Skip to content

Repository files navigation

Judgment Pack evaluator experiments

Empirical evidence about the evaluator proposed by the Judgment Pack Specification's RFC 0006 (Draft), in two tracks that answer two different questions and are judged by different standards:

  • Agreement — do independent implementations produce the same result? Clean-room evaluators derived from the specification text alone. Their agreement tests whether the prose actually pins the semantics; their divergences locate the places it does not. Deterministic, offline, CI-verified.
  • Efficacy — does representing a policy this way actually help? Preregistered experiments on third-party benchmarks. API-dependent, costly, non-deterministic, and run manually — never in CI.

The two are kept visibly separate on purpose. Conformance evidence is something a specification project is the right author of; efficacy evidence is something it is the wrong author of, and the only remedy is method: preregistration, public harness and data, and reported negative results.

Nothing in this repository claims any JPS conformance. JPS 0.1.0-draft defines no evaluator conformance class and forbids evaluator-conformance claims outright (Core §3.4). Every implementation here is experimental, may change or be removed without compatibility promise, and evaluates nothing normatively: a disposition is data, not an authorization, a decision, or an executed action.

Layout

Path What it is
CLEAN-ROOM-PROTOCOL.md The reusable recipe every implementation must follow to count as evidence
reference/ Pinned snapshots of the specification texts the rooms are built from
harness/ The post-hoc referee: runs identical inputs through implementations and diffs dispositions
python/ Agreement track. Implementation #2 (the first lives in judgment-pack-runtime): stdlib-only Python, written clean-room by a coding agent from the reference texts, with its interpretation log in python/DECISIONS.md
studies/ Efficacy, expressiveness, and interoperability track. Preregistered experiments on third-party benchmarks and independently developed systems, from policy expressiveness (001003) through trustworthy-input lineage (005008) to receipt-protocol interoperability (013, 014). See the studies index for the full matrix and the open-source projects each one builds on.
docs/adr/ Decision records for this repository — why a given study is being run, and what comes next

Future implementations (TypeScript, Rust, …) get sibling directories — but the clean room itself never happens inside this repository: an implementer working here could read the existing implementations. See the protocol.

Studies at a glance

Navigation and status only — every finding, and every qualification that finding carries, lives in the study's own directory. Statuses below are the ones the studies index records; it remains the canonical matrix, with the external source each study builds on.

Study Track Status Evidence
001 Expressiveness / efficacy Prepared expressiveness result; comparison arms not yet run Results, study
002 Expressiveness Completed expressiveness result Results, study
003 Expressiveness Completed escape census Results, study
004 Composition Completed composition-closure study; no efficacy claim RESULTS.md, run-log.md
005 Trustworthy input Completed semantic-source-discovery study README.md, RESULTS.md, ANALYSIS.md
006 Trustworthy input Deterministic phase completed; model-authoring phase terminated before inference because of infrastructure, so model usability is not estimable README.md, RESULTS.md
007 Trustworthy input Completed evidence-lineage model replication README.md, RESULTS.md, ANALYSIS.md
008 Trustworthy input Completed offline replay and probe with no model runs; read with the study's stated qualifications README.md, RESULTS.md, ANALYSIS.md
009 Blinded authorship Completed corrected retained attempt as a constructed existence witness — not discovery evidence, not general pipeline validation README.md, RESULTS.json, ANALYSIS.md
010 Blinded authorship Run (frozen) ANALYSIS.md, RESULTS.json
011 Blinded authorship Run (frozen) ANALYSIS.md, RESULTS.json
012 Blinded authorship Frozen + run — R1 unsupported; retracts a published claim ANALYSIS.md, CORRECTION.md
013 Interoperability Frozen + run — R1 holds (both strata) ANALYSIS.md, results/
014 Interoperability Frozen + run — R1 holds (both strata) ANALYSIS.md, results/
015 Interoperability Draft — five review rounds, not frozen PREREG-REVIEW.md
016 Interoperability Frozen + run — R1 holds (both strata) ANALYSIS.md, results/
017 Currency governance Frozen + run — R1 holds (both strata) ANALYSIS.md, results/
018 Currency governance Frozen + run — R1 holds; reviewer holdout diverged on three preregistered cells ANALYSIS.md, results/

This repository claims no JPS conformance for anything in it, and the table above adds no aggregate headline: each study answers a different preregistered question and must be read with its own qualifications.

Results so far

Agreement track

  • 13/13 semantic agreement between the Go reference runtime's experimental evaluator (v0.2.0) and python/ on RFC 0006's nine appendix instances plus three probes — identical kind, outcome, reason set, and handoff state (harness/README.md).
  • Two specification gaps found and recorded as RFC 0006 unresolved questions: number representability, and the disposition's concrete JSON serialization (the two implementations agreed on all semantics while serializing handoff incompatibly — see python/DECISIONS.md entry 3).

Efficacy track

Later studies have retained results; each answers a different preregistered question and must be read with its study-level qualifications. On the expressiveness arms below, no comparison arm has been run in either study. Two expressiveness results, on deliberately opposite policy types and by different model families:

Study 001 — CBA, arithmetic-dense Study 002 — airline, qualitative Study 003 — census, 12 decisions
Prepared facts 124 5 58
Prepared determinations 13 1 40
Decisions with ≥1 determination 1/1 1/1 12/12

The escape is universal in the census frame because real request inputs are collections — and one device dominates: quantification over collections forced 25 of 40 determinations (arithmetic: 6). Encoders were hypothesis-blind; classification was two-way independent with zero fact-class disagreements. See Study 003, Study 002 and RFC 0007.

License

Apache-2.0. This repository is maintained alongside, but is not part of, the normative specification; the specification repository owns the standard and its conformance corpus.

About

Clean-room evaluator implementations generating RFC 0006's evidence. Experimental; claims no JPS conformance.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages