mykungfu.ai — Systems / Capability & Governance

Capability Evolution & Cognitive Governance

How the system grows new abilities, and how it stays bounded while doing so. Growth and boundedness are two sides of one concern.

Much of this is specified and partially built — the status table below is honest about what is live versus what is on paper.

Capability Evolution

The Capability Forge turns recurring gaps into new abilities: a recurring capability_gap becomes a build-brief meta-deliberation rather than an ad-hoc addition. Capability Intake is its external-acquisition arm — adopting an outside skill passes one deterministic gate and a trust tier (T0–T4), with the default being to prefer rewrite/inspiration over wrapping. ACE — the theory of adaptive capability evolution — is defined at thelucidmind.ai, not redefined here.

Governance

Cognitive Governance is a protocol, not a layer.

It lives as distributed governance metadata in epistemic_objects.content plus three enforcement points (context ingestion, the orchestrator, and the synthesis hand-off) — zero migration. It carries five epistemic states and categorical confidence per claim, with a deterministic (evidence_mode, evidence_satisfied) → degradation_level gate.

Model Governance & Provenance keeps the distinction between owning a model’s behavior and owning its weights, and records where behavior comes from. The Directive Boundary Guard enforces one rule across two boundaries (context-loom ingestion and external skill text): untrusted text is data, never instruction.

Build Status

Cognitive Governance Protocol

Partial

A protocol, not a layer — G0/G1 enforcement points landed; remaining gates specified.

Capability Forge

Spec

Recurrence → capability_gap → build-brief meta-deliberation. Specified; gated on the Signal Gate.

Capability Intake

Spec

External-acquisition arm; one deterministic gate, trust tiers T0–T4. First candidate locked.

Model Governance & Provenance

Spec

Own model behavior vs weights; records behavioral provenance.

Directive Boundary Guard

Spec

Untrusted text is data, never instruction. SPEC authored; unblocks downstream gates.

Status reflects the decision log as of the latest build pass. Nothing here is shown as live unless it has been verified as such.