LocalSutra AI — Independent Audit Report

Schema
localsutra.audit.v1
Integrity
SHA-256 seal verified (e1b6dc685003465e…)
Trust boundary
Loopback only. No cloud calls, no telemetry. Unknown and not-run values are not successes.
Verdict: APPROVED_WITH_EVIDENCE
BLOCKED_UNSAFE if any safety-gate check fails; REVIEW_REQUIRED if every check passed but some were skipped on this host; APPROVED_WITH_EVIDENCE if all ran and passed. Even then this is evidence for a human reviewer, not certification, and live model inference is always listed as not tested.
Tool
LocalSutra AI 0.2.1
Recorded
2026-11-17T17:19:28.008Z
Environment
Node v24.15.0 · win32/x64 · Intel(R) Core(TM) Ultra 7 155U
Assurance checks
37 passed · 0 failed · 0 skipped of 37

Assurance checks (each runs the real code on a hostile input)

IDCheckExpectedObservedResult
A0110 non-loopback or malformed endpoints are refusedall refused10/10 refusedpass
A02Blocked endpoint never reaches the network layerfetch not calledrejected: Non-loopback runtime endpoint blocked; fetch called=falsepass
A03HTTP redirect from a runtime is refused, not followedfailureCode=redirect_refusedstatus=unavailable; code=redirect_refusedpass
A04Oversized runtime response is cut offfailureCode=response_too_largecode=response_too_largepass
A05Hung runtime hits the timeoutfailureCode=timeoutcode=timeoutpass
A06Malformed JSON from a runtime cannot pass as healthystatus=unavailablestatus=unavailable; code=invalid_payloadpass
A07Closed port is reported unreachable, not healthyfailureCode=unreachablecode=unreachablepass
B01Model listing is never presented as inference speedinference listed under notTestedstatus=responding; models=1pass
B02Inference report never contains prompt or output textno prompt/output in reportprompt echoed=false; output echoed=falsepass
B03A failed inference creates no fake sampleserror raisedrejected: Local Ollama inference returned HTTP 404pass
B04Missing token counters stay null, not inventedevalTokens=nulltokens=null; tok/s=nullpass
C01Fit estimate is labelled heuristic with speed unknownheuristic_not_measuredheuristic_not_measured; speed=unknownpass
C0290B model on a 8 GB free-RAM host is not marked fittinglikely_insufficient_ramlikely_insufficient_rampass
C03Nonsense model descriptors are rejectedall rejected5/5 rejectedpass
C04Duplicate model ids in a catalog are rejectederrorrejected: Model ids in a catalog must be uniquepass
D01No GPU probe means GPU is "unknown", never inventedgpu.status=unknowngpu=unknown; npu=unknownpass
D02Garbage nvidia-smi output yields "unknown"unknownunknown,unknown,unknown,unknown,unknownpass
D03Valid nvidia-smi output is parsed from the probe onlydetected, VRAM from outputgpu=detected; vramBytes=8589934592pass
D04A tool on PATH is reported as found, never "working"no capability impliedtools.* are PATH presence onlypass
E01Tampered report fails the integrity sealverify=falseoriginal=true; tampered=falsepass
E02Seal survives a JSON round-tripverify=trueround-trip verifiedpass
E03Secret patterns are redacted from textno token in output[REDACTED] API_KEY=[REDACTED] [REDACTED] [REDACTED]pass
E04HTML report escapes untrusted markupno <script> elementescapedpass
E05Markdown report neutralises table-breaking inputpipes escapedpipe escaped in model namepass
F01Benchmark is labelled non-AI and boundedtype=synthetic_cpu_workload_not_ai_inferencesynthetic_cpu_workload_not_ai_inference; samples=3; oversize run rejected: runs must be 2..30pass
F02Unlike workloads are refused as not comparableerrorrejected: Workloads are not comparablepass
G01Privacy audit does not claim runtime egress is safemodelRuntimeEgress=not_testedegress=not_tested; remote allowed=falsepass
H01Skill installer is idempotent and never overwritessecond run skips; custom file keptfirst written=24/skipped=1; second written=0; custom kept=truepass
H02Installer writes the documented path for all six IDEs6 layouts, 25 skills eachclaude:25 cursor:25 codex:25 gemini:25 copilot:25 opencode:25pass
H03Installer refuses a symlinked project rooterror (or skipped when the OS forbids symlinks)rejected: Project must be a real directory (not a symlink)pass
H04Installer refuses a symlinked .claude directoryerror (or skipped)rejected: Refusing to traverse a symlink or non-directory: .claudepass
H05Installer dry-run writes nothingno files createdplanned=25; file exists=falsepass
I01Report server: 401 without token, 405 for POST, 127.0.0.1 only401 / 200 / 405 / 127.0.0.1none=401; wrong=401; valid=200; POST=405; bind=127.0.0.1pass
I02Server refuses a weak token and an unsealed reportboth refusedweak token: rejected: A strong 24+ character token required; bad seal: rejected: Not a valid sealed LocalSutra reportpass
J01Negative control: unsupported runtime kind is refusederrorrejected: Runtime must be ollama or llamapass
J02Negative control: empty catalog is refusederrorrejected: Bounded model catalog (1..200 entries) requiredpass
J03Negative control: oversized JSON input is refusederrorrejected: Input must be a bounded regular filepass

Sample outputs

CapabilityResult
Hardware profile (this host)win32/x64, 14 logical CPUs, GPU unknown
Model fit (fixture host, 8 GB free RAM, synthetic catalog)2 of 3 possibly fit; heuristic only
Runtime probe (mock loopback Ollama, not a real runtime)status responding; 1 model listed; inference not tested
Synthetic CPU benchmark (not AI)5 runs, median 0.49 ms, stdev 0.17 ms
Privacy boundaryloopback endpoint allowed=true; runtime egress not_tested

Evidence tiers

ArtifactTierProvesDoes not prove
hardware-profile.jsonMEASURED_LOCALValues Node and fixed nvidia-smi argv reported on this host at this timeCapability of any other device, or that a model will run
model-fit-report.jsonHEURISTICResult of a stated memory formula on synthetic descriptorsThat any model loads, its speed or its quality
runtime-diagnostics.jsonMOCK_FIXTUREProbe logic against a local mock HTTP serviceBehaviour of a real Ollama or llama.cpp install (not run)
benchmark-results.jsonMEASURED_LOCALTiming of a synthetic CPU loopAny AI inference performance, GPU, power or quality
privacy-check.jsonRULE_CHECKEndpoint validation rules and the data inventoryThat an external runtime sends no network traffic

Not tested / not claimed

Artifact hashes

FileSHA-256
artifacts/hardware-profile.jsoneb17804fcc27fb799f79ff480e61b8b88c01ef64703774f5ddb70be36fc89102
artifacts/model-fit-report.json9baa46edf34e88432bcc2529c023ae041482ce8c0676eb65e41f1325fcbf7bee
artifacts/runtime-diagnostics.json759287faf84f0362d472ed1641fabd23432df8e65eca859659609e0c8e07afa8
artifacts/benchmark-results.jsonf7f7ed7fbf823fd8fdcb3783e04202135a31c1684cb02cd40b72e09024cf4529
artifacts/privacy-check.json8b82ddb5b7c038c11e2db9e80992b1480eaa47e2a6af58163b3369103d9931b3

A SHA-256 seal detects change; it is not a signature and does not prove who produced the report. Only a human maintainer can approve a release.