Yash Khoury — Index 2026
pearl
000
Yash KhouryApplied AI builder ·
Exhibit 07 — Agent tooling / provenance← All work

Tamper-evident local execution ledger for AI agents: hash-chained tool calls, decisions, retries, results and verifiable HTML reports.

Stack
TypeScript · JSONL
Year
2026
Last updated
Source ↗
RunLedger: iridescent glass study

See it run

Terminal: three events recorded, verify passes, one line edited, verify fails at event 1

Why it exists

When an autonomous run fails, normal console logs are often mutable, incomplete or scattered across providers. It can be difficult to establish what the agent actually called, retried, decided and returned.

How it’s built

RunLedger stores agent execution as append-only JSONL events linked by SHA-256, then verifies the chain and renders a human-readable run report.

  1. Write one structured event per line to a local JSONL ledger.
  2. Link every event to the previous event with SHA-256 so edits, deletion and reordering are detectable.
  3. Keep the core independent of any single model provider or agent framework.
  4. Render the same ledger into a readable HTML report without requiring a hosted observability service.

Capabilities

Append-only JSONL eventsSHA-256 hash chainTool-call and result eventsDecision and retry eventsTamper verificationHuman-readable HTML reportsNo database requiredLocal-first provenance

Principles

Boring formats are durable
Local evidence before hosted telemetry
Every event should be independently inspectable
Integrity checks should not depend on the agent that created the log

Questions

Does it require a hosted observability service?

No. The current core is a local primitive: JSONL, SHA-256 and an HTML renderer.

How does tamper detection work?

Each event hash includes the previous event hash. Editing, removing or reordering a recorded event breaks subsequent verification.

Does it need a database?

No. The core stores append-only JSONL locally and can verify and render it without a database.

Next exhibit — 08
ActionMesh →