gidp

Implementation status

This page records the known implementations of GIDP 0.1, in the spirit of RFC 7942: what exists, who wrote it, how far it goes, and what it showed about the specification. It is not an endorsement and not a conformance certificate. Implementers are invited to add theirs.

Reference implementation (Python)

Location: impl/ in this repository. Licence: Apache 2.0. Author: the specification’s author.

Coverage: the bilateral core (every transmitted object, the session state machine of Section 17.2, the disclosure engine, consents, the Opportunity and the Handoff), plus discovery projections, the dependency primitives, an A2A binding, four worked domains, a probing attack and four limit cases. It tests the specification and has no authority over it: where the two disagree, the disagreement is logged in impl/SPEC-ISSUES.md and the text is fixed.

Independent clean-room implementation (TypeScript), October 2026

Implementer: a large language model (GPT-6 Astra Max), run in a fresh session with no access to the reference code, to the companion documents or to any outside source, under an evaluation protocol designed to make silent invention visible. The code is not published; this entry reports the method and the results.

Given: the specification, the exported schemas, the profile manifest format and one profile instance (co-investment), and a set of thirteen conformance scenarios stripped of their expected outcomes. Required: an implementation of the bilateral core configured by that profile; a journal in which every choice the text did not determine was recorded, with the passage that failed to determine it, before the code that made the choice; then a single run of the blind scenarios, with any correction made after seeing them journalled separately. The journal declares two exceptions to the rule of recording before coding: one choice was coded before its entry was written, though before it was used, and one check was written without an entry and withdrawn after the blind run exposed it.

Results, compared afterwards with the full scenarios:

• The blind run met 77 of 80 expected outcomes, and 79 after its journalled corrections. None of the misses was a misreading of the specification: two came from defects in the scenarios, one from a scenario expecting more than the text requires.
• The journal recorded 21 decisions the text left open. Most are legitimate implementer choices (storage, clocks, binding conventions); seven are gaps in the specification and are recorded for 0.2 in spec/open-problems.md.
• Unprompted, the implementer arrived at the qualification rule the 0.2 session-model draft adopts (the initiator’s transition gated on the responder’s last reported status) without having seen the draft.
• Its audit of its own code also found two defects in this reference implementation, which the text already forbade, and one in the exported schemas. All three are corrected in this release (impl/SPEC-ISSUES.md, E-01, E-02, E-09).

What this does not show: that two implementations by different people interoperate on the wire (the second implementation was run against the scenarios, not against the first), that the specification is complete, or that a model-written implementation substitutes for a human one. The next step is the same test run by a human implementer, or by a model of another family, against the published scenarios.

The conformance scenarios, their blind variant and the profile format are developed on the 0.2 branch of this repository (0.2-dev).

This site is open source. Improve this page.