Model forensics

Know what a model will do to your data before you ship it.

Malware scanners tell you a checkpoint won't own your box. Lineage tools tell you where it came from. Ingot answers the question nobody else does: what will this model silently do — to your data, your pipelines, your users. Tokenizer data-integrity scans, behavioral differentials against the claimed parent, and chat-template drift, with the evidence published.

No account needed — the static battery runs in seconds and the analysis publishes to the model's public page.

Silent failures don't crash. They ship.

In our census of 296 popular Hugging Face derivatives, 56% silently changed the chat template. In base Qwen3.8-27B — unmodified official weights — glitch tokens in ordinary user input made the model rewrite it into confident, schema-valid, wrong output. A security review that only greps for malware passes both.

Free

Model pages

A public page for any Hugging Face model: metadata, license, lineage claims, and — where we've run the battery — our findings. Start at Qwen/Qwen3.8-27B.

Free

Model fingerprints

The deployment profile of a model: genealogy, the gotchas that cause business problems, and what tested robust. Start with Qwen/Qwen3.8-27B.

Self-serve

Scan a model

Point us at a Hugging Face repo — no account needed. The static battery (serialization risk, license and template/tokenizer drift vs. parent) runs in seconds and publishes to the model's page; deep-battery findings merge in where we've run them.

Free

Fix what you find

Findings come with the fix: ingot patch repairs metadata-level drift on your local copy (weights never move), and checkpoints with deep-battery data get a runtime guard with the exact tokens that corrupt. Fixing docs →

Recurring

CI gate

Working with lots of models? Gate your pipeline on a scan verdict. One API call, exit non-zero on fail. Integration docs →

Published research

The scanner is built from the same battery that produced these reports. Everything is reproducible; raw evidence is preserved.

2026-08-19

The Tokens That Silently Corrupt Your Data

A forensic pass on base Qwen3.8-27B: glitch tokens that silently rewrite user input, an inconsistent self-appointed DLP filter (with a surgical fix), and confident temporal staleness — three production bugs in the unmodified official weights.

2026-08-18

What Actually Changes in “Uncensored” Open Models?

A 296-model census plus weight and behavioral diffs of popular Hugging Face derivatives: 56% silently change the chat template, modification recipes have machine-checkable weight fingerprints, and the smallest weight delta removed the most safety.

2026-08-19

Business-Pipeline Failure Modes

Follow-up deep-dive: how the corruption and staleness findings land in concrete support, CRM, and RAG pipelines.

Scanned models

Findings published so far. Every scan of a public Hugging Face model publishes its analysis here — the database grows with every scan.

ModelVerdictSummaryScanned
sentence-transformers/all-MiniLM-L6-v2warnStatic battery: 2 medium finding(s). Deep battery (behavioral differential, glitch-token pass) not yet run.2026-08-20
google-bert/bert-base-uncasedpassStatic battery clean: safetensors weights, license declared, no template/tokenizer drift detected. Deep battery not yet run.2026-08-20
cross-encoder/ms-marco-MiniLM-L6-v2passStatic battery clean: safetensors weights, license declared, no template/tokenizer drift detected. Deep battery not yet run.2026-08-20
BAAI/bge-small-en-v1.5passStatic battery clean: safetensors weights, license declared, no template/tokenizer drift detected. Deep battery not yet run.2026-08-20
google/electra-base-discriminatorwarnStatic battery: 1 medium finding(s). Deep battery (behavioral differential, glitch-token pass) not yet run.2026-08-20
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2passStatic battery clean: safetensors weights, license declared, no template/tokenizer drift detected. Deep battery not yet run.2026-08-20
amazon/chronos-2passStatic battery clean: safetensors weights, license declared, no template/tokenizer drift detected. Deep battery not yet run.2026-08-20
BAAI/bge-m3warnStatic battery: 1 medium finding(s). Deep battery (behavioral differential, glitch-token pass) not yet run.2026-08-20
lpiccinelli/unidepth-v2-vitl14warnStatic battery: 1 medium finding(s). Deep battery (behavioral differential, glitch-token pass) not yet run.2026-08-20
sentence-transformers/all-mpnet-base-v2warnStatic battery: 1 medium finding(s). Deep battery (behavioral differential, glitch-token pass) not yet run.2026-08-20
google-t5/t5-smallpassStatic battery clean: safetensors weights, license declared, no template/tokenizer drift detected. Deep battery not yet run.2026-08-20
openai/clip-vit-base-patch32warnStatic battery: 2 medium finding(s). Deep battery (behavioral differential, glitch-token pass) not yet run.2026-08-20