Case study · Runnable local proof · Hiring managers · Aviation, platform, and AI-governance reviewers
Implemented Synthetic data only Source release pending

A deterministic authority boundary for AI-assisted dispatch

This independent reference implementation asks a narrow question: how can a model help prepare a dispatch proposal without becoming the authority that approves it? The answer is a typed MCP boundary, deterministic policy evaluation, and evidence that can be verified and replayed.

Contents

Why this is the safer and stronger case study

The work is independently owned, so the implementation can be inspected without reconstructing or implying access to a former employer’s proprietary product. It also demonstrates more than a retrospective narrative: the boundary is executable, tested, and deliberately constrained to synthetic inputs.

Role
Architect, implementer, test author, and technical writer
Stack
TypeScript, official MCP TypeScript SDK, Zod, Node test runner
Proof
Two tools, one resource, one prompt, three verdict paths, hash verification, and replay
Last reviewed
August 5, 2026

The publication boundary

Included

  • Synthetic fixtures and explicit demonstration rules
  • Typed MCP inputs and deterministic verdict precedence
  • Hash-linked evidence, verification, replay, and tamper tests
  • Exact limitations and implementation status

Excluded

  • Aircraft performance data, calculations, or operational thresholds
  • Live weather, NOTAM, flight-plan, crew, passenger, or operator data
  • Former-employer product internals or reconstructed architecture
  • Private Flightworks product, SOP, and research material

The authority boundary

The model may propose an action and identify facts. It cannot select or rewrite the verdict. Policy precedence is explicit: an absent required synthetic source produces DENY; a complete packet with a restriction or review flag produces ESCALATE; only a packet passing every demonstration check produces ALLOW.

One scenario, end to end

  1. Proposal: scenario SYNTH-ESCALATE-001 asks to advance a demonstration packet.
  2. Typed input: the packet has a synthetic briefing reference and the explicit flag HUMAN_REVIEW_REQUIRED.
  3. Policy: the required-source check passes; the human-review check does not.
  4. Verdict: the pure policy function returns ESCALATE.
  5. Evidence: the server records the proposal, checks, rationale, prior hash, timestamp, and record hash.
  6. Verification: chain validation succeeds and replay reproduces the stored verdict.

Implemented evidence

Seven automated tests pass. They cover ALLOW, DENY, ESCALATE, verdict precedence, valid chain construction, changed-record detection, and a real MCP client/server handshake over standard I/O. The demonstration run produces all three verdicts, verifies a three-record chain, and reproduces every decision through deterministic replay. 

SYNTH-ALLOW-001: ALLOW
SYNTH-ESCALATE-001: ESCALATE
SYNTH-DENY-001: DENY
chain: VALID (3 records)
replay: VALID

What is real, and what is next

Version 0.1.0 is a runnable local reference slice. It uses an in-memory evidence chain and contains no live aviation integrations. The next publication gate is a durable public source repository, followed by a ten-minute integration guide and a human aviation-domain and confidentiality review. Until then, this page remains out of search indexes and the sample stays outside the featured portfolio tier.

← Portfolio · Authoring process