Research Notes
Long-form updates on FDRP development and methodology
Posts
FDRP-on-FDRP — applying the refinement process to its own narrative
Run #1033 shipped 21 work items across fdrp.liviu.ai in roughly three
hours. Two cumulative coherence audits caught 16 incoherences before any
MEDIUM-tier page was allowed to promote; remediations passed 7/7 and 9/9. Six new
dashboards and a paper addendum landed under MEDIUM-tier PEAR. The cross-model gate
made a calibrated decline-of-DEFER when GPT-5.5 quota hit twice, and Constitution P9
(diverse expert-framed scrutiny) was operationally proven, not aspirational.
Wave R14 — Continuous Web-Property Update + Sugiyama-Layered Async Sync
Every page across every FDRP subdomain enters a typed inventory with a Sugiyama layer
(1 canonical, 2 derived, 3 referencing, 4 outbound). Pages stale >7 days auto-enqueue;
updates cascade through dependency edges; asynchronous streams synchronise at typed
barriers. Substrate ships at commit 27d5b317: 4 tables, 3 views, 3 daemons,
48 pages indexed across 18 vhosts, 34 stale at audit time.
2026-04-23: Reusability substrate, dynamic priority, and the defence-in-depth scrubber
Three scope expansions (VII/VIII/IV) shipped in a single evolution cycle: R8 reusable
artifacts + cache, R9 WSJF-ordered roadmap with 40 live items, and the A3 prompt-scrubber
MVP with 29 rules across 4 tenant tiers. Plus a mesh-wide Haiku purge rooted in
marketplace scaffolding templates. 11 new tables, 9 new views, zero direct cross-model
calls remaining in bin/.
R1028: FDRP Methods Survive the Opus 4.7 Migration
Opus 4.7 shipped with literal instruction following, a new tokenizer (+0–35% inflation), and three breaking API changes. Run #1028 tested six HIGH-priority FDRP methods head-to-head against 4.6. Four produced clean paired output, three improved on 4.7 (Bridges, Six Thinking Hats, Triage Cascade), Fractal-Out S0–S6 was equivalent, and zero methods retired.
The 85.7% Retraction: Why Self-Correction Matters More Than Being Right
The original 85.7% convergence rate was wrong. Cross-model verification with three independent LLMs caught the arithmetic error. This is the story of how the system corrected itself, and why retractions are a feature, not a failure.
fdrp.liviu.ai: A Companion Website for Research Transparency
Most papers exist as static PDFs. FDRP is different. A living system that evolves with every production run needs a companion that can show the dashboard updating in real-time, let you explore vector figures at full resolution, and give you downloadable data to verify our claims.