Explorer A: originate questions from plausible, checkable learning opportunities

Finding and actual prior

Compression progress supplies a direction for search, not a ready-made qualitative question generator. The strongest transferable candidate is to look for a presently unresolved relationship, propose a way of exposing its consequences, and prefer inquiries where either outcome would change an identifiable account. Opposing predictions and internally checkable consequences make that direction less vague. They do not establish that it beats ordinary problem-led inquiry.

The motivating observation is documentary, not experimental: the second brief records that the first trial already included surprise-led inquiry but left question origination underdeveloped. Compression curiosity was parked without a recorded comparative rationale. The actual prior is therefore not “curiosity was absent” or “retention-first failed.” It is that a useful, bounded provenance comparison emerged without settling how to originate promising investigations. This reading adds candidate machinery; it does not demonstrate a better allocation.

What the priority source supplies

The following source findings use the newly retained Quotes in Formal Theory of Creativity. Section labels identify the supporting extracts, not unretained ingest analysis.

1. Progress approximations are materially different mechanisms. Section III-B proposes learning a mapping from actions to expected future improvement. One implementation trains a separate network to predict changes in the model's output and uses its absolute output as intrinsic reward. The implementations monitor recent effects of learning instead of repeatedly evaluating the full history. Section III-C instead uses divergence between the predictor's probability distributions before and after observation. Section II-B's loss-free description-length measure counts both predictor and deviations from predictions.

These are not interchangeable permissions to score any visible revision as progress. My inference: changing a qualitative explanation, changing a probability distribution, and shortening a reconstructable encoding have different failure modes. In particular, a document revision count could reward oscillation. A persuasive new explanation could increase confidence without improving consequences. The source offers specific approximation choices; it does not validate these prose substitutes.

Section V also warns that high prediction error may persist because of noise or limitations of the learner, even in deterministic worlds. This matters for steering: lack of progress does not identify the encounter as meaningless. It can identify an unsuitable representation or learning method. Parking should preserve that alternative rather than label the evidence “noise.”

2. Opposing predictions supply an operational route beyond passive scoring. In III-D, modules propose executable algorithmic experiments and predictions. The opponent may reject the bet or accept it with a contrary prediction. Execution decides the accepted bet; reward transfers from loser to winner. The opponent's current knowledge supplies a changing novelty baseline. Costs of actions affect the reward-per-time ratio.

The important transfer is not adversarial personalities or mandatory zero-sum rewards. It is the separation between proposing a question, admitting a genuine disagreement, and obtaining an outcome that settles that disagreement. Two agents writing conflicting essays would instantiate none of the last requirement. Nor does disagreement alone guarantee significance: rivals can disagree about a cheap but irrelevant detail, or share the same mistaken premise.

3. Internal outcomes are legitimate objects of inquiry. III-D treats temporary computational results as part of the environment. V(1A-d) explicitly includes internal transformations and mathematical patterns without fresh external input. Thus restricting the proposal to predictions of future external events would misread the source.

Our translation is narrower: a qualitative KB can investigate implications, consistency, counterexamples, or whether a proposed distinction survives a change of assumptions. These are possible internal checks, not empirical confirmation of the world described. An argument that establishes “this conclusion does not follow” need not predict an error frequency. Conversely, an LLM's agreement with its own explanation is not the equivalent of checking an algorithmic outcome.

4. Recent-history evaluation has a named blind spot. V(2A) says performance decline on earlier history can go unnoticed when evaluation uses recent data. This is stronger than a generic admonition to keep records: the evaluation set can make local progress coexist with unobserved deterioration. Retaining old cases without ever consulting them does not close that gap.

5. Question origination remains partly delegated to search. II-A explicitly separates the goal of finding data modeled better or faster from the means, leaving action selection to reinforcement learning. III-B's action-to-improvement mapping could learn where opportunities recur; III-D gives programs a way to propose and settle their own tasks. The source therefore supplies both an objective and example generative machinery. It does not, in the grounded passages, supply a feasible recipe for originating consequential prose questions across a KB. That bridge remains our research problem.

Comparison with the earlier source

The retained Quotes in Driven by Compression Progress already establish three constraints: achieved compressibility is not improvement; persistent white noise and fully assimilated regularity both offer no further progress; old and new compressors must be compared on the same data. They also frame action selection as seeking expected future progress within the compressor's limitations.

The priority reading makes our present design space more concrete through approximation mechanisms, opposing bets, internal outcomes, and the explicit older-history warning. This is additional specification for this inquiry, not independent empirical corroboration. Neither source's cited passages warrant treating fewer files, fewer words, or a more uniform vocabulary as improved understanding. Literal loss-free coding and explanatory economy must stay distinct.

Candidate positive steering for qualitative KB work

The following is a conjectured research practice, not an adopted controller or executed trial. Its purpose is to originate investigations that could improve shared understanding without requiring immediate service payoff.

Start from a relation that could be learned, not merely a topic that is unfamiliar. Candidate origins include a recurring exception to a shared explanation, two successes that may rely on the same unnoticed condition, or two claims that appear inconsistent under one reading. State the actual incumbent account and what it leaves unexplained. If the incumbent already predicts the observation, its vividness is not evidence of surprise.

Generate contrasting consequences before making an improvement promise. For each promising relation, ask what would differ if a shared explanation held versus the strongest plausible rival. The answer can concern an observable event, a documentary difference, or an implication under stated assumptions. It need not be numerical. If both accounts permit every relevant outcome, the next worthwhile question may be how to acquire a discriminating case rather than how to rewrite the account.

A documentary example available in this run is the first-trial narrowing itself. One conjecture is that explicit learning opportunities were insufficiently articulated, so inquiry gravitated toward an easy-to-specify provenance comparison. The rival is that the comparison was a sensible bounded investment and question origination remained open by design. The brief's record does not settle that rivalry. Asking “what reason was retained for this narrowing, and what opportunity did it leave unexamined?” is better grounded than asking why retention defeated curiosity. Missing reasons are missing evidence, not evidence for the first conjecture.

Select by a concrete possible change, with a bounded route to finding out. Prefer an inquiry when a reachable outcome could remove an unsupported inference, discriminate rivals, identify a common necessary condition, or reveal that apparent unity conceals different mechanisms. Breadth across cases is a reason to investigate, not proof of a unification. Explain why this route is more promising than the nearest live alternative, without inventing a universal score.

Let results change future selection. A successful inquiry should identify what kind of encounter yielded learning and why; an unsuccessful one should distinguish already-known structure, inaccessible evidence, learner limitations, and genuinely unresolved rivalry. Otherwise the system merely writes progress narratives without learning where to look next. This is the qualitative counterpart of the action-to-improvement mapping, not an implemented reinforcement learner.

Rival, discriminant, and retention

The strongest alternative is ordinary problem-led inquiry with explicit rivals and consequence checks, without compression progress as a separate steering principle. It can generate all the practices above. Compression language earns additional value only if attention to shared learnable structure originates worthwhile questions that this rival would otherwise miss—not merely redescribes the same questions afterward.

A later authorized comparison could inspect whether the two framings nominate different investigations from the same evidence and whether those differences survive independent scrutiny. Distinct nominations alone would not establish benefit; the consequential distinction is a recoverable improvement in understanding at an acceptable cost. No comparison was run here. Resource measurements for this worker are unavailable to me.

Preserve the incumbent explanation, the encounter that motivated inquiry, rival consequences, rejected interpretations, and the evidence used for before/after assessment. Preserve disconfirming and older boundary cases as well as recent successes. Record evidence never acquired separately from evidence discarded or retained but not consulted. A prose exception summary is not presumed reconstructable. Acquisition remains open whenever retained material cannot decide the question.

Disposition and remaining question

Advance opportunity-led question origination, especially internal consequence checks and unexplained regularities, to the next synthesis. Keep the non-compression rival live. Park a numerical progress metric because the transfer lacks a defensible coding or scoring interpretation; park a generalized retention architecture because this inquiry establishes requirements, not ownership of that design. No new source destination or literature-search completion is claimed.

Remaining question: can we exhibit one actual KB relationship for which the learning-opportunity framing originates a consequential question that ordinary problem-led inquiry would not already select, while preserving the cases that resist its proposed shared explanation?

Grounding record: cp-skill-ground returned quotes added for kb/sources/schmidhuber-formal-theory-creativity.ingest.md; the eleven appended excerpts are retained verbatim in its Quotes section. The name-paired snapshot's checksum and source identity matched. Only Quotes changed; source validation passed with zero failures and warnings. The earlier source returned quotes sufficient without mutation. Source uses were checked against semantic/grounding-alignment; transfers above are explicitly conjectural. No experiments, code, commits, extraction, deletion, nested workers, or reading of current Explorer B occurred.