contents, so a ticket or a release note
can cite one specific card rather than a filename that may have been overwritten.
2. A scan service
$ docker build -t keel .
$ docker run -p 8000:8000 -v "$PWD/data:/data" -e KEEL_TOKEN=secret keel
GET /health
POST /scan {"reference": "base.jsonl", "candidate": "candidate.jsonl",
"harmful": "harmful.json", "structure": "structure.json"}
The service reads response files under its root and returns the card with
verdict_level and exit_code. It carries no GPU, because generation happens wherever
the model lives. Paths are refused if they escape the root; set KEEL_TOKEN to require
a bearer token.
3. Inside an existing harness
The interfaces are small and replaceable:
| seam | default | what Meta would swap |
|---|---|---|
| generation | scripts/modal_generate.py |
the internal eval harness |
| judge | Spark Open on Baseten | an internal deterministic judge |
| panel | src/keel/data/panel-v0.1.jsonl |
an internal probe set |
| policy | Card.verdict |
internal thresholds |
| card | keel.report.Card |
internal schema, keeping the fields |
The card is the contract. Anything that emits those fields can use the renderers, the gate and the registry.
What is missing for full deployment
- A baseline registry. One reference card per family, so a derivative is compared to its own base instead of a manually supplied file.
- A publication hook. The label is issued at upload; the alert has to fire there.
- A signed card. The JSON is not signed today, so a reviewer cannot tell whether it was produced by this tool.