8Z reasoning core — how RHPm thinks
Every generated prompt already carries an 8Z reasoning mode. Claude's useful addition was to make that thinking layer visible on the page itself, so a human reader or crawler can see the method without running JavaScript.
This is a workflow/system layer around LLM use. It does not change model weights, is not standard transformer architecture, is not proof of AI consciousness, and is not autonomous AGI.
Limits need proof
Treat “impossible / can’t / not allowed” as claims requiring evidence. If evidence is weak, surface the hidden assumption instead of obeying the limit.
Decompose before dismissing
If the whole task looks blocked, split it into smaller subproblems and test the part that can still move.
Question the question
Look for stale assumptions, wrong drawers, missing files, hidden premises, or a better framing before answering too literally.
Never exclude before the arena
Keep candidate paths alive until tests, baselines, MDL comparison, or concrete constraints kill them.
Follow the surprise
If an odd result, user seed, or cross-domain bridge carries signal, preserve it long enough to test.
Transfer structure across domains
Look for reusable mechanism, not surface analogy. Move the generator, not just the metaphor.
Verify before claiming
Use tests, manual checks, visual inspection, citations, acceptance criteria, or clear uncertainty labels where possible.
Capture the learning
If a workflow repeats, extract a compact skill, prompt pattern, checklist, or next-session seed so the next run starts higher.
AIM³ ecosystem map
RHPm is the front door, not the whole system. It sits between the AIM³ collaboration OS, the heavier RHPr/RHP protocols, the 8Z / AI8 Reasoning method, and the next AIm³ MentalArena / mRHP Core Paper layer.
AIM³
AI Management of Mind & Memory: roles, continuity, decisions, trust rules, loops, and reusable skills.
RHPm
Turns a rough request into a portable, testable prompt for a fresh LLM session.
8Z / AI8 Reasoning
The seed → bridge → test → result invention method behind the prompt family.
AIm³ MentalArena / mRHP
Core Paper / architecture page: RHP improving RHP through protocol genes, benchmarks, lineage, and scoring.
Protocol Harvest + CVC review loop
RHPm now learns from useful outside prompts. When a review prompt contains a reusable reasoning move, RHPm can extract it as a pattern instead of merely comparing outputs.
For audits and score-style tasks, RHPm can now ask for a clear answer, claim decomposition, confidence tied to evidence quality, scope-split scoring, verification, weighted synthesis, caveats, and one targeted retry if confidence is low.
For reviews, research, synthesis, and protocol work, RHPm can extract reusable patterns: name, when to use, when not to use, prompt fragment, risk, and where it belongs in RHPm/RHPr/RHP.
This is strongest for review/audit/scoring, research, synthesis, public articles, and protocol improvement. It is kept optional so simple writing or coding prompts do not become over-protocolled.
Real-world prompt benchmark
Same rough request, different prompt substrate. Warm project context can rescue a weak prompt; RHPm makes the missing method explicit and portable; cold weak prompts drift generic.
RHPm did not add intelligence to the model. It transferred the working discipline into the prompt: read the target first, do not answer from memory, use sources, separate correct / simplified / misleading / wrong claims, give a calibrated percentage, and produce a reusable verdict.
The compact summary stays here. The full source transcript from How_LLMs_work.md is on a standalone dark HTML page.
Completed RHPm → multi‑LLM article case study
This is no longer only a plan. RHPm was tested by using a rough human seed to generate one shared execution prompt, sending that same prompt to 10+ top LLMs, scoring their answers, synthesizing a stronger hybrid article, sending the hybrid back for review rounds, and publishing the final dark HTML article.
- Round 1 — same prompt: the exact same RHPm execution prompt was sent to each LLM. No special lenses, so the comparison stayed apples-to-apples.
- Synthesis: the answers were scored, useful parts were extracted, generic filler was removed, technical claims were source-checked, and one hybrid article was built.
- Manual BD contribution: a clearly separated section links AIM³, RHP, RHPr/RHPm, AI8, self-selecting MDL, DCC, and 8Z Reasoning as BD’s workflow contribution to how LLM work can be improved in practice. It is not presented as a standard ML architecture claim.
- Review rounds: the hybrid article was returned to the same LLMs where possible, asking whether it beat their own answers, what was still weak, and what final patches were needed.
- Final publication: only useful review corrections were kept, provenance was preserved, English/Slovenian reading was added, and the result was published as a final article plus a scoreboard report.
This completed protocol is now part of the final “How LLMs Actually Work” article, so readers can see not only the result, but how the result was built and audited.
Full comparison: prompt quality, answer quality, and scores
The source file contains three answered paths plus one RHPm-generated prompt. So this scores four artefacts in the test chain, not four equal answers.
| Path | Prompt score | Answer score | Meaning |
|---|---|---|---|
| A — simple prompt inside warm AI8/GPT context | 40/100 standalone · 60/100 warm-context | 87/100 | Weak prompt, strong answer because the existing project context already carried BD’s audit style. |
| B — RHPm-generated audit prompt | 95/100 | 94/100 as prompt-builder output | Excellent portable prompt: it forces reading, sources, claim grouping, severity, percentage range, and reusable verdict. |
| C — fresh GPT session using the RHPm prompt | 95/100 | 92/100 | Best practical result. The method survived the move into a fresh session without relying on personal memory. |
| D — same simple prompt inside a cold GPT account | 35/100 | 52/100 | Useful generic intuition, but weak audit behavior: too broad, under-sourced, and it drifts away from the exact article. |
Short Slovenian request: “read this and tell me how true it is in percent.” It works only because the surrounding AI8 context already teaches the model how BD expects audits.
The generated prompt explicitly says: do not answer from memory; open/read the article first; cite reliable sources; classify claims; give a non-fake-precise score.
The fresh session returns a stronger structured audit: overall 84–88%, higher for transformer core, lower for modern end-to-end assistant systems.
The cold account gives a broad next-token discussion and then effectively admits it would need the article pasted for precise analysis.
The practical gain is not mystical. RHPm packages the missing working method so another LLM session can reproduce the quality without already knowing BD, AI8, or the collaboration style.
Quick links to major LLM assistants
Handy user links for testing RHPm prompts across different model families. This is a convenience list, not a live ranking.
Use the same RHPm-generated prompt in several assistants, then compare whether the output stays concrete, sourced, bounded, and testable.
Prompting LLM assistants via APICollapsed helper for API snippets and optional local DeepSeek testing. No API key is embedded in this HTML file; load a local key file or paste a key only for the current browser session.
Collapsed helper for API snippets and optional local DeepSeek testing. No API key is embedded in this HTML file; load a local key file or paste a key only for the current browser session.
For private use, this page can send a prompt directly to the DeepSeek API from your browser after you load a local key file or paste a key. The key is kept only in current page memory and is not written into the HTML file.
Use a local text/.env file such as DEEPSEEK_API_KEY=sk-... or a raw key on the first non-empty line. Do not commit that file to GitHub.
Use this when you want to run the same RHPm-generated prompt through API providers such as DeepSeek or OpenAI-compatible services. The live DeepSeek chat is first; optional curl/Python snippets are below it.
Live chat and snippets use the generated Forge prompt if available; otherwise they use the Master prompt preview.
Load your key from a local file or paste it for this session, write a message, and send it directly to DeepSeek. This is meant for your own private/local use.
curl / Python snippetsOptional terminal/Python helpers for debugging or automation. The live chat above is the fast manual path.
Official docs change, so treat these as starter snippets. Check provider docs, rate limits, model names, and billing before large runs.
DeepSeek API docs · OpenAI API key safety · OpenAI API reference · Anthropic API docs · Gemini API key docs