Brain-Implementation-Evidence.md

Anda Brain implementation evidence — 2026-09-23

Historical record. It describes the revision it names. The later memory-brain revision removed the retained draft packages (a single cognitive-memory@2.0.0 remains), dissolved the Cognitive Consistency companion into the Specification and the brain/ companions, and replaced TrialState and DerivationState.

The earlier design review was based on this protocol repository. Reading the actual Brain changes the implementation assessment: a complete service already exists. This report separates that evidence from the new KIP contracts.

Inspected revision and actual run

Repository: ldclabs/anda-brain. Inspected HEAD: f723f96 (clean checkout). Its workspace pins anda_kip = 0.13.1 and uses anda_cognitive_nexus 0.13.4; the Worker uses its own published engine.

Executed locally from that repository:

RUST_MIN_STACK=16777216 cargo test -p anda_brain --lib --all-features --offline

Result: 538 passed, 0 failed, 0 ignored, test execution 21.32 seconds. This ran Rust library tests with real in-memory/local Nexus/storage paths and controlled models/executors. It did not call a live model provider, run the Worker suite, deploy anything or measure held-out learning. The Brain repository was not modified.

What the implementation already supplies

Reviewed concern	Existing implementation	Consequence for this revision
No runnable Brain	Formation/Recall/Maintenance agents plus native Nexus and APIs	Correct the review: the missing evidence was in KIP, not the implementation's existence
Formation ordering	src/agents/processing.rs, src/agents/formation.rs, shared writer admission	Preserve sequential processing; specify portable source-causality requirements for other bindings
Recall limits	src/agents/recall/budgeted.rs, src/recall_budget/, mandatory constraint tests	Preserve honest bounded packets; a packet is not semantic completeness or permission
Learning	src/learning/plan.rs, paired.rs, contracts.rs, native journal/evaluator	Existing pre-registered paired plan runs baseline before trial; prospective controls are an additive contract
Dispatch	src/action/dispatch.rs, native begin_wake_dispatch and durable permit	Define its native-admission boundary honestly; receiver fencing is a separately advertised stronger capability
Learning validation	learning::contracts::paired_trial_replays_a_real_protected_adoption_transaction passed	Evidence of the mechanism, not measured improvement by a real model
Evaluation integration	experiments feature and documented external MIB workflow	Reuse that workflow; do not recreate the retired evaluator inside Brain

Remaining deployment gates

Superseded naming. Package 2.2.0 and the retained 2.1.0 artifacts no longer exist; read "2.2.0" below as the current cognitive-memory@2.0.0 identified by its digest. Current downstream work is listed in the memory-brain revision.

The 538 existing tests do not certify package 2.2.0, new schema IDs, prospective trial enrollment, recording repair, selection-dependency tokens or receiver fencing. An engine/Brain must implement and advertise each corresponding new contract and run KIP2-REL-* scenarios through an actual adapter before claiming that coverage. Do not import the reference-model functions into an engine adapter.

A new package alone grants no deployed capability. Existing 2.1.0 artifacts and schema locks remain available unchanged. Cross-engine Capsule exchange and current-validation mapping need real Rust/Worker adapter runs; REL-016 defines the acceptance scenario.

Real behavioral measurements remain not_run. fixtures/brain-evaluation-not-run.json is deliberately unchanged in status/scores. No corpus, model-provider spending or production executor binding was configured by this protocol change. The release gate is the existing BrainEvaluation workflow, including costs and relevant-memory ablation.