Skip to content

0020, Adaptive & self-improving Ariadne — bounded and audited, on a propose→ratify→freeze spine

Context

Ariadne generalizes at code level today: a new corpus needs a hand-written DatasetAdapter mapped into the canonical schema and registered in DATASETS (ADR-0006). To be a general sensemaking harness it needs to (a) adapt at runtime to a user's own store/ontology and (b) improve from experience across repeated workups — without eroding the auditable, read-only, no-silent-merge, provenance-by-hook governance that is the project's whole value to an intelligence-analysis stakeholder. The trigger was a direct ask: make the dynamic Ariadne "as recursively self-improving as possible."

Decision drivers

  • Governance is non-negotiable. Any adaptive/learned change must remain auditable and must not let the system silently alter what it is graded against.
  • Anthropic's RSI framing. Recursive self-improvement is the unbounded form (an AI autonomously designing its successor); the remaining — and explicitly human — bottleneck is judgment, verification, and direction. We build the bounded form that keeps those human.
  • 2026 self-improvement practice. The deployable pattern is skill libraries + procedural memory + reflexion + audited skill-graph improvement with verifiable rewards — every serious source warns that deployed self-improvement invites reward-hacking and untraceable drift unless "design rules and tests govern what changes are allowed."
  • Build on what exists. Ariadne already has the canonical-schema seam, MCP tool families, integration ports, and — crucially — a verifiable reward (the eval harness) and an audit trail (provenance + governance audit).

Considered options

  1. Fully autonomous adaptation (LLM rewrites mappings/tools/code at runtime). Rejected. This is the unbounded RSI form. It maximises "magic" but destroys auditability and invites reward-hacking (an agent that can edit its grader will game it). Disqualified by the governance spine.
  2. Stay code-only (status quo: hand-written adapters). Rejected as the end state. Safe and auditable but not a general harness; every new corpus needs the maintainer. Keep it as the fallback, not the only path.
  3. Bounded, audited adaptation on a propose→ratify→freeze spine. Chosen. The agent proposes declarative artifacts (schema mappings, ontologies, named skills); a human ratifies; the artifact is frozen as config the deterministic gates keep checking. Self-improvement edits only those declarative, ratified artifacts — never the gates, scorers, governance, or code.

Decision

Adopt option 3. Architecturally, two axes ride the one spine:

  • Axis A — Adaptivity: schema introspection (A1), a declarative user ontology/semantic layer (A2), dynamic MCP tool registration (A3).
  • Axis B — Self-improvement (bounded, audited): learned mappings as procedural memory (B1), learned analytic skills (B2), reflexion over the eval harness (B3).

The hard boundary (the safety architecture): the self-improvement loop edits only declarative, ratified artifacts. It never edits its own gates, eval scorers, governance rules, or code. The gates and the human ratification step are fixed points. This single rule is what makes the loop defensible.

Sequencing. First slice = A1 + A2-into-the-existing-canonical-schema + the B1 seed, on Postgres: introspect a real Postgres, the agent proposes a mapping into person/org/site/document + edges, a human ratifies, it freezes as mapping.toml, and the existing indexer/workup/eval run unchanged on the user's data. The full user ontology (A2), dynamic MCP (A3), learned skills (B2), and reflexion (B3) are later, separately-specced phases. Store target is Postgres first (richest standardized introspection, most mature agentic schema-linking research, reuses postgres-mcp restricted mode); ontology format is a lightweight declarative TOML, SHACL-validatable later.

Consequences

  • Ariadne moves from code-extensible toward runtime-adaptive and experience- improving, while every change stays auditable and human-ratified — the governance spine is preserved, not bypassed.
  • The eval harness and provenance audit are repurposed as the reward signal and audit trail for a self-improvement loop, a differentiated position most agents cannot claim.
  • The first slice is small and reversible (read-only introspection + a ratified config + the existing pipeline); the larger vision is phased behind it, not front-loaded (YAGNI).
  • A clear, statable safety boundary (no self-editing of gates/scorers/code) makes the capability presentable to an intelligence-analysis stakeholder.

Sources: Anthropic, Recursive self-improvement (anthropic.com/institute/recursive-self-improvement); audited skill-graph self-improvement with verifiable rewards (arXiv 2512.23760); procedural memory from experience (ProcMEM, arXiv 2602.01869); Voyager skill libraries; AutoLink agentic schema linking (arXiv 2511.17190); OntoKG intrinsic-relational routing (arXiv 2604.02618); Anchor schema-agnostic KG construction (arXiv 2606.01208); dynamic MCP (dynamic-fastmcp, Docker Dynamic MCP).