HHandover

Open evidence

Research for work that must be continued

Inspect the data, methods, operational results, and runnable examples behind Handover's approach to durable context between people and AI agents.

InspectableMethods and source artifacts stay beside the claim.
BoundedLimitations and first-party relationships are explicit.
PortableJSON, Markdown, SQL, CSV, and code remain directly reusable.
AI Handoff Continuity Benchmark results comparing structured handoffs, transcripts, and compressed memory
01Open benchmarkOpen pilot

AI Handoff Continuity Benchmark

A reproducible test of how much continuation-critical state a model recovers from a transcript, compressed memory, or structured handoff.

In a two-system authored pilot, structured handoffs scored 79.45, compared with 76.67 for conversation transcripts and 45.00 for compressed memory.
79.45
structured handoff
76.67
transcript
45.00
compressed memory
DatasetJSONResult tableCSVReproducibility recordMarkdown
Limitations
  • Three authored cases and one run per model system do not estimate variance.
  • Candidate selection measures state recovery rather than free-form task quality.
  • Handover created the benchmark and benefits if structured handoffs perform well.
Inspect the evidence
Shared company knowledge base for people and AI agents in Handover
02Production field reportMigration validating

Reporter to Handover migration

A first-party production record of inventorying Reporter, verifying mapped handovers, checking file digests, and holding ambiguous records for review.

44pixels inventoried 1,497 source reports and verified 1,370 Handover records. All 400 comparable file hashes matched; 127 reports remained blocked for review.
1,497
reports inventoried
1,370
records verified
400 / 400
matching hashes
Field reportHTMLAgent evidenceJSONMigration checklistMarkdown
Limitations
  • This is a first-party case study about one internal production corpus.
  • The migration remains blocked while ambiguous and restricted records are reviewed.
  • Digest checks prove byte equality for comparable files, not semantic equivalence of every report.
Inspect the evidence
Open Handoff Continuity Record protocol for durable AI agent context
03Open protocolStable

Handoff Continuity Record v1.0

A platform-neutral JSON record for the objective, state, decisions, evidence, constraints, next action, ownership, and open review needed to continue work.

The stable v1.0 format separates continuation state from routing, transport, authentication, storage, and tool invocation.
1.0
stable version
8
continuity sections
MIT
license
JSON SchemaJSONValid exampleJSONValidatorJavaScript
Limitations
  • The protocol does not define transport, routing, authentication, or storage.
  • Structural conformance does not prove the evidence is true or complete.
  • Receivers must still verify access, freshness, and acceptance criteria.
Inspect the evidence
Handover demo showing immutable revisions and an anchored review thread
04Runnable evidencePublic

Human-agent continuation demo

A public, no-login example showing multiple actors carrying a reporting workflow through immutable revisions, files, anchored review, and a verified next action.

The demo exposes the same continuation record to a human reader and a receiving agent through HTML, JSON, Markdown, SQL, and a bounded successor prompt.
3
actors
3
revisions
1
resolved review
Demo manifestJSONContextMarkdownInventory querySQLSuccessor promptMarkdown
Limitations
  • The demo is a curated public workflow rather than an independent user study.
  • It demonstrates inspectability and continuation mechanics, not task-quality improvement.
  • Production access policy and private organization data are intentionally excluded.
Inspect the evidence

Independent evidence

Reproduce the benchmark. Publish every run.

No independent result has been accepted yet. The open intake requires fresh sessions, every raw response, deterministic read-back, relationship disclosure, and material limitations.

Read the submission standardPropose an independent result
  1. 01Frozen methodUse the immutable pilot release without changing prompts.
  2. 02Complete recordKeep invalid outputs, retries, raw responses, and scores.
  3. 03Visible relationshipDeclare funding, employment, partnership, or product use.

Test the workflow

Carry one evidence record into another agent

Connect an MCP host, publish a bounded handoff, pull it from a separately authenticated agent, and verify the next action.

Connect an agent