One command. The smallest sufficient machinery.
You should not have to choose between a “creative protocol” and a “work protocol” before the problem is understood. The public surface is simply RHP, and the complete governing source is one standalone RHP.md. Internally it remains layered so reasoning, assurance and provider adaptation can be applied only where needed.
RHP reasoning kernel
Question Opening when representation is at risk, Dreamer/Child/Cartographer/Falsifier lenses, free Resonance, Bifurcation, late Reframing, Crystallize, Empiricist, Genome and skill extraction.
Work / assurance layer
Capability checks, bounded candidate lanes, isolation/exposure records, evidence discipline, budget, verifier separation, artifact freeze, delivery and provenance—only when the task needs them.
Provider adapter
The actual runtime facts: available tools, files, agents, isolation, web, code execution, costs and verification. No model label is treated as proof of capability.
Open → resonate → reframe → test → learn.
RHP is deliberately asymmetric. It expands early, but it does not stay vague forever. Ideas are allowed to remain different long enough to expose their real mechanisms; then tests, constraints and evidence decide what survives.
The first mistake is often answering the wrong version of the problem.
Q0 is not a compulsory round of clever questions. It fires only when there is representation risk: a meaningful chance that the problem is framed in the wrong coordinates, assumptions, abstraction or causal direction.
- the request is materially under-framed or ambiguous;
- another representation could change the search space;
- a cross-domain structural bridge may matter;
- a prior approach failed or collapsed into one attractor;
- a carried
next_session_seeddeserves reopening; - the user explicitly asks for
questions first.
- the task is high-stakes;
- the answer needs citations or strong verification;
- the build is large;
- the final artifact needs a separate verifier.
Assurance risk and representation risk are different axes.
A question only matters if it can change the trajectory.
RHP does not reward questions for sounding deep. A pure open question belongs in Q0; it does not replace the classic requirement to contribute a non-derivative idea during Resonance. A conjecture phrased as a question counts as an idea only when the content already states a mechanism or representation.
“What are we assuming?”
Useful for opening the space. Not yet a mechanism.
“What if the solver teaches the pyramid?”
Grammatically a question; substantively already a proposed causal direction.
Freedom stays in the middle.
After Q0, RHP returns to ordinary free Resonance. There is no forced Question → Answer → Critique ceremony. Functions may build, challenge, bridge, contradict or ask again. The protocol protects difference without requiring theatrical disagreement.
Transfer mechanisms
In full RHP, one Seed Dreamer demonstrates naive questions and structural cross-domain transfer during the first three applicable Resonance rounds. After that it becomes ordinary; Dreamer Mode remains open to every function throughout.
See it physically
Fold it, move it, illuminate it, press it, listen to it. Break abstraction lock with embodied/geometric intuition.
Attack the load-bearing assumption
Steelman the opposite. Preserve irreconcilable assumptions as Bifurcation Pairs until evidence can discriminate.
Map voids and bridges
Track explored regions, missing territory, convergence and structural gaps—without becoming the judge.
Compress what survived
Turn an idea into its shortest useful formal statement. If it cannot compress, it is not yet understood.
Stop arguing. Test.
After Crystallize, test surviving candidates and reasonable combinations instead of selecting the winner by rhetoric.
Open early. Adopt late.
Q0 proposes questions; it does not rewrite the problem. The late Reframing Gate sees the original seed, the Resonance field, and the questions that remained unengaged. Only then can a lower-complexity or better-grounded restatement replace the working formulation.
Reason freely; harden execution proportionately.
RHP Work is not “the part that actually works”. Classic RHP already includes execution and testing. Work is the capability-aware assurance layer around that reasoning: freeze what matters, keep branch exposure honest, bound actions/cost, preserve raw candidates, and verify the delivered instance.
When Q0 fires in partitioned Work, the question is checkpointed before candidate work starts. Later traces refer to that frozen digest, not to a prettier restatement written after success.
The Cold Challenger remains Q0-free. The Test Designer preregisters before Q0 output becomes visible. They preserve alternative framing and test independence.
When real isolation is unavailable, RHP does not pretend. Shared-context runs are labelled as shared and are not pooled with isolated runs as if they were the same treatment.
Results should teach the system how to ask better next questions.
The strongest version of the Question Laboratory is not a generator of more questions. It is a learning loop that retains only evidence-bound influence.
| Observed effect | Learning signal |
|---|---|
| A frozen question precedes a visible representation change | Eligible |
| It changes a decision, branch transition or test definition | Eligible |
| It can be linked to an artifact diff or result path | Eligible |
| The system later says “this was the important question” | UNKNOWN without trace |
Question-selection learning happens across sessions, not as self-confirming adaptation inside the same run.
The front door stays short.
Natural language remains valid. For a deliberate invocation, this is enough:
Use RHP. Outcome: [what I should receive] Sources: [files, links, connected sources, or none] Constraints: [what must be preserved or must not happen] Done when: [observable checks]
RHP chooses the smallest sufficient path. If you want the question field itself as the output, use questions first. If the request is rough and you mainly need a strong prompt for another session, use RHPm.
Different tools for different bottlenecks.
RHP 3.0 is self-contained: rough-request framing and Retrieval Escape are already inside RHP.md. RHPm and RHPr remain useful focused companion tools, not runtime dependencies of RHP.
RHPm · Prompt Builder
Optional prompt-forging companion: turn a rough request into a strong, structured session prompt when you specifically want a reusable prompt artifact.
RHPr
Optional focused anti-lock companion. Its core retrieval-escape logic is already integrated into standalone RHP 3.0.
8Z / AI8 Reasoning
Seed → bridge → build → alternative explanation → kill test → result → transferable operator.
AIM³
The wider human–AI continuity, coordination and research environment around these reasoning protocols.
A strong architecture still has to earn its place.
The Question Opening extension is a converged design candidate, not a field-validated superiority claim. The first controlled test deliberately separates “questions help” from “an extra protected opening phase helps”.
Current RHP
No Q0. Same task, model/runtime and total budget.
RHP + Q0
Triggered Question Opening when representation risk is frozen before the run.
Placebo divergence
Same protected opening budget, but an idea instead of a question. Classic: idea round 0. Partitioned: R1i → candidate → R1c.
| Pattern | What it would mean |
|---|---|
| B > C and B is non-inferior to A on deterministic correctness/cost | Evidence specific to the question form |
| B ≈ C > A | Protected divergence / delayed commitment helps, not questions specifically |
| B ≈ C ≈ A | No detected Q0 benefit |
| B materially < A/C | Q0 harms that task class or trigger |
For Q0, B uses R1q → candidate → R1c; C uses R1i → candidate → R1c. The opening budget is matched so the intended difference is question versus idea.
Initial minimum: at least 3 fresh-session runs per arm per task, with no cross-run visibility and equal total budget.
No small pilot can establish universal superiority. The candidate advances only if the benefit survives cost, controls and multiple task families.
Open enough to discover. Bound enough to test.
The Q0 architecture was challenged across GPT and Claude branches; after restoring Seed Dreamer and the exact placebo-control shape, the standalone 3.0-candidate R2 consolidation is accepted as CONVERGED_CANDIDATE.
The public page is a research candidate. Canonical promotion requires BD review and controlled evidence; the page does not self-authorize a protocol rewrite.
Technical note: the Q0 / Work freeze boundary
R1q is a pre-candidate question checkpoint inside the existing Work Divergence Prelude. The existing normative R1_FREEZE remains the candidate/result freeze R1c. A future Work revision can carry the original R1q digest in the Branch Card's existing input digest/content-ID field so later question traces cannot quietly rewrite the original question.