ArcLedger

Prove the human and AI-agent work behind every USD invoice—privately, locally, and with a complete audit trail.

Comment

Why I built ArcLedger

I work with AI coding agents almost every day. They can get a lot of valuable work done, but normal time trackers only understand what a person does at the keyboard.

That leaves a pretty simple problem. When a client asks what actually happened, what was done by me, what was done by an agent, and what should be billed, the answer is usually scattered across timers, agent sessions, repositories, and memory.

I built ArcLedger to put that work into one private, reviewable record.

I conceived and built ArcLedger independently. I am the founder of ArcanEdge LLC, but ArcLedger has not been assigned to the company, so I am submitting it in my individual capacity.

What it does

ArcLedger is a local-first Windows desktop app for tracking human and AI-agent work without spying on the work itself.

The submission follows one focused path:

  • Today recognizes active work and keeps human attention separate from agent runtime.
  • Work Log keeps the complete history, including corrections, review actions, and billing events.
  • Clients connects recognized work to the right client.
  • Review turns verified work into a client-facing summary and a USD statement that still requires human approval.

ArcLedger does not record screenshots, prompts, source code, keystrokes, clipboard content, page content, or raw browser URLs. I wanted enough evidence to explain an invoice without exposing a client's work.

How I built it

ArcLedger has a TypeScript service and domain layer inside an Electron desktop app. The renderer is sandboxed and only talks through a narrow allow-listed bridge. Local state is protected with coordinated file locking, and higher-volume evidence uses an append-only local log.

I had an experimental MVP before Build Week, built with help from GPT-5.4 and GPT-5.5. During Build Week, GPT-5.6 and Codex helped me turn that experiment into a real, focused product.

The biggest decision was cutting the submission down to one complete promise: prove the work behind an invoice. Planning features and the JSON REST/MCP transports are intentionally hidden and unavailable in this build.

How I used Codex and GPT-5.6

Codex was my engineering partner, but I made the product, architecture, privacy, billing, scope, and release decisions.

I used GPT-5.6 and Codex to audit the full workflow, reproduce attribution and integrity problems, make targeted fixes, add regression tests, harden local file handling, isolate the Build Week profile, and verify the exact Windows installer being shared with judges.

The most useful part was not just generating code. It was having a system that could challenge my assumptions, trace problems back to the current code, and help verify that a fix actually held up.

What was hard

The hardest part was keeping human effort and agent runtime honest. Agent time is not human labor, so ArcLedger stores and explains them separately instead of blending them together.

The second hard part was trust. An AI-generated summary should not become billing evidence just because it sounds good. ArcLedger ties each saved summary to the verified facts and refreshes it when those facts change.

The third hard part was knowing what to leave out. ArcLedger can grow into a broader platform, but the Build Week version is stronger because it stays focused.

What I am proud of

  • A privacy-first record of human and agent work without surveillance.
  • A full path from recognized work to a human-approved USD statement.
  • Automatic on-device summaries that stay tied to verified facts.
  • A purpose-built Windows package with checksums and exact release evidence.
  • A submission build where excluded capabilities are actually unavailable, not just hidden in the UI.

What I learned

I learned that AI-assisted development is most useful when the model helps with criticism and verification, not only implementation.

I also learned that polished AI output is not evidence by itself. Provenance, invalidation rules, and human approval matter.

Most of all, narrowing the product made the story clearer. ArcLedger does not need to do everything in this demo. It needs to prove the work behind an invoice.

What is next

After Build Week, I want to run a small signed Windows pilot, finish the legal and privacy review, and get feedback from service businesses that combine human and agent work.

From there, I can expand the platform without weakening the privacy and evidence rules at its core.

Share this project:

Updates