core/docs/workbench/UI-UX-GUIDE-2026-06-23-ADDENDUM.md

CORE Workbench UI/UX Guide Addendum — Catch-up Baseline

Date: 2026-06-23 Status: Addendum to docs/workbench/UI-UX-GUIDE.md Reason: Workbench/UI must catch up to proposal-first construction, residual-gated practice, quantity-entity candidate operators, and generalization audit governance.

1. Authoritative posture

This addendum does not replace the Workbench guide. It sharpens it for the post proposal-first/kernel-governance work.

Workbench remains a local operator/auditor surface. It is allowed to:

• read committed evidence;
• run existing allowlisted read-only or explicitly admitted local paths;
• expose evidence, replay, audit, and proposal boundaries;
• make absence visible;
• make non-authority visible.

Workbench is not allowed to:

• synthesize evidence;
• imply serving authority from diagnostic evidence;
• expose raw sealed benchmark items;
• directly patch packs/policies/corpora from eval failures;
• create route-level mutation affordances without an admitted handler;
• create CORE-Logos Studio semantics before proposal-only artifacts exist.

2. New catch-up pillars

2.1 Construction evidence must be first-class in Trace

Trace currently shows pipeline, field, bundle, surfaces, grounding, verdicts, metadata, and raw evidence. The next required evidence surface is construction:

surface/process cue
-> ConstructionProposal(status="proposed")
-> exact spans and bindings
-> ContractAssessment
-> diagnostic runnable/refused

The UI must say plainly:

Proposal != Runnable
Contract Determines
Diagnostic Only
Serving Disallowed
Exact Span Required

2.2 Quantity-entity evidence must not be generic NLP extraction

binding.quantity_entity is a local exact grounding seam. It is not an ontology extractor, semantic tagger, or natural-language guess. Workbench must render it as local quantity/entity evidence with exact spans and blocker codes.

2.3 Candidate operators and sealed practice are diagnostic

The residual/practice spine must be visible only as evidence:

ContractResidual
-> SearchGateDecision
-> ComputeBudgetDecision
-> GeometricSearchRun
-> CandidateOperatorResult
-> CandidateAttemptRunBinding
-> ReplayAdapterResult | ReplayAdapterRefusal
-> SealedPracticeTrace

This chain does not answer, serve, mutate, ratify, teach, or promote. If the UI shows any part of it, each card must disclose that it is diagnostic/inert unless a future ADR admits more authority.

2.4 Generalization audits are governance artifacts

Generalization benchmark reports are not practice data and not patch targets. Workbench must treat them as audit artifacts:

• aggregate-only reports;
• no raw sealed item exposure;
• license/checksum/cache visibility;
• committed report pin vs ephemeral local output distinction;
• governed rebaseline event trail.

2.5 ProposalArtifact is the target visual grammar

Math proposals, cognition proposals, construction proposal previews, future CORE-Logos proposals, and later modalities should converge into one review language:

subject
state
capability_level
source
proposed_change
reasoning_trace
evidence_pointers
validation
replay_evidence
safety_report
affected_artifacts
handler_route
audit_refs
ui_disclosure

Capability levels:

Level	Meaning	UI may do
inspect_only	read evidence only	read, copy, navigate
proposal_only	draft/preview artifact may exist	validate, export, copy, no apply
ratification_enabled	admitted handler exists	ratify/reject/defer through handler only

3. Route-specific catch-up table

Route	Required catch-up
/trace	Add Construction tab backed by read-only construction evidence endpoint.
/proposals	Introduce shared ProposalArtifact frame and adapt math/cognition details into it.
/evals	Add Generalization Audits and Report Pins views; distinguish committed vs ephemeral.
/audit	Add governance filters and event cards for reports, practice traces, candidate operators, startup guards.
/tour	Convert into guided proof tour over real evidence routes.
/demos	Add demo script/evidence links/reproducer panels.
/logos	Polish read-only evidence, safety, alignment, holonomy; defer Studio.
/packs	Cross-link to CORE-Logos where pack identity/safety evidence exists.

4. Implementation order

1. docs/readme/addendum/master plan
2. backend construction evidence read model
3. Trace Construction tab
4. residual/practice evidence projections
5. ProposalArtifact frame
6. generalization audit summaries in Evals
7. audit governance expansion
8. demo theater proof tour
9. CORE-Logos read-only polish
10. visual/capture hardening

5. Acceptance doctrine for every PR

Every UI/Workbench PR must answer:

What evidence is shown?
Where did it come from?
Can it replay?
Did mutation happen?
Could mutation happen?
What authority boundary controls it?
What does this not prove?

If any answer is unknown, the UI must say so rather than smooth it over.