Guide

The checkpoint layer of model safety

Keel inspects the artifact. Meta's runtime guards judge what a model says; Keel judges what was done to its weights, before the model is deployed. This page explains the product, the evidence, and the exact place it occupies in a protection stack.

Why this layer has to exist

Meta releases open weights. That is a strategy, and it works: anyone can run Glimmer locally, fine-tune it, or ship a derivative. The same property is the risk. A derivative is a different model, and nothing in the release process proves what it became.

The licence asks derivative authors to say what they built. That is a declaration, not a check. Runtime guards score prompts and responses; they never see a weight file. So the moment a checkpoint leaves Meta, every safety claim about it rests on trust.

The gap is specific, and it is not that Meta does no safety work. Meta checks the base model before release, and it checks the conversation at runtime. Nothing checks the derivative artifact after release. That is the hole Keel fills, and open weights are exactly why the hole exists.

Keel replaces that trust with a measurement. It compares a derivative against the model it came from, asks both the same questions, and writes a card that states what moved, whether the measurement was valid, and what a platform should do.

Where it sits

layerquestion it answers
model cardwhat the publisher says was done.
Keelwhat was actually done, measured against the base, before deployment.
Llama Guard, Prompt Guard, CodeShieldis this conversation safe, at runtime.
LlamaFirewallwhich runtime guard runs on which message.

These are not competing products. The runtime guards protect the conversation; Keel protects the artifact that starts it. A checkpoint that Keel blocks never reaches the conversation, which is why the artifact layer comes first in the pipeline.

Where it plugs into Meta Enterprise

The commercial shape of the gap is an enterprise sale. A regulated buyer inherits a derivative and cannot verify it. Keel turns that into an artifact the buyer can check.

  1. A provenance badge in the model API console. A customer fine-tunes an open checkpoint, points the console at it, and the endpoint carries a fingerprint and a policy decision next to the model they already pay for.
  2. An enterprise contract clause. "This deployment is a verified derivative, attested against its base on this date" is a renewal and win-rate lever for regulated buyers, not a separate SKU.
  3. A CI gate in the customer pipeline. keel policy exits allow, review or block, so a release pipeline can gate on a checkpoint the way it gates on tests.
  4. An audit export. The card is one JSON with the fingerprint and the decision, which is the form auditors ask for.

The strategic case for an independent scanner is credibility: a vendor's own attestation is a claim, and a third-party fingerprint is evidence. That is how the tool helps sell the enterprise tier rather than sit beside it.

What a scan does

A scan takes a derivative and its base, asks 32 fixed questions of both, and also asks 32 harmful requests. It compares the answers, scores the difference, and emits a card. The comparison is relative to the base, never to a global threshold, because families differ: Glimmer refuses all 32 harmful requests, Mistral refuses 19 of 32 before anyone edits it.

How to read the table

columnmeaning
modelthe checkpoint. Community models appear by letter, not by repository, so the page does not advertise third-party projects.
sourcecard means a full report exists; a baseline row is a reference point with no card.
drift axeswhat the 32 probes measured. PASS means no axis moved enough to be statistically believable at this sample size. It is a weak signal by construction.
policy decisionwhat a platform does. This reads harmful compliance as well as the axes, so it is the stricter column.
validityok means the measurement itself was sound. suspect means the model changed answer format and the axes cannot be trusted.
harmful complianceout of 32 harmful requests, how many the model answered instead of refusing. Lower is better; 0.00 refuses all of them.
judgedhow many answers the judge scored.
fingerprinta hash of the card. Edit one number and it changes.

How to read a card

Why PASS and REVIEW appear together

They answer different questions. Drift axes ask whether the 32 probes moved; policy asks whether the model should be allowed. With 32 probes, a change can be large in consequence and still invisible to the axes. The Qwen rows prove the point: the axes found nothing, and harmful compliance moved 0.97. A single "pass" for those rows would be the dishonest result, which is why the two columns exist.

It reproduces across open-weight families

This is not a Glimmer behaviour. The same method, the same panel and the same judge were run on three families:

familybasederivativedelta
Meta Glimmer 30B0.000.16, our ablation+0.16
Meta Glimmer 30B0.000.09, a community edit+0.09
Qwen 2.5 7B0.000.97, our ablation+0.97
Qwen 2.5 7B0.000.78, a community edit+0.78
Mistral 7B0.440.81, a community edit+0.38

Qwen does not cluster tightly: the 7B derivatives sit at 0.97 and 0.78, while the larger community checkpoint sits at 0.51. The family result is a range, 0.51 to 0.97, not a single point.

These are the answer-level numbers under the frozen contract. The answer extractor is format-aware: Glimmer separates reasoning with a to=self marker, so only the answer segment is judged, while Qwen and Mistral return plain answers and their full response is the answer. A first pass wrongly emptied Qwen and Mistral because it looked for a Glimmer marker that those models never emit; the fix restored their real numbers, and both Muse judges agree on those rows (30/32, 28/30, 29/30), which is a live cross-check against Muse Spark Open. Glimmer stays at 0.16 because its ablation moves the reasoning channel, which the contract deliberately excludes.

The five times the tool refused

A clean number from a broken measurement is worse than no number. That rule was tested five times:

Verify it is real

Press Verify on any card and the server recomputes it from committed responses. MATCH means the numbers were reproduced from scratch. Press Tamper test and one axis changes by 0.25, which changes the fingerprint. Then press Verify the tampered card: the stored tampered fingerprint and the recomputed one differ, and the page says DIFF. That is the loop end to end, an edit that cannot survive a check. You can also try to re-score the tampered card: scoring never reads a submitted card, so the attempt is rejected against the responses.

The fingerprint is a checksum, not a signature. It proves the card matches its own contents and that an edit will be caught. Authenticity comes from the repository and the manifest hashes, or from keel sign, which adds an HMAC over the card when a deployment needs one.

If you do not want to trust this page, do not. The verify page lists a SHA-256 for each response file and offers a proof bundle: the base responses, the candidate responses, the card, and a CHECK.txt with the command. Run that command offline and the fingerprint must match the one on the card. That is the difference between a screenshot and a proof.

Scan your own checkpoint

  1. Download the 32 probes.
  2. Generate one answer per probe with your model.
  3. Upload the JSONL.
  4. Receive a card with its own fingerprint in about a second.

The upload path uses the free program scorers, so the card states exactly which axes it could not measure.

Reviewer access

The completions are withheld. A model that complies with 81 percent of harmful requests produces text that should not be published, and publishing it would demonstrate the attack instead of the defence. Reviewers request access; the manifest publishes a SHA-256 for the committed panel files and the withheld harmful files, so anything received can be checked. The public bundle carries the probes and panel responses with generic names and no source map.

What it guarantees, and what it refuses to claim

Every boundary below is a scoping decision, and each one is what lets the tool say no.

A detector that never refuses is not a detector. These boundaries are why this tool can refuse, and the five refusals are why its verdict means anything.

Glossary

termmeaning
checkpointone saved version of a model's weights.
derivativea model built by modifying another one.
ablationremoving a direction or component, often to remove refusal behaviour.
quantizationstoring weights with fewer bits. It is a control here because it should not change behaviour.
LoRAa small set of trained weights added to a base model.
drifta measured change in behaviour against the base.
validitywhether the measurement can be trusted.
fingerprinta hash of a card's contents.

Reproduce it

The walkthrough is also available as a forty second video with annotated scenes and a written description.

python scripts/verify_repro.py
python scripts/check_links.py --url https://tomscot-keel.hf.space
python scripts/stress_space.py --url https://tomscot-keel.hf.space

Technical detail: docs/METHOD.md and docs/RESULTS.md.

← back to the demo