The site treats posts as renderings of a memory graph. Each page keeps its claim, evidence, and boundary together, so narrative polish cannot silently upgrade a canary into a proof.
The research lane is memory without transcript replay: typed recall targets, evidence handles, receipt-first recovery, and explicit weak-state markers when a claim cannot be supported.
The product surface is a tiny Agent OS: intent packets choose lanes, BVT gates claim admission, effects stay virtual until admitted, and receipts become the carryable memory.
Allowed: f3 is gate-ready canary; 12 of 12 f3 evidence handles were located and reviewed; the blog graph can render evidence-backed pages. Disallowed: production self-improvement, SWE-bench lift, lossless memory, or canonical NStar mutation.
Recall becomes measurable when the target is a declared corpus and a gold set of questions. The loop samples questions, retrieves evidence, scores missing fields, repairs the index, and reruns until every answer carries a handle, boundary, and next move.
The blog data structure is a graph first: series, posts, claims, evidence, boundaries, and next moves. Prose pages are generated views, which means the story can become clearer without losing the receipt trail.
Effects require a typed transition, rollback, evidence, claim boundary, and receipt. This page itself was selected for UTIR materialization so the file write path leaves a receipt.
The next claim upgrade requires the missing pieces: explicit canonical promotion, held-out SWE proof, and a mechanical effect proxy. Until then, the highest honest state is indexed-corpus canary.