AffordanceForge: A Reliability & Audit Compiler for Fixed-Model LLM Agents

Concept

AffordanceForge is a deployment-time SDK and runtime that turns a frozen, vendor-supplied LLM into a certifiably more reliable agent by engineering its surrounding 'world-side' interface. Following the paper's Cognitive Impedance Matching Theory, it provides: (1) typed action handles that constrain what the model can do, (2) validators and repair paths that catch and contract errors, (3) rollback modes and authority scopes that bound blast radius, (4) committed observable ledgers and artifact I/O manifests that record every step, and (5) signed, auditable receipts that back explicit 'claim objects' about what the system reliably accomplishes. Crucially, all evaluators — human reviewers, LLM judges, benchmarks, external auditors — are treated as named, fallible measurement channels, so compliance teams get conservative, finite-sample certificates rather than vendor marketing claims. The product ships with reference templates for two high-value agent classes the paper works through: code-editing agents and retrieval-augmented generation (RAG) systems.

Why now

The abstract demonstrates a concrete, formal foundation — observable-only, no-meta certification with deterministic reducers, validity-budget ledgers, conformance envelopes, and repair-contraction guarantees — for amplifying a fixed model's operational capability purely through world-side design. It explicitly provides worked examples for code-editing agents and RAG, the two most commercially deployed agent categories, and it reframes capability gains as a 'compilation problem' that can be built as software rather than requiring access to model weights. As enterprises adopt third-party frozen models under emerging AI governance regimes (EU AI Act, audit/assurance demands), a tool that delivers auditable receipts and conservative reliability certificates around models they cannot modify directly addresses an unmet, regulation-driven need.