A deterministic software demonstration
A source-linked administrative draft.
Fictional pump-job records show what happens when a coordinator is missing, dates and status disagree, and a source changes after review.
SYNTHETIC DEMONSTRATION. DRAFT. NOT AUTHORIZED FOR MAINTENANCE.
Prepared JSON fields, not extraction or interpretation of real maintenance documents. No equipment connection, safety decision, customer outcome or operational deployment is represented.
The draft stays incomplete.
Job J-SYN-100, equipment P-SYN-101. A similar asset, P-SYN-101A, belongs to another job and is kept separate.
Missing
No current source supplies the coordinator. The draft leaves it missing.
Conflicting
November 3 and 4, 2026 both appear, with draft_ready and deferred status. Neither is chosen as approved.
Human review
Unreadable and seeded hostile input are visible exceptions. A recorded review retains unresolved items and authorization NONE.
Each included value links to its document ID, revision and passage, with file and passage hashes. Superseded, unrelated and similar-asset records cannot fill the missing current facts. Currency comes from the supplied manifest; it is not inferred from a higher revision number.
The browser-readable pack is a frozen snapshot. Source-linked views have display-only wrapping for readability. The original export and records are available in the public evidence. Current freshness is checked by the CLI retrieval/query path, not this web page.
Inspect the test scope and rerun it.
The recorded worker run reports 34 tests passed, zero failures and errors on Python 3.14.0. A fresh archive extraction also ran 34 passing tests. These small authored fixtures do not establish representative coverage, production reliability, safety or savings.
Independent static review found two real defects in the earlier version: ambiguous export filenames and competing current revisions omitted when ingestion failed. The reviewed revision includes explicit delimiter/case collision and unreadable, missing, invalid and quarantined-current regression cases. The static reviewer checked the corrected package but did not independently run the code or validate maintenance semantics.
- Reviewed Python implementation and test code
- Fictional source manifest, fixture checklist and expected-answer key
- Actual recorded results, test log and extracted-package rerun log
- Review invalidation and abstention scenarios and original draft pack JSON
- Version provenance and reproduction guide and capability limits
Reviewed implementation: 0b5f39af903ee3fed80ab62031365c681d737fc6
Frozen evidence publication: 4cbd1f380f08af387ebf489a64cd1387125d1cf4
Check out the reviewed implementation, then run locally with the Python standard library:
python -m unittest discover -s tests -v
python run_evidence.pyNo API key, model or dependency installation is required for the fixture test. Use fictional records only.
Where the demonstrated ability stops.
The prototype groups four predefined administrative fields, retrieves exact job/equipment/date matches, stores local drafts and reviews, and preserves source references. It does not understand arbitrary documents. Three quarantine patterns are tested against seeded hostile text; this is not general AI defense. Hashes detect changes, not trusted authorship or correctness.
No OCR, PDF ingestion, LLM/RAG accuracy, authenticated sign-off, multiuser operation, immutable audit, production backup, managed-site offline behavior or operational integration is established. Real domain review, data handling and customer acceptance remain independent requirements.
Questions to settle before scope.
Recorded software evidence answers a narrow question. Customer facts, qualified domain judgments and contract decisions need their own answers.
Demo
Is this reading PDFs or using AI?
No. The input is a small, curated set of fictional JSON records with four administrative fields already supplied: asset description, planned date, coordinator and administrative status. The code uses deterministic rules, exact identifiers and local SQLite. There is no PDF extraction, OCR, language model, embedding or RAG evaluation.
Demo
Does a recorded review authorize work?
No. Every output remains a synthetic draft with authorization NONE. Missing fields and conflicting values stay unresolved. A source change makes the old review stale; the stale fact-query path abstains and a new draft starts unreviewed. The reviewer name is an unauthenticated CLI string. Files and SQLite are editable, and source/review checks are not atomic with external edits.
Domain
Does this demonstrate mining maintenance experience?
No comparable mining deployment or named mining-domain reviewer is established by this example. It demonstrates a narrow software pattern over prepared administrative records. A qualified maintenance, document-control and safety reviewer must define real source authority, required fields and approval responsibilities. The demonstration never determines equipment fitness, selects isolation points, supplies maintenance instructions or authorizes return to service.
Customer
Can it integrate with our systems or run at an offline site?
No CMMS or document-management integration, authenticated multiuser review or customer deployment has been implemented or tested. The local demonstration needs no model/API service. Tests observed no connection attempts at three socket entry points during selected flows; this is not network isolation or managed-site offline validation. Actual system versions, supported access, permissions, hardware and operating requirements need assessment.
Customer and security
What would need answering before real records are used?
The customer must define permitted information and outputs, authoritative revisions, representative formats, reviewer roles, acceptance thresholds and failure handling. The data/security owner must approve access, storage and processing locations, logs, backups, retention, deletion and incident responsibilities. This example uses only fictional local files; it establishes no production security certification or data-residency arrangement.
Contract
What are the scope, ownership and support terms?
Those are settled in a written scope and quote. Implementation, ongoing management, third-party costs, ownership and licensing, account handover, support coverage and continuity must be explicit before commitment. The demonstration does not include a custom build, unlimited assessment or support entitlement. Start with one workflow and consider a simpler process or existing-system feature first.
SYNTHETIC DEMONSTRATION. DRAFT. NOT AUTHORIZED FOR MAINTENANCE.
Bring one administrative workflow, its current systems and the questions the evidence has not answered.
See how we workCompare fit and responsibilities