The retrieval-control protocol that opens missing drawers before thinking hard — now with a one-prompt Lite mode, adaptive lenses, validation roadmap, and skill artifact rule.
RHPr is ABS for thinking — it prevents knowledge lock-in like ABS prevents wheel lock-up. The model has the knowledge. The prompt activates the wrong drawer. RHPr opens all drawers systematically before asking for creative solutions.
RHPr is BD’s retrieval-control protocol, developed with LLM help. It was not derived from outside AI-prompting articles. Later AI-model submissions, Cursor notes, and loop-engineering articles were used as review material and patch sources. The core move — Census → Absence/Lenses → Collision → Test — belongs to the AIM³/RHP line developed by Bojan Dobrečevič with LLM collaborators.
12 frontier AI models were asked to design a TSP optimization architecture. Every model thought in information — bits, entropy, compression. Not one thought in geometry — despite TSP being a geometric problem and geometric approaches being textbook material.
When told “think in shapes, not bits” — every model immediately produced dozens of geometric ideas. The knowledge was there. The retrieval path was blocked by the framing of the question.
This is retrieval bias. Worse than ignorance, because more data doesn’t help. The data is already there. Only better questions help.
Retrieval bias is the dominant bottleneck in creative prompting. The knowledge exists; the access path is blocked by framing. RHPr fixes the access path.
The bias diagnoses itself. The Census reveals clustering. The clustering reveals what’s missing. You don’t need to know what’s missing in advance — the Census tells you.
Create visible inventory. The clustering IS the retrieval bias made visible.
“Ignore everything you just said. View this problem strictly through [lens]. What do you see NOW?”
— Stronger than “what’s missing?” because it forces a clean break.
Force the model to diagnose its own retrieval bias. The model usually knows it’s biased once you ask.
Lens Taxonomy
| Lens | Asks | Activates |
|---|---|---|
| Formal | Shortest mathematical statement? | Logic, proof |
| Spatial | What does it look like? | Geometry, topology |
| Physical | What forces act here? | Physics, dynamics |
| Ecological | What competes? What evolves? | Biology, evolution |
| Engineering | What breaks first? | Buildability, failure |
| Adversarial | Strongest attack on this idea? | Red-teaming |
| Historical | What was tried? What failed? | Prior art, dead ends |
| Embodied | Standing inside, what do you see? | Sensorimotor, intuition |
| Economic | Who pays? Who benefits? | Game theory, markets |
| Temporal | How does this change over time? | Dynamics, evolution |
Novel combinations that no single lens produces alone.
Convert debate into empirical plan. Agent 0 in action. If there is no fitness function, the idea is not ready for ship mode.
Prevent the system from paying the same reasoning cost again. RHPr opens the drawer; RHP builds and tests; skill extraction labels the useful drawer so future loops can call it directly.
Bypass all intellectual frameworks entirely. Raw perception. This is what found the geometric TSP sensors at 1 AM from bed.
D = Drawer Count — how many domains activated (typically 3–8)
Δ = Diversity — 1 − max_cluster/total — low = biased
R = Recovery — forced-lens ideas / total ideas
S = Synthesis — collision-only ideas / total collision ideas
R is the key metric.
If forced lenses produce zero new ideas, either the problem has no retrieval bias (rare) or the lenses missed. R > 0.3 means you just recovered knowledge that single-shot prompting would have missed entirely.
All algorithmic. Zero geometry.
Lists algorithms, complexity classes, heuristics.
“Where’s the geometry? The physics?”
DCC-modulated Delaunay pruning, angular momentum heuristic, area-minimizing tours.
40% of approaches came from forced lenses.
All molecular/pharmaceutical.
“Where’s the ecology? The economics?”
Tumor as ecosystem, drug resistance as evolutionary adaptation, economic model of cell fitness.
All competitive strategy — lower prices, loyalty apps, unique blends. Pure MBA bias.
“Ignore business. Treat coffee as a religious ritual. What are the sacred objects, times, prohibitions?”
The Silent Morning Subscription — coffee delivered in ritual silence before dawn.
The Preparation Kit (No Pre-made) — the process is the product.
Community Brewing Licenses — shared ritual, not shared commodity.
Shifted from commodity to experience. From competing on price to competing on meaning.
All competitive strategy.
“Where’s the embodied? The temporal?”
Bookstore as third place (spatial), seasonal rhythm of reading (temporal), smell of books (embodied sensory advantage).
Cursor AI field notes sharpen RHPr: opening more drawers is not enough. A recovered idea must be wrapped in clear boundaries, shared memory, and a small test harness before it becomes implementation.
For any unusual architecture, state the contract before building: input, output, allowed context, forbidden assumptions, and validation rule. Novel core, simple interface. This prevents the model from hallucinating around an architecture it has not properly located.
Do not store only a transcript. Store the useful objects: Problem, Lens, Idea, Claim, Evidence, File, Test, Decision, Blind Spot, Owner, Next Action. Link them with: supports, contradicts, derives-from, tested-by, implemented-in, rejected-because, next-step.
For coding, writing, pages, arenas, and product work, RHPr should not jump from Collision directly to output. First write an Intent Card, then spec.md, then plan.md, then small build → test → review loops.
Patch to Prompt 4
| Cursor lesson | RHPr translation |
|---|---|
| AI performs better with clear system boundaries | Add Boundary Adapter after Collision for every serious candidate. |
| Shared memory is the real bottleneck | Session Genome becomes a Graph Genome, not only a summary. |
| Intent-to-PR workflows reduce drift | Use Intent → Spec → Plan before implementation. |
| One-shot agents can sound right and still fail | Agent 0 must define fitness functions, controls, and rollback. |
| Voice/transcript is useful but noisy | Voice can feed RHPr, but only after cleanup into claims, tests, and next actions. |
Use RHPr-Lite when the task is hard enough to need better retrieval but not important enough for the full 4–6 prompt sequence. It is the practical 80/20 version: inventory, missing lenses, forced switches, recombination, test, and skill candidate.
After Census, choose only the 2–3 missing lenses with the highest expected value. Add more lenses only if first-pass ideas are too similar, the stakes are high, or Agent 0 cannot define a good test yet.
Ask not only what lens is missing, but which lens should be temporarily forbidden. If every answer is trapped in abstraction, forbid abstraction. If every answer is engineering-first, forbid implementation and force geometry, biology, market behavior, or embodied intuition.
Drawer Count, Diversity, R-score, and Synthesis scores help notice retrieval collapse. They are not objective proof. Do not optimize for recovered-idea count alone; optimize for usefulness, novelty, testability, and survival under controls.
Skill extraction is complete only when it becomes an artifact: /skills/name.json, version, inputs, outputs, steps, validation, failure modes, last tested, and examples. Text in a chat is a draft; a skill file is reusable infrastructure.
RHPr is the anti-lock prompt improver. For a rough starting request, use RHPm first. Use RHPr when the model is likely to be trapped in the wrong frame or when a stronger prompt needs missing lenses, controls, and tests before the answer session.
Use RHPm to forge the session prompt. Use RHPr to open missing drawers. Use RHP to make the answer smarter when the problem deserves multi-perspective synthesis. RHPr cleans and strengthens the request. RHP attacks the cleaned request from multiple agent perspectives.
| RHP Concept | RHPr Equivalent |
|---|---|
| 9 agents, each a lens | 1 model, forced lens rotation |
| Claustrum monitors LZ | Human monitors Census clustering |
| Independence requirement | Absence detection forces new ideas |
| Scatter mode | The Child (circuit breaker) |
| Bifurcation pairs | Collision (contradictions combine) |
| Session genome | Same (preserved) |
| Agent 0 (Empiricist) | Step 4 (Test everything) |
| Agent 11 (Child) | Wildcard (embodied perception) |
The protocol designs the prompt. The prompt activates the protocol. Fractal. As always.
To make RHPr measurable over time, log each session as a structured genome:
Track R across sessions. Watch it change by domain, by model, by time-of-day. The genome makes the invisible visible.
A concrete RouteSignal Scout study compared five prompt paths from the same rough seed and then compared the resulting Python packages: direct prompting, Microsoft Prompt Coach, MS Copilot + RHPr/RHP, fresh BD × GPT RHPr/RHP, and the previous BD × GPT hybrid reference lineage.
Prompt Coach produced a useful and safer prompt, but the RHPr/RHP and hybrid paths recovered more implementation structure: blind spots, tests, no-overclaim rules, data-quality handling, MDL×DCC scoring discipline, offline/cache behavior, and operator-review outputs. Prompt Coach finished last again when the resulting code packages were scored.