OrgX · Reproducible evidence
Run the example. Inspect the limits.
Three packages show what a portable receipt can carry, what survived one cross-client handoff, and why human acceptance stays separate.
The evidence classes are labeled separately: current local execution, historical first-party observations and a synthetic review example. None is independent adoption or evidence of business results.
Reproduce locally
Download and extract the package, then run these commands inside it. The public npm SDK 0.3.0 was checked on October 27, 2026; the checked-out SDK 0.2.1 also passed the same positive and negative cases. The validator runs locally without an account.
npm install --ignore-scripts @useorgx/[email protected]
node reproduce.mjs01 · Measured local execution
Validate a receipt without an OrgX account.
The local runner checks five existing public fixtures, writes the measured output, and emits a receipt that references those exact bytes. A missing-evidence version fails validation.
Schema and reference conformance do not prove that an agent’s claims are true. The local receipt leaves human acceptance pending. No hosted import or paid model request is part of this example.
90-second walkthrough · captions and transcript · no audio. The handoff video explains historical records; it does not record a new live client session.
02 · Historical first-party observations · July 11, 2026
Inspect a failed handoff and the recovery.
One published Claude Code → OpenCode task recovered 0/5 fields without a shared record, 0/5 during an OrgX server outage, and 5/5 in the post-fix run. The five fields were decision, rationale, owner, artifact and next action.
This is one task and one client pair, not current parity across clients. The baseline and OrgX arms used different writer models. These observations do not isolate a causal effect. The outage and raw source records remain visible.
90-second walkthrough · captions and transcript · no audio. The handoff video explains historical records; it does not record a new live client session.
Pinned original run set, failure analysis and raw transcripts03 · Synthetic review teaching example
Keep a valid receipt separate from acceptance.
A receipt can record agent completion and a passing conformance check while a reviewer rejects the work. The downloadable rejection is explicitly synthetic; the actual local run stays pending human acceptance.
The example does not establish reviewer identity, authorization or a real customer outcome. Use actual criteria, evidence and an authorized person’s ruling before accepting work. A producer-supplied acceptance field is not an authenticated human decision.
90-second walkthrough · captions and transcript · no audio. The handoff video explains historical records; it does not record a new live client session.
Before a person accepts the work
- Confirm the requested criteria and the exact artifact version being reviewed.
- Inspect the evidence behind each result, including failures and missing observations.
- Establish the reviewer’s actual identity and authority through the review system.
- Record the person’s ruling separately from agent completion and automated checks.