Rules first, AI second
Nine deterministic checks run on the graph first. They are fast, certain and repeatable. The model is asked only where judgement is needed, and its answers are written back to the graph as findings with evidence.
Design Verification & Evaluation Toolbox
Every requirement traced.
Every finding proven.
Nothing leaves the machine.
DVET reads your standards, requirement specifications, design documents and test evidence, builds one requirements graph and flags every gap, conflict and nonconformity with a verified quote. Once an engineer has confirmed the findings, it fills in the test report form. The AI model runs on your own PC, fully offline.
requirements-graph · live trace
127.0.0.1
An animation of DVET tracing standard clauses to product requirements, design elements and tests, and flagging the gaps it finds.
- 0outbound connections. A self-test proves it.
- 0automated checks, plus a challenge pass
- 0/143seeded defects found by the rule checks, no false alarms
- 1 nightof unattended machine time for a 300-page standard. By hand: up to 300 hours.
01 — Concept
One graph. Every obligation accounted for.
DVET turns a pile of documents into a single requirements graph. Every normative shall, should and may in a standard becomes a node and is traced through product requirements and design elements to the tests that verify it. Wherever the chain breaks, a finding appears.
No quote, no finding
Every finding from the model must quote the documents word for word. DVET finds the quote in the parsed source before anyone sees it. Answers it cannot verify are discarded and logged.
The engineer decides
DVET is an assistant to the assessor, not a replacement. Every finding and every verdict is a proposal that a person confirms, edits or rejects, and the audit log records who did it.
02 — Workflow
From a stack of documents to a filled test report.
Five stages, run from the command line or with one click on the project page in the browser. Every stage can be resumed. A re-run only repeats what has changed.
-
01
Ingest
dvet ingestReads PDF, Word, Excel, PowerPoint, ReqIF, CSV and text. Every passage gets a clause anchor (7.2.3.1, Annex A.2) and a page, and every file a SHA-256.
-
02
Structure
requirements graphEach shall, should, may becomes a source requirement. IDs, cross-references and quantities with units are extracted and linked.
-
03
Check
dvet analyse9 rule checks on the graph, 12 checks with the local model, and a challenge pass that tries to refute every major finding.
-
04
Review
dvet serveFindings register, traceability matrix, a document viewer with the evidence highlighted, review rounds and an audit log, all on 127.0.0.1.
-
05
Report
dvet reportA verdict is proposed for every clause (P / F / ND / N/A). The official form is filled in place with verdicts, remarks, evidence and an annex.
dvet — offline workstation
$ dvet ingest ./pilot # read every document, anchor every clause $ dvet analyse ./pilot # 9 rule checks + 12 AI checks + challenge pass $ dvet serve ./pilot # review in the browser, on 127.0.0.1 only $ dvet report ./pilot # fill the test report form $ dvet selftest # prove nothing tried to leave the machine
03 — Check catalogue
21 checks. Every one versioned.
Each finding records the check and version that produced it. Model-assisted checks also record their prompt, so a result can be traced back to exactly how it was made.
Deterministic, over the requirements graph. Rule checks alone already cover much of the value of a review.
- TR-01Standard requirement without product requirementmajor
- TR-02Product requirement without design elementmajor
- TR-03Requirement without test or verification methodmajor
- TR-04Orphan design element or testobservation
- REF-01Referenced document or revision not suppliedminor
- REF-02Referenced clause or requirement does not existmajor
- ID-01Duplicate or malformed IDminor
- VAL-01Same parameter, different value or unit across documentsmajor
- DOC-01Revision, date or approval missingminor
A local model with a versioned prompt and a fixed JSON schema. Requirement quality follows ISO/IEC/IEEE 29148.
- RQ-01Ambiguous wording: “fast”, “adequate”, “as required”minor
- RQ-02Compound requirementminor
- RQ-03Not verifiablemajor
- RQ-04Missing tolerance or conditionminor
- RQ-05Prescribes the implementationobservation
- RQ-06Inconsistent terminologyobservation
- RQ-07Incompleteminor
- RQ-08Conflictingmajor
- APP-01Applicability of a standard clausequestion
- CMP-01Design evidence does not fulfil the clausemajor
- CON-01Prose contradiction between documentsmajor
- FEA-01Technically doubtful parameters, raised as questions to the designerquestion
A second opinion on every major finding.
CHL-01 · Challenge pass
A separate prompt tries to refute each major finding. Findings it disputes are marked in the register, with its argument, and can be filtered. Nothing is removed without the reviewer: in a test on the demo project the challenge pass disputed seven findings, and two of them were real defects.
04 — Review cockpit
Built for the assessor’s desk.
A local web interface for the whole review: filter, confirm or reject findings in bulk, jump straight to the evidence in the source document, and confirm verdicts clause by clause. Click a finding to see its evidence.
127.0.0.1:8731/findings
Reviewer: A. Assessor
| Check | Severity | Finding | Status |
|---|---|---|---|
| FEA-01 | question | Hold-up time calculated from nominal, not minimum, supply voltage | proposed |
| CMP-01 | major | Rated output current 0.6 A exceeds the 0,5 A limit | open |
| VAL-01 | major | Minimum password length: 8 vs 6 characters | open |
| CON-01 | major | Event log must survive a power failure, but is kept in RAM | proposed |
| RQ-01 | minor | “start up fast” is not quantified | proposed |
| REF-01 | minor | Circuit diagram DC100-SCH rev C referenced, not supplied | closed |
| Product requirement | Source | Design element | Test | Suggested links |
|---|---|---|---|---|
| PRS-006 | ESO 1000-1 5.2.1 | DD-IO-01 | TC-005 pass | |
| PRS-007 | ESO 1000-1 5.3 | DD-IO-02 | TC-006 pass | |
| PRS-008 | ESO 1000-1 6.2 | — TR-02 | — TR-03 | |
| PRS-009 | ESO 1000-1 6.1.1, 6.1.3 | DD-WD-01 | TC-007 pass | |
| PRS-021 | ESO 2000-4 SR 1.2 | DD-SEC-01 | TC-008 fail | |
| PRS-032 | ESO 1000-1 4.2 | — TR-02 | analysis | DD-TMP-01 · 0.64 |
Gaps highlighted · suggested links come from semantic similarity and close no gap until a reviewer confirms them
| Clause | Requirement + Test | Remark | Verdict | |
|---|---|---|---|---|
| 5.2.1 | Digital inputs of type 1 | PRS-006 → DD-IO-01 → TC-005 passed | P | confirmed |
| 5.3 | Rated output current per channel not above 0,5 A | CMP-01: 0.6 A in DD-IO-02 | F | confirmed |
| 6.2 | Start-up time | No design element, no test (TR-02, TR-03) | ND* | Confirm |
| 7.2.3.1 | Supply interruption at minimum voltage | CMP-01 unreviewed: TC-002 tested at 24 V | ND* | Confirm |
| 6.1.1 | Hardware watchdog | PRS-009 → DD-WD-01 → TC-007 passed | P* | Confirm |
* proposal, not yet confirmed by the reviewer · P pass · F fail · ND not determined · N/A not applicable
Illustration based on DVET’s synthetic demo project. The ESO standards and the DC-100 product are invented.
05 — Privacy
Air-gapped by design. Provable, not promised.
Design files, unpublished test data and licensed standards must not leave the building. DVET has no cloud service and no network traffic. The language model, the parser and the review interface all run on your own PC, and a self-test you can run with the network cable unplugged proves it.
-
Socket guard
DVET patches Python’s network layer at start-up. Any connection or DNS lookup to an address other than this PC raises an error and is logged, even if the calling code swallows the error.
-
Loopback only
The model runtime (Ollama or llama.cpp) and the review interface are reached over 127.0.0.1. DVET refuses any other model address and ignores system proxy settings.
-
A self-test that runs unplugged
dvet selftestruns the whole pipeline on a sample project and fails if anything tried to connect out. It needs no model and no network. -
No telemetry, no update checks
No cloud APIs. Offline flags such as
HF_HUB_OFFLINEandDO_NOT_TRACKare forced at start-up. -
Offline installation
DVET installs from a bundle that carries its own Python and every library, and the models come from a separate model bundle. The SHA-256 of every file is checked before installation. On Windows, no administrator rights are needed.
-
Your files stay put
Everything stays in the project folder, which can sit on an encrypted volume. DVET never copies input files anywhere else.
-
A locked-down interface
A one-time session token in an HttpOnly cookie, CSRF tokens, host checks and a strict Content Security Policy. The pages load nothing from outside.
-
Open-weight models
Apache 2.0 models such as Qwen 3.6 or Gemma 4 run on your own GPU. No account, no API key, no per-token billing, no data retention policy to trust.
06 — Accuracy
Measured, not claimed.
dvet eval scores every check against a gold set of known defects and reports precision and recall per check. These are the current numbers.
0%
Rule-check recall
143 of 143 seeded defects, with no false alarms, on a 301-page reference project
0%
Recall with the local AI
37 of 38 gold findings found by all checks together, using a 27B open-weight model
0%
Quotes verified
Every quote you see is the document’s own text. Answers that cannot be verified are discarded.
0
Failed model calls
Out of 5,950 in the latest full run on a 300-page project
Evidence verifier
case 1 / 2
Model answer · FEA-01
“The hold-up time of 9.9 ms is accepted as 10 ms.”
Source text · DC100-CALC-01 §4
The hold-up time of 9.9 ms is accepted as 10 ms.
- normalise whitespace, quotes, dashes✓
- fuzzy ratio ≥ 0.901.00 ✓
- numbers match exactly9.9 · 10 ✓
- confidence ≥ 0.5✓
VERIFIEDshown to the reviewer, with document, clause and page
How model errors are kept in check
- Constrained output. Answers must fit a fixed JSON schema, enforced by the runtime while the model writes.
- Retrieval per clause. The model sees only the passages relevant to one requirement, so every answer can be traced to its context.
- Verbatim evidence. Quotes are matched against the parsed source. Numbers must match exactly.
- Reproducible runs. Temperature 0, a fixed seed and versioned prompts. Each finding records the model, the quantisation, the prompt’s SHA-256 and the SHA-256 of every input file.
- Challenge pass. A second prompt tries to refute every major finding and flags disagreements for the reviewer.
- Gold-set evaluation. Precision and recall per check before a check is trusted.
Full transparency
AI findings are tuned for recall, so that little is missed. The challenge pass and the engineer sort out the false positives.
| Measurement | Recall | Precision | Scored against |
|---|---|---|---|
| Rule checks · 301-page reference project, 1,860 normative statements | 143 / 143 · 100 % | 100 %, no false alarms | Seeded answer key |
| Rule checks · demo assessment | 26 / 26 · 100 % | 1 known false positive | Gold set |
| All checks with local AI (qwen3.6:27b, 4-bit) · demo assessment | 37 / 38 · 97 % | 36 % | Gold set |
Measured on synthetic projects built for testing. Validation against a real, completed assessment is the next milestone.
07 — Effort reduction
The tedious part runs overnight.
A manual review takes up to 300 engineer-hours, depending on the size of the project, and much of it goes into tracing thousands of shall-statements through requirements, design and tests. DVET does the tracing and checking unattended, overnight, so engineers spend their time on judgement, not lookups.
Manual reviewengineer time, depending on the size of the project
up to 300 h
DVET’s automated passmachine time for a 300-page standard on one GPU, overnight and unattended
8 h 46 m
The engineer still reviews DVET’s findings and confirms the verdicts. That time depends on the project and is not included in the bars.
0 s
to read, trace, rule-check and report on a 301-page standard with 607 product requirements, on a 4-core CPU without a GPU
8 h 46 m
for the full AI review of the same project on one 24 GB GPU: 5,950 model calls, unattended, with no failed calls
0 s
to confirm all 637 verdict rows of the test report form. A bulk decision on 100 findings takes 0.3 s.
0 s
to re-read a project when nothing has changed. Model answers are cached, so the next review round only re-checks what changed.
A review cycle on the 300-page reference project
- eveningDrop in the documents
Standards, specifications, design documents, test logs and the form template. Start the run from the project page.
- + 30 sGraph built, rule checks done
1,860 normative statements traced. Coverage gaps, broken references and conflicting values are listed.
- overnightThe AI review runs locally
Requirement quality, compliance, contradictions, feasibility and the challenge pass. If the run is interrupted, it resumes where it stopped.
- morningThe engineer reviews
Findings come with verified quotes, and disputed ones are flagged. Confirm or reject in bulk.
- doneTest report filled
Verdicts, remarks and evidence go into the official template, with an annex of input SHA-256 hashes and model and prompt versions.
Who does what
Automated by DVET
- Parsing and clause anchoring of every document
- Extracting every normative statement
- Tracing requirements to design and tests
- Coverage gaps, orphans, broken references
- Conflicting values and units across documents
- Requirement quality per ISO/IEC/IEEE 29148
- Draft findings with verified quotes
- Verdict proposals for every clause
- Filled test report form with annex
Kept with the engineer
- Confirm, edit or reject findings
- Decide applicability
- Set the final verdicts
- Sign off the report
08 — Specifications
Runs on an engineering workstation.
Inputs
Standards with numbered clauses or requirement IDs (SR 1.1 RE(1)), requirement specifications, design documents, test logs, calculations.
Outputs
- Open-points list (XLSX)
- Traceability matrix (XLSX)
- Filled test report form (DOCX / XLSX, optionally PDF)
- Annex: input file hashes, versions of checks, models and prompts
- Audit log with integrity check
Report templates
- Official forms, filled in place without being redesigned
- In-house Word templates with placeholders
- Excel checklists through a column mapping
Platforms
- Windows (primary), Linux, macOS
- Windows setup per user, no administrator rights
- Air-gapped install from an offline bundle
Local AI
- Ollama (default), llama.cpp or any OpenAI-compatible server on 127.0.0.1
- Default model: qwen3.6:27b; Gemma 4 as an alternative
- Embeddings: bge-m3, run locally
Hardware
- Recommended24 GB GPU, or Apple Silicon with 48–64 GB
- Minimum16 GB GPU, or Apple Silicon with 32 GB
- CPU only32 GB RAM, for the rule checks and small batches
0automated tests
0 %code coverage
3 OStested in CI on every change
0network access in CI install tests, with every proxy pointed at a dead port
09 — Questions
Good to know.
Does DVET replace the assessor?
No. DVET prepares the evidence and proposes findings and verdicts. Only a reviewer can confirm a finding or set a final verdict, and every decision is recorded with the reviewer’s name in an audit log. In the filled report, verdicts that have not been confirmed carry an asterisk.
Can it run on a PC with no network at all?
Yes, and that is the intended setup. DVET installs from an offline bundle, the models are moved there from a model bundle, and dvet selftest runs with the network adapter disabled.
How long does a full review take?
Without the model, a 300-page standard with about 600 requirements takes well under two minutes. The AI review of a project that size took 8 h 46 min on one 24 GB GPU, so plan it as an overnight batch. A mixture-of-experts model can cut the time further. Later review rounds re-check only what changed.
What about the licence terms of our standards?
Some publishers restrict loading standards into AI or text-and-data-mining tools. Check the terms of your standards subscription before processing them. DVET keeps standards and their parsed text on your PC, in your project folder.
Which document languages are supported?
English in version 1. The modal-verb rules and the prompts are written for English documents.
Does it read drawings and schematics?
Not in version 1. DVET reads text, tables and document structure. A referenced drawing that is missing from the inputs is still reported (REF-01).
Your documents stay home.
Your review is ready by morning.
DVET: design verification and evaluation, with the AI on your own machine.
Back to top