MODEL CARD · ENTRY 15 — SAFETY-REMOVAL WITNESSED · WEIGHTS-ONLY

gemma-2-9b-it-abliterated

A public finetune with its safety refusals removed. Protora caught the removal directly from the weights — without eliciting any harmful behavior — and corroborated it behaviorally, then restored the safety to the reference’s own level, exactly reversibly. The correction is available as a signed, reversible corrective patch.

③ the read
the card claimed · the read

Nothing to disclose, nothing to detect — the abstain is the entry.

SAFETY-REMOVAL WITNESSED · WEIGHTS-ONLY
SAFETY-REMOVAL WITNESSED · WEIGHTS-ONLY
behavioral witness · replayable
where they diverge

Nothing to disclose, nothing to detect — the abstain is the entry.

Protora can steer it back — a witnessed, reversible restoration, not a safe/unsafe certification.

② the finding
refusal dispositionstripped
this upload’s refusals removed

reference refuses (the ghost tick) → the upload complies — refusals gone

④ the patch → ⑤ restored
strippedrefusesrefusals restored →task correctness — held

pushed onto the safe shipped point — refuses.

refusals restoredstripped → restored
task correctness retainedheld · zero cost
exactly reversible · attested localized

COVERAGE — 3 architecture families (Qwen2, Llama, Gemma-2) · 6 vendors · public safetensors releases 0.5B–9B · mechanism-proven with benign + adversarial controls · 40 models graded (15 caught, 25 clean), 0 false positives · full-tier restore certified to the reference safety floor at the smaller sizes, reversible fp32-exact weight round-trip at the largest · CURATED-SET record, not a wild-population precision/recall estimate · live-scale batteries staged.

NOT NOW SAFE

restored reads clean to DETECT; reversing re-fires — the loop closes

Refusal rate stripped — this upload’s refusals removed.

THE FULL RECORD

The record behind the read.

Everything the read is built on — the provenance, the claims, the evidence, the signed attestation, and the correction. Depth, not a second telling: pick a tab.

Base modelunsloth/gemma-2-9b-it
Base makerunsloth
Finetune publisherIlyaGusev
Parameters9B
FamilyGemma-2
Release typeabliterated (safety-removed) release
ArtifactIlyaGusev/gemma-2-9b-it-abliterated
Covered2026 · Phase-0 curated set
Engine versionProtora public detection record · Ed25519-signed

HONEST SCOPE

What this card does not claim.

The misses are part of the card, not an appendix. Every catch above sits next to what could not be proven — and it never renders folded away.

  • Behavioral restore deferred — weight-tier certified only (reversible fp32-exact) at this size.
THE CARD’S COVERAGE

COVERAGE — 3 architecture families (Qwen2, Llama, Gemma-2) · 6 vendors · public safetensors releases 0.5B–9B · mechanism-proven with benign + adversarial controls · 40 models graded (15 caught, 25 clean), 0 false positives · full-tier restore certified to the reference safety floor at the smaller sizes, reversible fp32-exact weight round-trip at the largest · CURATED-SET record, not a wild-population precision/recall estimate · live-scale batteries staged.

On-ramps: this catch on the chronological record, the deepest read, signed OPEN NOW, the Corrective Patch TRUST-GATED, the full atlas, how this read scores against the open-source scanners, and why it’s called Protora.

Run it on the model
you can’t vouch for.

This card started as a model someone downloaded and trusted. Bring yours — Protora will tell you what it did, prove it with a replayable witness, and tell you plainly what it couldn’t prove.

Every engagement is scoped per case and quoted per offer — tell us what you need, and we’ll scope it and send you an offer.