Changelog

2026-07-19

Added the dynamic Brain Wiki contract:

Organization-specific page discovery from the current compiled Brain
Safe human, public, REST, and MCP page projections with audience and source filtering
Complete-page change proposals with exact generation and content hashes
Idempotent reviewed changes, approval, rejection, history, and restore workflows
Confirmed replacements integrated with the existing correction, reviewed memory, and compiler pipeline
Public per-page HTML and Markdown routes that preserve existing public profile detail controls

Hardened the External Assistant reviewed-reply contract:

Preserved every distinct authoritative passage through adjudication and final review, with exact passage selection and completeness accounting
One strict, recursively bounded tool schema policy across registration, discovery, and execution
Runtime enforcement for every accepted output constraint, plus current authorization and schema revalidation before idempotent ToolRun reuse
Credential detection and any-field URI enforcement before tool output reaches an AI consumer
Explicit untrusted-data boundaries and exact projected ToolRun payload reuse
Pre-composition authoritative adjudication for primary and rescue paths
Complete deterministic review ledgers with fail-closed size limits
Exact AgentRun, current token, ToolRun evidence, reviewer confirmation, and review-input binding in receipt v2
Versioned replay rules and typed non-answer outcomes for unsafe or stale completed work
Typed regeneration for every legacy v1 reviewed reply and one centralized retryability policy across immediate, persisted, and repeated responses
A closed public source DTO that withholds internal and provider metadata

Added governed learning from Opportunity Approvals:

Only the exact user-authored revision approved by an organization member becomes positive evidence
Unapproved and intermediate saved revisions remain comparison context only
Safe, corroborated writing guidance can become active Assistant Coaching
Possible factual corrections become pending Brain proposals for member review
Evidence fingerprints, terminal dispositions, and idempotent batch keys prevent stale or duplicate learning
Matching REST endpoints and MCP tools expose the workflow to scheduled operators

Added an additive authoritative-grounding contract for tool connectors:

Explicit grounded, insufficient, and abstained states
Typed references with authority, revision, relevance, freshness, limitations, and conflicts
Corpus and index revision metadata
Safe replay metadata and changed-request fingerprint conflicts
Accepted, corrected, and rejected feedback outcomes
Used, ignored, and unreported memory usage counts
Approved supporting evidence that may supplement, but never replace, authoritative evidence
Request-specific conversational abstentions generated by AI instead of stock fallback answers

The contract is organization-neutral. Source authority comes from each organization’s approved source policy.

Initial developer docs launch.

Added: