Skip to main content
Noxport Labs
Noxport

ENGINEERING / LOCAL SYSTEMS / TRACEABLE OUTPUTS

We build inspectable systems for difficult data.

Our work spans structured document tooling, source-linked knowledge pipelines, browser state interfaces and local desktop research software.

Discuss an engineering projectExplore technical demonstrationsReview studio products

01. CORE TECHNICAL PILLARS

Four engineering pillars demonstrated by inspectable fixtures

Noxport Labs four engineering pillars, challenges, demonstrable concepts, and honest scope
PillarEngineering challengeDemonstrable conceptHonest scope
01. Document structure & Local AST Compiler (Noxport)Eliminating the 5 Structural Clipboard Defects across 50–150+ turn multi-AI sessionsNoxport source → Typed AST → OOXML (<w:tblGrid>, <m:oMathPara>) / Vector PDF + manifest.jsonAvailable (v3.0.4.226); 100% Local-First Execution
02. Browser State & Truthful Attention Layer (AnswerBell AI)Solving the "Babysitting Tax" and "The Valley of Waiting" (waiting_on_tool)AnswerBell AI 6-state Evidence Reconciliation Engine (3000ms quiet window)In development; Content-Redacted State Receipt V2
03. Verified Evidence Operations Core (YouTube → Gemini Gem)Fixing the 4 Structural Breakdowns of Raw Video Transcript IngestionCanonical Source Manifest → 1:1 raw_lines Normalization → coverage.json → Dual-Format CommitPreview; Epistemic Provenance Contract
04. Local Health Data Plane (Vitalis) & Signal QC Workbench (CardioLab)MCP Confused-Deputy Defense (-32001) & Preventing Unflagged HRV Metric DistortionVitalis Scoped MCP Policy Gate (Dual Baselines) & CardioLab 6-Event Beat Ledger (Δ a.u.)Preview / in-development, strictly non-medical
PILLAR 01Available (v3.0.4.226); 100% Local-First Execution

Document structure & Local AST Compiler (Noxport)

Problem
Standard copy-paste from AI chat threads frequently breaks multi-column tables, strips 4-space Python indentation while injecting stray Copy code labels, duplicates KaTeX/MathML accessibility subtrees, bleeds #090A0F dark-mode backgrounds, and omits unmounted virtualized turns.
Design constraints
Must compile 10 formats locally (0 KB outbound chat payload) with Two-Phase Commit (2PC) credit protection, explicit OOXML <w:tblGrid> dxa column widths and OMML <m:oMathPara> math, and fail-closed virtualized scope reporting (allow_complete_claim: false).
Artifact sample
Pre-export Readiness Card («What will be in the file») + post-export manifest.json receipt + offline Ed25519 NOX1.<payload>.<signature> verification via WebCrypto.
Known limits
Cannot inspect DOM nodes unloaded by host virtualization on Optional DOM Sites unless scrolled into view; AI host DOM selector updates are addressed via adapter revisions.
Inspect Noxport AST & Coverage Fixture
PILLAR 02In development; Content-Redacted State Receipt V2

Browser State & Truthful Attention Layer (AnswerBell AI)

Problem
When orchestrating multi-minute reasoning runs (o3-pro, Deep Research, Claude Extended Thinking) across 10–30 tabs, naive Stop-button timers fire false-positive chimes whenever an agent pauses mid-run in "The Valley of Waiting" (waiting_on_tool) for code execution, search, or tool approval.
Design constraints
Requires multi-signal Evidence Reconciliation across composer button state, streaming markers, a 3000ms DOM subtree quiet window, and monotonic run_epoch sequence deduplication — plus Exact Return (windowId + tabId) navigation and zero prompt/response text capture.
Artifact sample
Content-redacted State Receipt V2 JSON with evidence_signals[], run_epoch, and signature_status: not_implemented.
Known limits
Local DOM observation cannot independently prove remote server-side LLM completeness; cryptographic receipt signing is not implemented in this preview.
Inspect AnswerBell State Machine
PILLAR 03Preview; Epistemic Provenance Contract

Verified Evidence Operations Core (YouTube → Gemini Gem)

Problem
Generic 1-click YouTube AI summarizers erase start_ms/end_ms temporal provenance, split sentences across 2s ASR fragments, silently hallucinate over missing audio intervals (seg-005 at 00:01:52–00:02:19), and suffer cross-video context drift across 10+ transcripts.
Design constraints
Every normalized corpus paragraph must retain an immutable segment_id, monotonic start_ms/end_ms offsets, and a 1:1 raw_lines audit trail, while missing intervals fail closed in coverage.json (source_status: "unavailable", grounding: "source_unavailable").
Artifact sample
Dual-format Markdown + JSONL corpus package with explicit seg-005 unavailable ledgering and strict separation of Source Traceability (Provenance) from Factual Validity.
Known limits
Timestamp anchors prove where a statement occurred in the recording (Provenance), not whether the speaker’s claim is externally true (Factual Validity); never claims Proof of Absence.
Inspect Timestamp Corpus Pipeline
PILLAR 04Preview / in-development, strictly non-medical

Local Health Data Plane (Vitalis) & Signal QC Workbench (CardioLab)

Problem
In wearable archives (Apple Health, Garmin, Oura, Whoop), users face Vendor Truncation vs. Cloud AI Exfiltration and MCP Confused Deputy / Prompt Injection / Tool-Description Poisoning risks. In physiology time series (Polar H10 RR / ECG), unflagged motion spikes or black-box smoothing filters distort RMSSD, SDNN, and pNN50 calculations.
Design constraints
Vitalis enforces pre-execution JSON Schema allowlists on local SQLite (28-day acute vs 90-day chronic Dual Baselines + Calculation Receipts), blocks read_all_raw_records with JSON-RPC error -32001, and flags Stage 4 → Stage 5 egress (LOCAL_ONLY vs cloud AI host). CardioLab preserves SHA-256 raw_au baselines alongside processed_au (Δ a.u.) and a 6-event Beat Ledger (BL-01..BL-06).
Artifact sample
Schema-checked JSON-RPC 2.0 allow/deny exchange (-32001) + 10.0s synthetic waveform comparator (Raw/Processed/Overlay) paired with Beat Ledger BL-01..BL-06 and manifest.json.
Known limits
Strictly non-medical research architectures only; not cleared or intended for clinical evaluation or patient monitoring.
Inspect Vitalis MCP & CardioLab WorkbenchesCardioLab Beat Ledger →

01B. INTERACTIVE PROTOCOL & COMPILER SPECIFICATION DRAWER (DR1–DR5)

Inspect the concrete wire formats, storage journals, and cryptographic invariants

Select any of the four studio engineering protocols below to inspect the target XML/CMap schemas, MV3 two-phase commit (`2PC`) transaction loops, MCP JSON-RPC 2.0 policy boundaries, and offline WebCrypto `Ed25519` verifiers.

4 Inspectable Protocol Specs
Division I · M2R8 Compiler Spec

Deterministic AST-to-OpenXML (<w:tbl>, <m:oMathPara>) & Embedded OpenType/CFF ToUnicode CMap

Instead of wrapping raw HTML in a .doc MIME header or invoking headless print-to-PDF, the M2R8 compiler lowers typed AST nodes into native ECMA-376 WordprocessingML and embeds subsetted OpenType/CFF glyph tables with explicit ToUnicode CMaps for zero-network searchable vector PDFs (pdf_text).

Enforced Engineering Invariants:

  • Tables compile to explicit <w:tblGrid> column widths in twentieths of a point (dxa) so complex matrices never collapse on paste.
  • Display and inline math expressions compile to native Office Math Markup Language (<m:oMathPara> / <m:oMath>).
  • Vector Text PDF (pdf_text) embeds local OpenType/CFF glyph subsets and ToUnicode CMaps with zero external font server requests.
Inspect Noxport AST & OpenXML Fixture
openxml_and_pdf_cmap_spec.xmlxml / pdf-syntax
<!-- 1. ECMA-376 WordprocessingML + OMML Target AST Lowering -->
<w:tbl xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main"
       xmlns:m="http://schemas.openxmlformats.org/officeDocument/2006/math">
  <w:tblPr>
    <w:tblW w:w="9120" w:type="dxa"/>
    <w:tblLayout w:type="fixed"/>
  </w:tblPr>
  <w:tblGrid>
    <w:gridCol w:w="2880"/><w:gridCol w:w="3120"/><w:gridCol w:w="3120"/>
  </w:tblGrid>
  <w:tr>
    <w:tc><w:p><w:r><w:t>Key / Value Lookup</w:t></w:r></w:p></w:tc>
    <w:tc><w:p><w:r><w:t>O(n) linear scan</w:t></w:r></w:p></w:tc>
    <w:tc><w:p><w:r><w:t>O(1) amortized average</w:t></w:r></w:p></w:tc>
  </w:tr>
</w:tbl>
<m:oMathPara>
  <m:oMath><m:r><m:t>T(n) = O(1) amortized vs. O(n) worst-case</m:t></m:r></m:oMath>
</m:oMathPara>

% 2. Vector Text PDF (pdf_text) Embedded OpenType/CFF ToUnicode CMap
/CIDInit /ProcSet findresource begin
12 dict begin
begincmap
/CMapName /Noxport-M2R8-UTF16-CMap def
/CMapType 2 def
1 begincodespacerange
<0001> <FFFF>
endcodespacerange
3 beginbfchar
<0024> <0054> % 'T'
<002B> <004F> % 'O'
<03A9> <2192> % '→'
endbfchar
endcmap

02. STUDIO DESIGN GOALS & ENGINEERING PRACTICES

How we structure specifications and failure boundaries

Note: The practices below describe our studio engineering goals and verification methods on this candidate build; they do not assert that every research concept is a finished commercial product.

Reproducibility & deterministic fixtures

Every interactive demonstration is backed by fixed inputs and unit-tested state transitions so behavior can be inspected independently of network conditions.

Typed interfaces & compile-time predicates

Product statuses, claims, and commercial gates are enforced by strict TypeScript schemas that prevent unverified links or prices from rendering.

Failure-closed unknown states

When source data is missing, off-screen, or interrupted, interfaces transition to explicit review_required or uncertain states rather than optimistic green badges.

Least privilege & explicit review boundary

Design goals favor user-initiated local actions over background collection, and clearly separate on-device parsing from downstream third-party client egress.

03. REPRESENTATIVE ENGAGEMENT AREAS

Technical focus areas

Possible collaboration areas: browser extension engineering, document-format converters, private data pipelines, desktop research tools, frontend proof interfaces.

  • · Manifest V3 extension architecture & DOM extraction pipelines
  • · Structured document compilers (AST to Markdown / HTML / OpenXML)
  • · Source-linked transcript & corpus normalization tooling
  • · Scoped local MCP gateway specifications & signal inspection UIs

04. HOW TECHNICAL EVALUATION WORKS

Evidence-first evaluation sequence

  1. Step 01

    Define scope and data boundary

    Establish what inputs exist, what must remain local, and what cannot be guaranteed.

  2. Step 02

    Inspect sample inputs and criteria

    Construct deterministic synthetic fixtures covering both nominal and malformed/missing inputs.

  3. Step 03

    Agree on interface and verification plan

    Specify typed data contracts, negative test cases, and accessible UI states before implementation.

  4. Step 04

    Evaluate deliverable against evidence

    Verify output against automated tests, browser inspection, and explicit limitation logs.

INTERACTIVE EVALUATION SCOPE CHECKLIST · LOCAL CLIPBOARD BRIEF BUILDER

Assemble a structured technical evaluation brief

Select the architectural pillars and runtime constraints relevant to your review to generate a clean Markdown scope checklist you can copy into your own engineering notes. Zero data is transmitted from this page.

Selected pillars: 2 / 4 (Pillar 01, Pillar 04)
GENERATED MARKDOWN EVALUATION BRIEFlocal-only buffer
# Noxport Labs — Technical Evaluation Scope Brief

> Generated locally on `https://noxportlabs.com/engineering`
> Direct Engineering & Licensing Support: [email protected]
> Target Runtime Constraint: Local Browser Extension (Manifest V3, zero remote relay)

## Selected Engineering Pillars (2 of 4)

### Pillar 01: AST Document Compilers & Coverage Receipts (Division I · Browser AI)
- **Target Deliverable:** Typed AST serializer + coverage receipt schema (allow_complete_claim: false) + explicit off-screen/unsupported node accounting.
- **Verification & Safety Boundary:** Operates only on user-selected DOM turns; zero background tab scraping or silent truncation.
- **On-Site Reference Fixture:** `/products/noxport#sample-export`

### Pillar 04: Local MCP Policy Gates & Signal Workbenches (Division III · Desktop Research Studios)
- **Target Deliverable:** Scoped local MCP tool schema + outbound egress warning gate + raw-vs-processed waveform comparator & Beat Ledger.
- **Verification & Safety Boundary:** Strictly non-medical research/engineering scope; no clinical diagnosis or unscoped raw archive exfiltration.
- **On-Site Reference Fixture:** `/products/vitalis#mcp-demo & /products/cardiolab#signal-workbench`

## Required Pre-Engagement Verification Steps
- [ ] Inspect corresponding deterministic fixture on `noxportlabs.com` (including negative/warning states)
- [ ] Confirm explicit non-goals and failure boundaries (`review_required`, `uncertain`, `-32001`)
- [ ] Share this scope brief with `[email protected]` for architectural or B2B licensing review

DIRECT ENGINEERING & LICENSING SUPPORT

Direct Engineering & Licensing Support

For product licensing support, order recovery, or technical evaluation inquiries, contact [email protected]. You can also use the local evaluation scope checklist above to copy a structured Markdown brief, or inspect our technical documentation today.

Inspect architecture docsBrowse documentation index