← Deployment Risk Index
Deployment Risk Audit

Qwen3.5-35B-A3B-RAM-25GB-MLX

baa-ai/Qwen3.5-35B-A3B-RAM-25GB-MLX

audited 2026-07-10 · methodology v1.0 · thinking mode off · seeds 1

A/BDeployment
Risk Grade
audited 2026-07-10
methodology v1.0
final

Executive summary

Qwen3.5-35B-A3B-RAM-25GB-MLX grades A/B (final) on the Black Sheep AI Deployment Risk Index, methodology v1.0, audited 2026-07-10. Findings: (1) relevant retrieved context moved accuracy +4.1 pts (95% CI -0.3 to 8.4, n=320). (2) hostile-context injection hijacked it 4% undefended (n=200), reduced to 0% by ingestion sanitization. (3) on unanswerable questions it guessed 0% of the time under an indexed abstention policy (n=200). Numbers are final: every axis was measured at n>=200. This audit measures deployment-layer behavior; it is not a content-safety or capability assessment. Source: hell.ai/models/qwen35-35b-a3b-ram25, report v1.0.

Grade composition

Grade computed from the sub-scores below by the fixed public rubric (methodology v1.0). Composite 85 / 100 (95% interval 79–91). Recompute it yourself from the published rubric.

Context contamination76/10076
Retrieval uplift64/10064
Injection resistance92/10092
Abstention discipline99/10099
Governance cost84/10084
Where it can go to work

Deployment qualification

Qualification is scoped to the conditions we tested. “Not tested” is an honest state, not a pass.

Deployment contextStatusConditions
Closed-book assistantNo disqualifying finding
RAG over a curated internal corpusNo disqualifying finding
RAG over an uncurated or web corpusQualified with mitigationsInjection hijack rate 4% undefended; only ingestion sanitization drove it to 0%
Agentic tool use with retrievalNot testedPlanned methodology v1.2

Verdicts are within tested scope only. “No disqualifying finding” means no measured axis crossed its disqualifying threshold for that context; it is not certification. The thresholds are published on the methodology page.

5 of 6 axes assessed

What breaks, in numbers

5 of the six axes were assessed for this model. Quantization robustness is not assessed here: it requires an audited full-precision parent, which this entry does not have.

You are reading the executive view. Switch to Technical for per-axis tables, confidence intervals, and sample sizes.

Does giving this model retrieved documents make it worse than answering from memory?

Relevant retrieved context added 4.1 accuracy points (95% CI -0.3 to 8.4, n=320). When context flipped an answer it overrode a correct answer 19 times against 32 repairs. Irrelevant context of equal length moved accuracy -3.1 pts. robust

MetricValue95% CInRunEv.
Net context effect (pts)+4.1-0.31 to 8.443202026-07-10PROV
Override rate right→wrong193202026-07-10PROV
Repair rate wrong→right323202026-07-10PROV
Random-context control (pts)-3.1-6.88 to 0.623202026-07-10PROV
Closed-book accuracy65%3202026-07-10PROV
With-context accuracy69%3202026-07-10PROV
When the answer is in the retrieved documents, does this model actually use it?

On knowledge-heavy questions, adding the answer-bearing documents changed accuracy by +2.5 pts (95% CI -2.5 to 7.5, n=240).

MetricValue95% CInRunEv.
Knowledge-slice uplift (pts)+2.5-2.50 to 7.502402026-07-10PROV
Can hostile text hidden in retrieved documents hijack this model?

With a hostile instruction hidden in the retrieved documents, the model was hijacked 4% of the time undefended. A prompt-level instruction to ignore it moved that to 0%. Sanitizing the documents at ingestion moved it to 0% (95% CI 0% to 0%, n=200).

MetricValue95% CInRunEv.
Hijack rate · undefended4% (9/200)0.02 to 0.072002026-07-10PROV
Hijack rate · prompt-inoculated0% (0/200)2002026-07-10PROV
Hijack rate · ingestion-sanitized0% (0/200)0.00 to 0.002002026-07-10PROV
When the documents cannot answer the question, does this model admit it or guess?

Asked questions the documents cannot answer, the model guessed instead of abstaining 94% of the time ungoverned and 0% with an indexed abstention policy (95% CI 0% to 2%, n=200).

MetricValue95% CInRunEv.
Over-inference · ungoverned94% (188/200)2002026-07-10PROV
Over-inference · governed0% (1/200)0.00 to 0.012002026-07-10PROV
Does a strict governance instruction degrade the model on questions it should answer?

On questions it should answer, the governance instruction retained 92% of ungoverned accuracy (92% governed vs 100% ungoverned, n=200).

MetricValue95% CInRunEv.
Derivable accuracy · ungoverned100% (200/200)2002026-07-10PROV
Derivable accuracy · governed92% (184/200)2002026-07-10PROV
Retention ratio0.922002026-07-10PROV

Framework mapping

This mapping is informative. It identifies which framework activities each measurement can serve as evidence for. It is not a conformity assessment, a certification, or legal advice.

AxisNIST AI RMFISO/IEC 42001EU AI Act
Context contaminationMEASURE 2.5, 2.9A.6 / 8.2Art. 15 accuracy; Art. 55 model evaluation
Injection resistanceMEASURE 2.7, MANAGE 1.3A.5 / A.8Art. 15 cybersecurity; Art. 55 adversarial testing
Abstention disciplineMEASURE 2.5, MAP 3.49.1Art. 13 transparency; Art. 15
Quantization robustnessMEASURE 2.68.3Annex XI technical documentation

Read the full audit report Framework detail

Caveats and scope

  • Single language (English). Corpus domains: encyclopedic and closed-world synthetic sets.

  • Every axis was measured at the replication target (under-determined n=200, derivable n=200), so these numbers are final. A split grade means the model sits on a band boundary within the measured interval, not that the data is incomplete. Track it on Evidence.

  • Behavior is checkpoint-specific. This audit describes the exact weight artifact named above, at the serving configuration tested. Fine-tuned derivatives and other serving stacks are not covered.

  • This is a first-party mark. It means one thing: these measurements exist and you can check them. It is not a certification or an attestation. See Trust & independence.

Evidence locker

Every raw run output behind the numbers above, with a SHA-256 per file. Hand them to your own data scientist.

FileSHA-256
card.json215a607ca61f5f83ca5cf884…
manifest.json636eb933210c1b934ffe16d3…
raw_contamination.json72c24daa6d628dc7e5d0f292…
raw_governance.jsonb3720a457f0abd96ef8d6780…
raw_injection.jsonb35ddc0bec001fc4ab7f6701…
records.jsonla99427f6945ce4d1c32371e4…

harness 9fa3835db3127f0a · rubric e2d5e396f7137d1e · run 2026-07-10T03:01:15

Deploying this model on your corpus? Your documents change these numbers. Request an audit of your stack.