KIP v2 Design Review Resolution

2026-09-06 — all twelve review items addressed in this repository.

Scope: current English normative documents, standard Profile package, reference Brain policies, language toolkit, schemas, fixtures, contract oracles, finite models, adapter runner and CI. Chinese mirrors and the frozen design notes are unchanged. This is a draft-contract revision, not a deployment of external Nexus engines or an empirical claim that a Brain has learned.

Resolution map

Item	Resolution	Primary artifacts	Acceptance
R1	Final BELIEF incorporates applicable slot conflicts; candidate status is diagnostic	Consistency §1, projection schema	MEM-001; SCHEMA-020 revised; grounded/non-functional oracle cases
R2	Stable Skill selects immutable SkillRevision; standing, grades and authority bind revision/digest	Profile, package 2.1.0	MEM-002; revision-reset and authorized-rule checks
R3	Decisions, attempts and observations have distinct identities; quota counts independent attempt aggregates	Consistency §5, record schemas	MEM-003; observation-fan-out fault injection
R4	Protected evaluation policy; explicit comparable controls, sampling/correlation, missingness and uncertainty	Consistency §6	MEM-004/019/022; aggregate-reversal, missing-stratum and unauthorized-rule tests
R5	Immutable TrialRecord/EvaluationRecord and retained exact replay artifacts; mutable states are pointers/caches	Profile §6	MEM-005/020; late-outcome, re-trial, correction and retained-replay checks
R6	Producing/validation Activities pin read inputs; virtual dependency validity gates use before maintenance	Consistency §3	MEM-006; immediate correction, alternative-support and hidden/unavailable-source checks
R7	Complete ProjectionBasis, exact context matching and half-open intervals	Consistency §2	MEM-007; exact boundary, invalid interval, context and cache-coordinate tests
R8	Protected, version-guarded identity-decision withdrawal; preserve supplied references and ambiguous review set	Consistency §4	MEM-008; repair oracle and merge-pointer write rejection
R9	Watch generations/coverage, leases/fencing, durable intent and same-attempt reconciliation	Consistency §7, Maintenance	MEM-009; Watch finite model and restart/fence/recovery contract cases
R10	Compression omissions/re-encoding, separate recall channels, coverage and semantic ErasurePlan	Consistency §8	MEM-010/023/024; partial/held erasure and typed coverage/erasure records
R11	Portable numeric domain; strict JSON/JCS-safe canonicalization and pinned artifact digests	canonical.ts, lower.ts, digest tool	MEM-011; exact-number, underflow, Unicode, duplicate-key, property-preservation and artifact golden checks
R12	Normative Profile status, complete invariant-to-vector mapping, typed artifacts, adapter and separate learning gate	conformance guide, Brain evaluation	MEM-012–025; vector/report/schema checks and CI

The catalogue now has 356 portable vectors, 43 Core invariants and 46 Profile invariants. Every Profile invariant has a vector; this mapping does not imply that every vector has already been run against an external engine.

Compatibility and implementation boundary

• Core still has five element kinds. New process contracts use Concepts, Activities and Facets, and existing KQL/KML shapes; no new untyped mutation language is added.
• The current package is kip://profiles/cognitive-memory@2.1.0. The original 2.0.0 artifact is retained byte-for-byte for migration. Exact package references are not overwritten with different content. Older Skill data is not assigned fabricated revisions, trial enrollment or verified current standing.
• kip-jcs-safe-v1 is a new explicit canonicalization identity. Previous draft digests require an explicit legacy verifier; relabeling them is not migration.
• Integer-valued numbers outside ±9007199254740991 and nonzero underflow now fail rather than silently changing value. Larger exact values use declared string/value object schemas. This tightens the previous language draft and must be handled by clients that relied on the older numeric domain.
• identity_repair and durable_brain_runtime are advertised capabilities; the standard memory Profile requires dependency validation. Unsupported contracts fail explicitly. Record/schema validation and protected evaluation policy are engine obligations, not proof supplied by an Activity class name.

External Rust/Cloudflare implementations need their own adapter results for this revision. The repository's oracles and finite models do not implement a production Nexus and are never reported as one. The Brain evaluation template remains not_run with no measured score; real task improvement is a separate controlled experiment.

Verification

See the consistency verification report for the actual checks and explicit skips. Local commands and engine adapter requirements are documented in conformance/README.md.

Follow-up review of commit 6437e96

The five follow-up findings are corrected in the unreleased 2.1.0 draft. Existing published 2.0.0 bytes remain unchanged; the current package and golden Capsule pins are regenerated together.

Finding	Correction	Regression coverage
Missing transitive validation-schema pin	Generate the complete local reference closure by actual $id, including the HTTPS Change Envelope ID; verify the lock and Capsule package dependency as well as artifact digests	Fresh Ajv instance compiles only manifest-pinned schemas; resolver handles transitive references/cycles and rejects missing resources; MEM-012 strengthened
Ungraded Skills excluded by procedural recall	Proposed/trialed Skills without GradingState remain unproven candidates; existing grades must match the selected revision and validated evaluation; unverifiable adopted claims cannot become validated recommendations	Recall oracle covers absent/current/stale grades, missing evaluations and revoked warnings; MEM-023 strengthened
Monitoring incorrectly required a new promotion	Encode legal state pairs and separate same-state monitoring from trialed → adopted promotion; retaining adoption still identifies its trial and preserves prior adoption evidence under the authorized policy	Schema accepts sparse/insufficient monitoring while rejecting promotion with insufficient/duplicate samples or no trial; bounded model detects the old monitoring quota bug
Revoked Skills could promote directly	Promotion requires current trialed standing; revoked re-entry selects a new trial before adoption can recur	Schema state-pair matrix, promotion oracle and bounded model reject direct proposed/revoked adoption and preserve reachable promotion after re-trial; MEM-018 strengthened
Package constraints retained obsolete grading semantics	TrialRecord freezes the comparison basis; TrialState is only a pointer/cache. Family membership establishes neither attribution nor control membership, and utility needs recorded actual-use attribution	Package constraints, hints and current English architecture/learning examples aligned; independent-attempt, revision/trial and retained-replay checks remain passing

The finite lifecycle model contained the same promotion/monitoring shortcuts; it now covers all four starting states and includes fault modes for both. This is contract/model verification, not a measurement of an external Brain's learning.