State of Local AI Audit — 2026 H1

Historical report (SIGIL 0.1 semantics). Produced by SIGIL 0.1.0 with the API probe on. PASS did not consider coverage then. Not comparable with v2 results; not re-evaluated. See reports/2026-h1/README.md.

Published 2026-06-07. A 5-model sample of the Ollama public registry, audited with SIGIL 0.1.0. We report license layer presence, SPDX detection, manifest integrity, and verdict distribution — derived deterministically from each model's on-disk evidence, with no LLM in the verdict path.

Aggregate PASS verdict for the 2026 H1 sample: all five models (gemma3:4b, llama3.2:3b, mistral:7b, phi3:mini, qwen2.5:7b) PASS. 5 Ollama tags audited: 5 / 5 PASS; 100% layer presence · 60% SPDX; 5 / 5 manifest digests matched.

Headline numbers

Four numbers a compliance reviewer can quote. The raw per-model AI-BOM JSONs are linked below so the figures can be re-aggregated independently.

Measure	Value	What it counts
Sample size	5	Ollama tags, see Methodology
License layer presence	100 %	5/5 models shipping a license layer
SPDX detection rate	60 %	3/5 license layers matched to an SPDX id
Manifest integrity	5 / 5	models where every blob SHA-256 matched the manifest digest

License analysis

License-layer states partition into three buckets: absent, present but not matched to an SPDX family, and present + matched.

License layer state	Count	Share of sample
Present, SPDX detected	3	60 %
Present, SPDX undetected	2	40 %
Absent	0	0 %

SPDX family distribution

For the models with a license layer matched to an SPDX shortname, the family distribution.

SPDX id	Count	Models
`Apache-2.0`	2	`mistral:7b`, `qwen2.5:7b`
`MIT`	1	`phi3:mini`

Verdict and findings

Each model's AI-BOM carries a deterministic PASS / WARN / FAIL verdict aggregated from analyzer-emitted finding severities. The distribution across the sample is reported as-is.

Verdict	Count
PASS	count: 5
WARN	count: 0
FAIL	count: 0

Total findings: 0. Models with at least one finding: 0.

Runtime exposure

Reported as the default Ollama install on the audit workstation — the host was not specially configured. This is the as-shipped picture, not a hardened deployment.

Default OLLAMA_HOST: http://127.0.0.1:11434
Observed listener class: localhost (single bind 127.0.0.1:11434)
Source: /proc

Per-model index

One row per scanned model. The per-model AI-BOM JSON is the authoritative source for everything above.

Model	Verdict	License	Layer count	Model bytes
`gemma3:4b`	PASS	present, SPDX undetected	5	3.34 GB
`llama3.2:3b`	PASS	present, SPDX undetected	6	2.02 GB
`mistral:7b`	PASS	`Apache-2.0`	5	4.37 GB
`phi3:mini`	PASS	`MIT`	5	2.18 GB
`qwen2.5:7b`	PASS	`Apache-2.0`	5	4.68 GB

Methodology

Five Ollama tags were chosen by license diversity rather than by popularity ranking: Gemma + Llama for vendor-custom community licenses; Qwen + Mistral for Apache-2.0; Phi for MIT.

Sample size: 5 tags from the Ollama public registry
Tool: sigil aibom generate --runtime ollama --format json (SIGIL 0.1; removed in v2)
SIGIL version at run time: 0.1.0
AI-BOM schema version: 1.1
Audit date: 2026-06-06 UTC; summary re-aggregated with the SPDX fast-path fix at 2026-06-14T11:45:40Z in summary.json
Ollama version on the audit workstation: 0.30.6
Verdict path: deterministic severity aggregation in crates/sigil-core/src/ollama.rs:349-356 (at the commit of the report's data)
SPDX detection: shortname + body match over 10 families
Manifest integrity = no finding with id in {ollama.blob_digest_mismatch, ollama.invalid_blob_digest, ollama.blob_missing, ollama.model_not_found}. Surfaced as manifest_integrity.integrity_pass_count in summary.json.

How it was produced

The runner and aggregator are checked in, next to the data. The model list is a plain text file with one Ollama tag per line. They use the SIGIL 0.1 CLI, so they run only at the commit the report was produced with:

git clone https://github.com/ultra-supara/SIGIL && cd SIGIL
git checkout b30d0a2                     # SIGIL 0.1, as this report was produced
sudo apt-get install -y clang jq         # Linux
# Edit scripts/audit-models.txt if you want a different sample.
./scripts/audit-run.sh                   # pulls + runs sigil on each model
./scripts/audit-aggregate.sh             # writes reports/2026-h1/summary.json

Sources and raw data

Per-model AI-BOM JSONs are the authoritative source for everything in this report.

Per-model AI-BOM JSONs: reports/2026-h1/raw/<model>.aibom.json — one per model, schema 1.1.
Aggregated summary JSON: reports/2026-h1/summary.json — the numbers in this report come from here.
Runner: reports/2026-h1/audit-run.sh — pull + scan + emit per-model JSON.
Aggregator: reports/2026-h1/audit-aggregate.sh — jq pipeline that builds summary.json.