entropy

name

entropy

description

Broaden a solution search without sacrificing quality by running mechanically blind, semantically orthogonal model lanes and converging through a frozen evidence-based evaluator. Use for naming, strategy, architecture, research hypotheses, creative direction, or other consequential multimodal problems when ordinary LLM answers feel correlated, conventional, prematurely convergent, or trapped in one framing; also use when the user asks to introduce entropy, explore a wider solution space, seek genuinely different perspectives, or combine Sol, Fable, Opus, or other available model families.

Entropy

Treat entropy as a way to decorrelate hypotheses, not as a creativity score. Create different search paths first; converge only after testing them against the same frozen quality bar.

Read references/protocol.md before running the full protocol. Use its templates and evidence rules.

Choose the run size

Use a normal single pass when the task has one clear answer, a hard deterministic procedure, no credible evaluator, or too little consequence to justify forks.

Otherwise choose the smallest useful run:

• Use a mini run for a reversible decision: three blind lanes, 3–5 candidates per lane, and one evidence pass.
• Use a full run for a contested or costly-to-reverse decision: three to five blind lanes, 8–15 candidates per lane, a conventional control, and 6–10 evidence-pass survivors.
• Do not add lanes merely because model capacity is available. Stop when the next lane is unlikely to open a new semantic region or change the decision.

Run the protocol

1. Freeze the evaluator

Write these items before generating any candidate:

1. One-sentence observable outcome.
2. Knockout constraints.
3. Three to six weighted criteria totaling 100.
4. Exact reality checks and interpretation rules.
5. One set of shared evidence and sources plus permitted read-only tools and data routes.
6. Exact collision, deduplication, shortlist, tie-break, and lane-diversity rules.
7. Search budget, survivor count, and stop condition.

Do not edit the evaluator after seeing candidates. If reality reveals a missing constraint, start a clearly labelled new run with the revised evaluator.

2. Partition the solution space

Locate the semantic commitment points where plausible solutions genuinely fork. Assign each lane one positive commitment such as a source metaphor, causal model, evidence set, morphology, register, or risk posture. Give every lane the same outcome, constraints, and quality floor.

Always include the conventional center as a lane or control. Do not equate entropy with eccentricity. Avoid briefs such as “be creative,” “be unusual,” or “avoid the obvious”; they produce novelty theatre rather than useful coverage.

3. Execute mechanically blind lanes

Launch each lane in a fresh context. Pass only the frozen spec and that lane's brief, including the identical shared evidence and sources and read-only tool scope. Do not expose candidates, rankings, summaries, or reasoning from another lane before merge time. A persona instruction inside one shared context is not blindness.

Before dispatch, confirm that every provider and tool is an approved route for the shared inputs. Remove secrets and unnecessary private identifiers, and send only the minimum evidence needed for the lane. When cross-provider routing is not allowed, use fresh isolated lanes within an approved model family and disclose the reduced family diversity.

Use true model-family diversity when it is actually callable:

• Keep GPT-5.6 Sol or the primary frontier reasoning model accountable for the frozen spec, lane design, evidence, and final integration.
• Use Fable as an independent conceptual or adversarial lane by following the available fable-advisor skill. Require and preserve its verified model, effort, persistence, and terminal verdict metadata.
• Use an exact Opus route as a separate model-family lane when available. Verify the returned model identity, run it in a fresh nonpersistent context, and label the result as Opus rather than Fable.
• Use other actual families or fresh same-family contexts when they add a distinct lane. Disclose the substitution.

Vary reasoning effort only to suit the work: reserve high/max effort for deep, adversarial, or integrative lanes and use a lighter pass for the conventional control when appropriate. Never report effort variation as model independence, and never simulate an unavailable family by asking another model to role-play it.

Require each lane to return candidates, a short approach descriptor, risks, and falsifiers. Record the actual model and effort from orchestrator, launcher, or verified provider metadata; do not trust a lane's prose self-identification.

4. Merge before judging

Apply knockouts first. Deduplicate with the operational task-specific rule that was frozen before generation, not by surface wording or vague similarity. For naming, a valid frozen rule can treat candidates sharing a root morpheme or governing metaphor as one family. Resolve collisions with the frozen rule and record cross-lane collisions as observed attractors rather than independent votes.

Assign anonymous candidate IDs before shortlist evaluation. Score the remaining candidates without lane origin, then freeze those scores. Only then may an allocator reveal lane labels and apply the predeclared lane-diversity quota. Refill a lane once when more than half of it collapses or fails knockouts; apply the same frozen rules to the refill.

5. Test against reality and integrate

Evaluate the survivors without showing their lane of origin. Run the frozen checks before scoring soft criteria. Mark missing evidence as unverified; do not substitute model confidence for a lookup, test, proof, or source.

Send any recombined candidate back through every knockout and reality check. Choose one accountable winner from the evidence, not by majority vote. State the runner-up and the evidence that would reverse the choice.

Report the proof

Include:

• the frozen evaluator, shared inputs, merge rules, and lane briefs;
• the actual models and efforts used;
• candidate counts, knockout counts, and cross-lane collisions;
• the conventional control result;
• verified evidence for finalists;
• the winner, runner-up, rejected near-misses, uncertainty, and stop reason.

Describe lanes as engineered-diverse, not statistically independent. For domain or trademark work, keep all checks read-only, distinguish registration evidence from registrar purchasability, timestamp the evidence, and state that screening is not legal clearance. Never register, buy, message, publish, or take another consequential external action without the authority required by the host.