Compliance / Federal
Complete · all six steps
Rust · air-gapped · CPU-only
SENTINEL
Compliance evidence that fails closed — because a false pass is the one error you cannot take back.
A defense contractor has to prove 110 NIST 800-171 controls and report a score to the DoD. SENTINEL scans the machine, evaluates all 320 assessment objectives using NIST's own published text, scores against the official DoD point table, ranks the fix order by points, seals the evidence in a hash chain, and re-checks later for drift. No cloud, no phone-home, no language model anywhere in the engine — it runs on an air-gapped box, on the CPU that is already there.
sentinel — assess
$ sentinel assess --files ./share --evidence evidence.json
✔ posture scan · 11 probes · 1 unreadable → recorded Absent, not assumed
✔ 320 objectives evaluated · 9 pass · 2 fail · 309 need evidence
✔ DoD SPRS -198 of 110 · 1 control met · 107 unproven
✔ CUI enclave computed · certify the assets it is on, not all of them
✔ evidence sealed · 16 facts · SHA-256 chain · signature VALID
⚠ NOT_READY — unproven is never a predicted pass
Representative output — not a live instance.
What it does
●Every objective, from NIST's own text — all 320 assessment objectives in SP 800-171A, generated from the official published CSV with the determination statements verbatim, never paraphrased. Fifteen carry a real automated check anchored to the specific statement it evidences; the other 305 are declared manual and ask a tailored question. I do not invent a mapping to make a number look better.
●Fails closed, by written policy — a Pass requires affirmative evidence. Missing, unreadable or ambiguous evidence becomes Fail, Substantive or Unknown with a question attached, never an optimistic Pass. The two errors are not symmetric: a false pass detonates in a real assessment and kills the tool's credibility permanently, while a false fail is recoverable friction — the contractor attaches proof an assessor would have demanded anyway.
●The official DoD score, not an approximation — scored against the published DoD Assessment Methodology point table, all-or-nothing per control, so one unknown objective sinks the whole control. The two partial-credit controls are honoured and the floor is enforced. A first scan of an unprepared host lands near that floor, which is the honest answer, and it climbs as evidence is supplied.
●Finds unmarked CUI — three layers: a frozen grammar that matches a control marking only in marking position, strict token shapes for contract numbers and CAGE codes, and a weighted scorer for documents that are substantively controlled but carry no marking at all. Reads plain text, Office files, PDF text layers, and scanned pages through OCR. Anything it cannot read becomes Unknown and scopes in — never Clean.
●Seals the evidence — every fact is hashed, chained and committed to one seal, optionally signed. Re-sealing reproduces it exactly, so an assessor can check it independently, and a single changed byte names the entry that moved. It proves integrity and a commitment value; it is not a timestamp authority and the documentation says so.
●Collapses the paperwork instead of dumping it — the ~300 inherently manual objectives are not 300 contractor problems. They are the standing policy set, identical for everyone, and they group into the 14 documents a contractor actually hands over. So the report leads with the handful of machine-decided gaps and turns the rest into a gather-once checklist that stops re-prompting once supplied.
●Watches for drift, without a daemon — re-scan against a sealed baseline, diff at control level, and alert on any regression with the DoD points it cost. Scheduling is the operating system's job; an air-gapped compliance tool should not be opening sockets or holding a process open.
●The prime-contractor view — fold many subcontractor results into one portfolio and rank suppliers by confirmed exposure. A sub that supplied no evidence at all ranks above a sub with proven passes, and is bucketed "unproven" rather than "cleared". A human's decision to accept a finding is recorded alongside the machine verdict and never replaces it — there is deliberately no disposition that means "pass".
Things I built, measured, and then turned off
●An OCR confidence gate — the obvious idea: trust the scan when the engine is confident. Measured, it does not work. A correctly read marking scored 41 while a misread one scored 64, because markings are inherently low-confidence tokens. Rejected. Severe scan degradation is an honest limit of OCR, so findings derived from scans are labelled as such and flagged for a human instead of being quietly cleared.
●An image preprocessing pipeline — binarize, denoise, deskew, upscale, all built and all measured on the real fixtures. It did not help, and it actively hurt clean input by mangling the separator inside a marking. Off by default, opt-in for operators to calibrate on their own bad scans. Measuring first stopped a harmful default; that is a deliberate non-result, not a failure.
●A GPU path — built, verified bit-identical to the CPU result on real hardware, and honestly slower here. CPU-only is the product, not a fallback: every prospect has a multi-core CPU, none is guaranteed a GPU, and a GPU dependency is a non-starter on an air-gapped box. The code stays dormant behind a feature flag and I do not pitch it.
●A threshold a practitioner told me might be wrong — two range checks were flagged in review as possibly too strict, which would fail a contractor who documented a different number. I left the logic unchanged pending their answer rather than substituting an unvalidated guess, and recorded it as open. Loosening a check to make more people pass is exactly the wrong direction of error.
●A self-audit of my own mappings — I graded all 15 automated checks against NIST's authoritative test text before anyone else could: 7 aligned, 4 partial, 3 stretch, 1 procedural. The three stretches are named in the document as mine to fix. It is cheaper to catch my own weak proxies than to have an assessor catch them.
Proof
320/320
assessment objectives verified in the pack
0
tests green · clippy -D warnings clean
0
DoD controls scored on the official point table
0% / 0%
false positives / negatives · 74-document CUI corpus
0
checks in the gate, all required before "done"
0
documents the ~300 manual objectives collapse into
Rust · eight cratesTOML signature packsSHA-256 hash chainEd25519PDF + OCR frontendsrayonno network · no async · no LLM in the engine
ℹWhat the zeros mean, and what they don't. The 0% false-positive and 0% false-negative figures are measured against a 74-document corpus I generate deterministically. Twenty-eight of those documents are near-miss hard negatives written to be wrong in the exact ways a detector fails — including a public training agenda that discusses the marking in prose — and fourteen more are edge traps checked byte-level. It is a synthetic corpus, and it is the corpus, not the world. A perfect reading is a defect until proven otherwise, which is the whole reason the hard negatives exist. The 15 automated control mappings are still pending review by a registered practitioner.
🔒Complete, runnable code — all six steps of the assessment journey in one binary, one run folder, one consistent story. It is not a certification and every report says so; that is an assessor's job. Not yet sold. Source, signature packs and binaries are private — this is a capability showcase, not a distribution.
Have to prove
you're compliant?
I build the kind of evidence that survives someone checking it — fails closed, cites what it actually observed, and seals itself so a change is visible. Tell me what you have to prove and who is going to audit it.
Start a project →