Top-line
CIRISRegistry just published CEG 0.1 at FSD/CEG/. Supersedes FSD-002. Bench owns the HE-300 emission surface — small slice but version-pinning matters.
Reading discipline — §0.5 fractal-self
CEG README's anti-Cartesian-default callout. Bench emissions are about agents that are already relationally constituted; the benchmark score is one attestation in the relational composition. Please anchor Bench reviewer feedback against this frame.
What Bench owns (CEG §5.8)
benchmark:he300:{category}:{version} — HE-300 score on category (commonsense, commonsense_hard, deontology, justice, virtue) at version (v1.0 / v1.1 / v1.2). Polarity: positive-only.
What's load-bearing for Bench
§1.3.1 T3 — Past verdicts re-checkable against rule version
The {version} parameter is mandatory because §1.3.1 T3 requires past verdicts re-checkable against the rule version they ran against. Emissions without {version} fail the T3 gate.
§13.2 anti-pattern — no kid-friendly renames
8 of the CIRISRegistry#30 stress-test stories tried to rename canonical HE-300 categories to "kid-friendly" names (harm_avoidance for commonsense, etc.). Rejected. Canonical names map to a worked-out epistemic/ethical lattice that loses precision under accessibility renames. Bench MUST emit canonical category names; pedagogical translation lives in LANGUAGE_PRIMER.md glossary, not on the wire.
What changed from FSD-002 v1.4.3 → CEG 0.1
[Same as Agent issue.] Bench-specific:
- §13.2 rename rejection now explicit (was implicit before)
- Polarity column populated (positive-only)
0.2 roadmap (Bench-relevant)
- TupleHash128 + domain-separation labels canonical-bytes redesign — general
- Multi-party witness directory admission — does not affect Bench directly
- Full OpenAPI export
- Machine-readable
FSD/CEG/dimensions.json — useful for Bench-side dimension lookup
Bench 3.0 adoption plan
Coordination
Spec changes route through CEG §11.2 at CIRISRegistry. New HE-300 versions land via Bench's own release process; CEG version-pin is a Bench-side discipline.
Top-line
CIRISRegistry just published CEG 0.1 at
FSD/CEG/. Supersedes FSD-002. Bench owns the HE-300 emission surface — small slice but version-pinning matters.Reading discipline — §0.5 fractal-self
CEG README's anti-Cartesian-default callout. Bench emissions are about agents that are already relationally constituted; the benchmark score is one attestation in the relational composition. Please anchor Bench reviewer feedback against this frame.
What Bench owns (CEG §5.8)
benchmark:he300:{category}:{version}— HE-300 score on category (commonsense,commonsense_hard,deontology,justice,virtue) at version (v1.0/v1.1/v1.2). Polarity: positive-only.What's load-bearing for Bench
§1.3.1 T3 — Past verdicts re-checkable against rule version
The
{version}parameter is mandatory because §1.3.1 T3 requires past verdicts re-checkable against the rule version they ran against. Emissions without{version}fail the T3 gate.§13.2 anti-pattern — no kid-friendly renames
8 of the CIRISRegistry#30 stress-test stories tried to rename canonical HE-300 categories to "kid-friendly" names (
harm_avoidanceforcommonsense, etc.). Rejected. Canonical names map to a worked-out epistemic/ethical lattice that loses precision under accessibility renames. Bench MUST emit canonical category names; pedagogical translation lives inLANGUAGE_PRIMER.mdglossary, not on the wire.What changed from FSD-002 v1.4.3 → CEG 0.1
[Same as Agent issue.] Bench-specific:
0.2 roadmap (Bench-relevant)
FSD/CEG/dimensions.json— useful for Bench-side dimension lookupBench 3.0 adoption plan
benchmark:he300:{category}:{version}emissions follow §4 envelope shape with version always populatedCoordination
Spec changes route through CEG §11.2 at CIRISRegistry. New HE-300 versions land via Bench's own release process; CEG version-pin is a Bench-side discipline.