Planning for complex systems lacks the convergence metrics and defect detection that manufacturing achieved through six decades of statistical process control. FDRP addresses this gap by treating each planning decision as a manufactured artifact with measurable convergence and gate-reviewed quality, applying SPC, 5S, RAMS, Andon, and configuration freeze directly to planning. Applied across 80 production runs (1,968 decisions, 23 CONVERGED, 25 COMMISSIONED) with 32 active subsystems, the system's strongest empirical contribution is the antimatter building programme: 68 LLM-generated expert specialists across two rounds produced 28,936 lines of analysis and 1,656 structured findings, discovering expert persistence as an emergent operational pattern. Cross-model verification with 3 independent LLMs revealed 4 systematic bias patterns, suggesting that N≥3 models are needed for blind spot detection. In April 2026 the platform completed R1027 “FDRP-on-FDRP” —the architecture applied reflexively to its own evolution, converging with CVT 0.000 and RAMS 0.953. As of 21 May 2026 the architecture has grown to 1,200+ database tables, 380+ analytical views, 77 skills, 76 active agents, 446 expert personas, and 20 concept-to-X pipelines. The observability spine emitted 120,143 officer_events across 184 distinct event types in the trailing 7 days (11.9 events/min). Four foundational doctrines were memorialised on 2026-05-13 (omni-scope override, skills ecosystem evolution, cybersecurity + testing PPG, event-driven not schedule-driven). Wave R11 architecture panels converged round-1 (coordination-layer, skills-lifecycle, PPG, event-bus, all CVT≥0.86). Three runs commissioned in trailing week: #1037 solution-train coordination (CVT 0.870), #1036 SAFe-as-improvement-candidate audit (CVT 0.700), #1035 autonomous flywheel restoration (CVT 0.870). #1034 closure-obligation-backbone remains ACTIVE (CVT 0.403, iteration 2, supervises >300 live obligations).
FDRP
Manufacturing Quality Control for Complex Project Planning
AI-assisted planning at scale produces unmeasurable artifacts and undetectable defects. FDRP treats every planning decision as a manufactured part — statistical process control, convergence metrics, configuration freeze, peer-reviewed gates. Self-evolving, cross-model verified, fully traceable.
80 production runs · 1,200+ tables · 446 expert personas · 77 skills · 76 agents · 8.65/10 avg 5S
Choose your path through FDRP
FDRP serves four distinct audiences. Pick the door that fits.
Peer-review the method
Read the v22.1 paper, BibTeX citation, retraction history, and the FDRP-on-FDRP convergence proof (CVT 0.000, RAMS 0.953).
Open the paper page → Operator · ProcurementSee it in your domain
Eleven domain case studies: particle physics, cyber defence, hospital twins, hospitality, grants, fire response, CJMM, AirGuard, APEX-OS, ISCIR, SingleKey.
Browse case studies → Investor · Stakeholder 🔒 InternalWatch the system run live
Real-time operator dashboard: decisions pending, runs in flight, convergence velocity, 5S audits. Every number a live SELECT against MySQL (BIND-021).
Open live dashboard → Engineer · IntegratorRead the engineering log
Wave write-ups, retraction notices, R1027 self-convergence, R1028 Opus 4.7 method validation. Architecture decisions in the open with their commit hashes.
Read the blog →Key Numbers
Abstract
Historical Milestone — R1027: FDRP-on-FDRP (16 April 2026)
Current run: #1033 fdrp-on-fdrp website refresh (10 May 2026, in progress). R1027 and R1028 below are recorded historical milestones; for the live operational state see /dashboard/ and paper v21.0 addendum.
16 April 2026 — Run #1027 “fdrp-on-fdrp-v2” applied the Fractal Diamond Review Protocol reflexively to its own continued evolution and CONVERGED in a single day. The run closed with CVT 0.000 (zero velocity tension, every axis stable), convergence score 1.0000, and a RAMS verdict of PASS (overall 0.953 — Reliability 0.978, Availability 1.000, Maintainability 0.913, Safety 0.922). The pipeline was directed to execute “30 waves of megaexpert expansion, n-fractal views, n-fixes and fine-tuning”, fed its own corpus back into the expert panel, and self-converged to a stable fixed point. This is the first publicly documented case of the methodology auditing and improving itself under its own gates.
New Subsystems Landed in R1027
- data-quality-auditor — nightly sweep over FDR/council/correction tables, flags schema drift and orphaned rows.
- proactive learning collector — fixes the
SYS_SESSION_LEARNING_RATE=0blind spot; harvests decisions, retractions and pivots into the knowledge index. - drift-detector overflow fix — replaces
decimal(5,4)saturation with bounded soft-clip so extreme deltas no longer mask drift. - traceability formula scoping —
SYS_TRACEABILITYnow computed only over runs with concerns (27 runs backfilled); overall traceability rose from 0.32 to 0.92. - BIND-052 PIO router — every human-facing session now operates as a Principal Intelligence Officer (PARSE → CLASSIFY → DISPATCH → MONITOR → REPORT), with specialist work offloaded to subagents.
R1028 — FDRP Methods Validated Against Opus 4.7
20 April 2026 — Claude Opus 4.7
(GA 2026-04-16) introduced three breaking API changes
(temperature/top_p/top_k rejected,
thinking.budget_tokens removed, assistant prefill removed),
a new tokenizer with 1.0–1.35× token
inflation on the same text, and a behavioural shift toward
more literal instruction following. Run #1028 ran the
full FDRP method catalogue against 4.7 head-to-head versus 4.6 to
measure the impact.
Net verdict: NEUTRAL with selective IMPROVEMENT. Of 6 HIGH-priority methods tested head-to-head, 4 produced clean paired output, 3 improved on 4.7 (Six Thinking Hats, Bridges, Triage Cascade), 1 was structurally equivalent (Fractal-Out S0–S6), 1 broke on 4.7 only (DDn — completion timeout, not content regression), and 0 methods were retired. Average 4.7 output is ~12 % shorter but carries more domain-specific citations per word — denser, not dumber.
Liviu-Named Methods: Bridges & Fractal-Out
Two methods in the FDRP catalogue were named directly by Liviu—Bridges (cross-domain absorption) and Fractal-Out S0–S6 (scale stress testing). Both were explicitly re-tested under R1028 against Opus 4.7.
- Bridges (method #7) — SHARPER on 4.7. 4.7 pulls more specialist citations (Siddiqi Intelligent Credit Scoring, FAO-56 Penman-Monteith, Makridakis M5, Elkan IJCAI 2001) where 4.6 stayed at foundational references. Transferable operators are more concrete and directly executable (adversarial-weight CV, ET0 monotone constraint in LightGBM, isotonic-over-pinball at multiple τ). Length −13 % (398 vs 459 words) but higher density. Zero hallucinations on spot-check. Keep prompt as-is.
- Fractal-Out S0–S6 (method #8) — 4.7-SAFE. The literalism concern from Anthropic’s migration guide did not materialise: 4.7 addressed all seven scales (S0 solo → S6 civilisational) without prompting, same as 4.6. The two models disagreed on breakpoint (4.6: first failure at S3; 4.7: first failure at S4) — a legitimate architectural disagreement, not drift. Keep prompt as-is.
Method Health Under Opus 4.7
| Method | 4.6 status | 4.7 status | Notes (from R1028 δ report) |
|---|---|---|---|
| Bridges (cross-domain) | ● validated | ● sharper | Tighter domain specialisation, concrete operators, 13 % shorter |
| Fractal-Out S0–S6 | ● validated | ● validated | All 7 scales addressed; S3 vs S4 breakpoint disagreement |
| Six Thinking Hats | ● validated | ● sharper | Fewer validation-forward phrases, broader doctrine integration |
| Triage Cascade | ● validated | ● sharper | PSI, Mahalanobis, ECE, bp lift — tighter technical notation |
| DDn (Double Diamond recursive) | ● validated | ◐ needs rework | 4.7 timed out at 180 s; cap phase word budget or use task_budget beta header |
| Pre-Mortem | ◐ prompt-bug | ◐ prompt-bug | Both models timeout; chunk 10 failure modes into 2 batches of 5 |
Opus 4.7 breaking changes absorbed: sampling-parameter rejections, adaptive-thinking migration, prefill removal. Tokenizer inflation of 1.0–1.35× noted for cost budgeting. Full method-by-method evidence: see the R1028 differential report linked from the research notes.
What shipped 2026-04-25 … 28. Nine waves landed in a four-day window. Substrate work dominates: orchestration, agent-agnostic abstraction, hardware roadmap, optimisation kernels, async web sync, per-project isolation, client-funded backbones, and destructive-action confirmation. R16 is applied; its harness is gated on dispatch. R18 Phase 2 is gated on seven CRITICAL countermeasures.
- R10 — Multi-account orchestration v1. shipped ·
5d8d5603 - R11 — FDRP v-next agent-agnostic abstraction. Phase 1 ingestion shipped ·
f5d91b4a - R12 — Local-model build-out + hardware roadmap. design shipped ·
67ddaa57 - R13 — PCRC optimizer + Sugiyamaⁿ + RLMⁿ. substrate shipped ·
76b36c54 - R14 — Sugiyama-async web continuous update. substrate shipped ·
27d5b317 - R15 — PPCIL per-project context isolation. shipped ·
45e1bc93 - R16 — Testing-tier framework. substrate applied 2026-04-28; harness blocked on dispatch
- R17 — CFBE Client-Funded Backbone Extraction. shipped
- R18 — DAFC Destructive-Action Fractal Confirmation. Phase 1 shipped ·
08043737+ adversarial review6ada3e81; Phase 2 gated on 7 CRITICAL countermeasures
Latest evolution · 23 April 2026. Three scope expansions shipped in a single cycle: reusability substrate (Scope VII, R8 Gate 8a), WSJF-ordered roadmap (Scope VIII, R9 Gate 9a), and the defence-in-depth prompt-scrubber MVP (Scope IV, A3) with 29 rules across 4 tenant tiers. Plus a mesh-wide Haiku purge rooted in marketplace scaffolding templates. 11 new tables, 9 new views, 40 live roadmap items.
How It Works
Convergence Velocity Tensor (CVT)
The CVT measures planning quality as a diamond with four axes: decision stability, coverage completeness, clash resolution rate, and expert consensus. Each axis is independently measurable using SPC control charts, and the composite tensor tracks convergence across gates.
Six-Phase Gate Lifecycle
Each gate enforces monotonic convergence: decisions can only tighten, never regress. Violations trigger Andon stops and root cause analysis.
32 Self-Evolving Subsystems
The pipeline is organized into 32 subsystems across 7 categories: Core (fractal-diamond, concept2x, collective coordination, cognitive architecture), Discovery (expert expansion, knowledge grafting, thinking types, cognitive opportunities), Quality (peer review, 5S auditing, web finetuning), Evolution (SPC monitoring, self-evolution, opportunity hunter), Integration (ecosystem maps, Rosetta Stone, multinode project, wiring check), Governance (payload constitution, RAMS safety, CERN compliance, cost accounting, daemon taxonomy, limits hunt), and Visualization (visualization engine, Unreal 3D).
Subsystem Health
Live health status across all 32 subsystems. Data from production MySQL database.
Featured Case Studies
Antimatter Building Programme
68 experts, 6,052 findings, CHF 147M facility. Five review rounds with emergent expert persistence.
CybersecurityAsymmetric Cyber Defence
67:1 to 347:1 force ratio, 451 CVE detections. Proportional response with graduated escalation.
Emergency ManagementFire Crisis Response
8 data sources, 6 live dashboards, real-time monitoring. FIRMS, EFFIS, weather, and air quality integration.
System Architecture
The production system as of 21 May 2026, measured from the live MySQL database (BIND-021 compliant — every number has a SELECT).
Framework Comparison
| Framework | Convergence Metrics | Traceability | Clash Detection | Grounding | Self-Evolution | Cross-Model |
|---|---|---|---|---|---|---|
| FDRP | ● | ● | ● | ● | ● | ● |
| AutoGPT | ○ | ○ | ○ | ○ | ◐ | ○ |
| CrewAI | ○ | ◐ | ○ | ○ | ○ | ○ |
| LangGraph | ○ | ◐ | ○ | ○ | ○ | ○ |
| MAPE-K | ◐ | ◐ | ○ | ○ | ◐ | ○ |
Native support Partial Absent
Correction Notice
Correction Notice (2026-03-15): The originally reported 85.7% convergence rate was retracted to 75.0% after cross-model verification caught an arithmetic error. This self-correction demonstrates FDRP's intellectual immune system in practice. Details →
Cite This Work
@article{olos2026fdrp,
title = {FDRP: A Self-Evolving Architecture for Planning Quality},
author = {Olos, L.},
year = {2026},
note = {Preprint, v22.1 (supersedes v22.0, retracted 2026-05-13)},
url = {https://fdrp.liviu.ai/paper/FDRP-Paper-v22.1.pdf},
date = {2026-05-13}
}