Skip to content

Work / Verdict

Verdict

A trust platform that grades the data an AI agent acts on, and holds the agent to the same review loop as the humans.

Period
2026-03 to present
Status
live, password-gated, demo on request
Repo
private, walkthrough on request
01

What I decided

  • Rather than trust the client, I put team permissions on the server. The nightly reseed fails closed when its guard isn’t set.
  • An AI agent authenticates with a scoped key and is graded by the same loop that grades the human reviewers.
  • A later correction goes on the metric and never regrades the action. The grade stands. The metric absorbs the loss, and Verdict walks the correction to every committed action that used the value.
  • A disputed input is excluded, not discounted. The dispute removes the contested evidence and the score is recomputed from what’s left.
  • Sessions decide and batches build. Every architecture ruling lands in a ledger whose anchors fail CI when a ruling is reworded. A code batch drains the queue after the week decides.
02

What shipped

  • Live on 2026-08-25 as a password-gated shared demo, a Hono API on Turso with a nightly reseed.
  • v3.0 closed on 2026-08-05. RBAC and governance persistence, the ingestion funnel, live connectors, and an agent in the loop.
  • Multi-tenancy v4.0 to v4.2 by 2026-09-01, after a root cause where the workspace resolver never read the organization id.
  • Build tiers 0 to 5 and organization onboarding W0 to W4 by 2026-10-03, taking the unbuilt ledger count from 34 to 11.
  • Four API keys found shipping in the client bundle on 2026-09-18. Treated as public and rotated, and the leak scan now catches the key prefix.
03

Proof

  • verdict.thetruthlayer.dev, password-gated
  • A user guide written by role, eight roles from VP of Finance to the AI agent
  • Lineage view
    Metric Trust Platform report view. A metric catalog on the left, a P&L table by region with conversion rate, CAC and orders, a banner about three unresolved severity events, and a lineage graph panel on the right.
    A P&L statement report with three pinned metrics, each cell carrying its verdict, and the lineage graph open on the right.
  • Governance loop
    Metric Trust Platform Loop Console with Agent Actions for Claims Bot, Finance Reviewer and Allocation Engine, a Stakes Gating panel of low, medium and high, and token JSON panels below.
    The Loop Console. Evaluate, judge, recalibrate and subscribe in one place, with three agents gated by stakes.
04

Captures

  • Verdict app showing a Reports table of seven rows with owner, folder, status chips such as Draft and Certified, a trust percentage bar, a seven day views sparkline and a schedule for each report, with an icon sidebar on the left.
    Verdict, local seeded demo, October 2026. The Reports list with a trust score beside every report that has metrics.
  • Verdict Governance page with five stat cards for certified, in certification, at risk, open decisions and owners, a row of tabs including Trust Loop and Loop Console, and an ownership matrix heatmap with green, tan and red cells per team.
    Verdict Governance dashboard. Certification counts at the top, the control plane and an ownership heatmap by team below.
  • Verdict Governance page on the Loop Console tab, with an Advance evaluation button, a metric selector, an Inject authority change button, an Agent Actions card listing Claims Bot and Finance Reviewer, and a Stakes Gating card with low, medium and high tiers.
    The Loop Console. Evaluate, judge, recalibrate and subscribe in one place, with agent actions gated by stakes.
  • Verdict Catalog page with concepts, org and source filters on the left and a P and L hierarchy tree on the right, each line showing a dollar figure and an Act, Caveat or Escalate pill, with Net Profit at the top marked Escalate.
    The Catalog in hierarchy view. Net Profit rolls up from leaves, and the trust floor is the worst leaf.
  • Metric Trust Platform report view. A metric catalog on the left, a P&L table by region with conversion rate, CAC and orders, a banner about three unresolved severity events, and a lineage graph panel on the right.
    A P&L statement report with three pinned metrics, each cell carrying its verdict, and the lineage graph open on the right.