Explorer B — useful execution is not yet reusable capability

Outcome and status

Increase useful, warranted work now, and separately test whether experience improves later work and later method development. Do not count executed methodologies, retained instructions, or automation coverage as substitutes for those outcomes. My main proposal is to make transfer a contested return on an investment, not an assumed by-product of use.

This is independent turn-1 exploration from the capability-development angle. I read the commission, parent framing, root instructions and work collection contract in full. I did not inspect any other turn-01 artifact. No delegation, external research, library changes or commits occurred. Findings below are arguments and bounded readings of existing evidence, not results from a new experiment.

The question merits investment because the proposed system can produce valuable deliverables while failing to acquire reusable capability. If those returns are conflated, successful use can indefinitely finance unsuccessful learning.

1. What should increase?

The declared objective in Commonplace as an instrument is more useful and better-warranted knowledge work per unit of human judgment, not automation. The new commission makes learning to automate methodologies part of the plan; it does not silently replace that objective.

I would distinguish four claims rather than combine them into a score:

1. Present service: more worthwhile work completed to the required standard with acceptable total cost and operator involvement.
2. Later service: earlier experience improves performance or reduces the cost of later work that was not simply the original task repeated with its answer supplied.
3. Adaptation: when conditions change, the system can discover that the old method no longer fits and recover without rebuilding every relevant judgment from scratch.
4. Method-development productivity: earlier experience makes the next useful methodology cheaper or easier to develop, not merely easier to execute once somebody supplies it.

These are proposed distinctions, not newly adopted metrics. Improving the first while degrading the others is possible. So is investing in the latter without an immediate service improvement. The operator must decide which tradeoffs are worth funding; no evidence here supplies the exchange rate.

For broad knowledge work, the most important expansion may be the range of changed situations in which the system can still reach a warranted result, including recognizing when it needs help. Counting task names or domains would conceal whether each new domain required another bespoke human-built method. The bootstrap compatibility argument already makes that distinction. It is an argument about the burden of a scaling claim, not independent evidence that Commonplace meets it.

A decrease can constitute progress: fewer rules, a narrower applicability claim, or less attempted automation can improve later work by preventing predictable misuse. Growth in a library is therefore neither necessary nor sufficient for growth in capability.

2. Execution exposes opportunities to learn; it does not perform the whole learning operation

The experience-responsive retention account supplies a useful borrowed decomposition: production experience must determine a reusable change, the change must survive, later production must depend on it, and the dependence must improve the declared outcome. I extend that argument here into five possible breaks:

Break	Why useful execution can coexist with failed learning	Consequence for this plan
No new information	The model already knew how to do the work; the method merely reminded it	Credit delivery, but do not infer acquired capability
Wrong explanation	Success came from a lucky tool result, hidden operator repair or a different part of the process	Preserve the episode without promoting its explanation to a general rule
No later uptake	A retained method is never selected, or its content does not change behavior	Retention is not a return until a later consumer benefits
Wrong transfer	A useful local shortcut is applied where its conditions fail	Scope recognition and abstention are part of capability
Net loss	Retrieval, checking, repairs and maintenance cost more than reconstruction	A causally effective memory can still be a bad investment

Two grounded source readings prevent this from becoming a purely skeptical argument:

• Positive but bundled evidence. Memento-Skills, in its retained experimental-setting and result passages, reports higher held-out GAIA and HLE accuracy for its full skill-optimization system than for a read-write ablation disabling failure attribution, rewriting and discovery. That supports the possibility of useful frozen-LLM adaptation through mutable skills. It does not isolate which of those operations caused the gain, establish cross-kind knowledge-work transfer, or demonstrate that every retained skill is useful. Its quoted claim that each write directly improves policy is stronger than rewriting alone warrants.
• A missing causal link. Large Language Model Agents Are Not Always Faithful Self-Evolvers reports controlled interventions across four frameworks, thirteen backbones and nine environments. Its retained passages report stronger behavioral dependence on raw experience than on condensed experience, including condensed-only conditions. This is paper-reported evidence that delivery of distilled lessons need not produce their intended behavioral dependence. It does not show that all summaries fail, or that raw traces are always economical or correct.

Together these do not vote for or against memory. They identify separate questions: can a learning arrangement help, and did the particular retained content cause the later behavior? Neither source demonstrates the commissioned broad bootstrap.

3. Transfer is not the same as copying a method to another task

My conjecture is that useful broad transfer will often preserve a relationship between a problem condition and a decision, not a complete procedure. For example, a source-comparison method may transfer its insistence on distinguishing observations from explanations while its exact source hierarchy, timing and review process do not transfer. This is a hypothetical example, not an observed cross-domain success.

The local-use/rationale argument says a successful configuration underdetermines which parts made it work. Retaining why a decision answered a constraint gives later work something to challenge. But I would resist a stronger reading that retained rationale is the only route: varied use can test transfer, and a sufficiently capable fresh model can reconstruct a useful explanation. The note itself recognizes varied-use testing. The economic question is which route costs less for the actual stream of work.

The neighboring sample-efficiency conjecture is appropriately conditional: retained theories may save new observations under shifts that preserve relevant structure. Its distinction between fewer observations and lower total cost is essential. It also exposes a bootstrap risk: the system may spend heavily discovering a beautiful reusable account of a structure that future users seldom revisit.

New emphasis from this exploration: the future task stream is part of the learning opportunity. A method is not economically reusable merely because a second application can be imagined. Recurrence, change in conditions, access to feedback and the cost of fresh reconstruction jointly determine whether reuse matters. This is reasoning from the mechanism, not an estimated frequency distribution.

4. Where present usefulness conflicts with future automation

Rescue versus diagnosis

An operator who fixes a failing run can save the current deliverable while removing the failure signal needed to improve the method. Conversely, withholding rescue for cleaner learning evidence may waste a user's time or risk a consequential error. The useful compromise is not mandatory nonintervention: distinguish the delivered result from what the system achieved before intervention. A rescue can itself be evidence if its decisive contribution survives.

Exploiting familiar work versus finding the boundary

Repeated familiar tasks offer immediate value and cheap reuse. They can also leave a narrow method looking general. A boundary case may be a worse current investment but a better test of future reliability. The selection issue is therefore not simply useful work versus research; it is which useful work produces discriminating experience.

Fast instruction versus revisable evidence

A concise instruction can reduce current execution cost but discard conditions and counterexamples needed to revise it later. Keeping everything imposes retrieval and review costs. The use-and-outcome records workshop already owns the record-design question. This exploration proposes no competing schema: it only identifies the later decision that makes selective retention valuable.

Reliable specialization versus adaptation

A tightly checked method can perform well on a stable task family while becoming expensive to adapt. Looser methods may adapt better but require more judgment now. This is a conjectured tradeoff, not a universal rule. Formal checks can also make adaptation safer; their value depends on whether they protect a real invariant or merely freeze an accidental implementation.

Automation versus observability

Removing human review may save attention but remove the only signal that detects subtle quality failures. A system can then appear cheaper precisely because its errors are no longer observed. The return on automation must include replacement feedback or explicitly accept the resulting uncertainty. More execution does not cure missing observation.

5. Failure modes particular to bootstrapping

• Self-confirming curriculum: the system chooses tasks its existing methods handle, then interprets those successes as increasing breadth. This is a plausible selection mechanism, not a measured finding here.
• Hidden human learning: the operator learns how to prompt, decompose and repair while the technical system's retained methods remain weak. The human-inclusive service improves, but the automation claim has not been earned.
• Evaluator inheritance: methods become better at producing outputs accepted by current checks while defects in those checks persist. Learning the evaluator is not the same as learning the user's purpose.
• Frozen abstractions: early distinctions shape every later diagnosis; failures are treated as missing instructions inside that scheme rather than reasons to replace it.
• Premature closure: a successful current method is promoted to mandatory policy, reducing opportunities to discover a better alternative.
• Regression hidden by averages: improvements on newly emphasized work conceal damage to old work or inability to improve again.

The last distinction has bounded external support. Do Agent Optimizers Compound? retains a phased protocol and table in which static improvement, transfer and re-optimization have different profiles across optimizers. I use that only to show that those outcomes can come apart in a reported benchmark. Continuing to improve after retaining an earlier result does not by itself show that retention made the next improvement more productive. A fresh-start comparison is needed for that stronger inference; this is my causal reading, not a reproduced result.

6. Strongest rival: a strong fresh solver with cheap records

The strongest practical rival is not a system that forgets everything. It is a current general model with good tools, authoritative project facts, ordinary episode records and task-local reconstruction, but little automatic methodology accumulation. Human maintainers add a reusable method only after a concrete repeated bottleneck appears.

This rival can win because reconstruction uses current capabilities without carrying stale instructions, speculative abstractions and their maintenance burden. It can still benefit from model improvements purchased externally. It also does not need to solve general method selection before doing useful work.

The bootstrap could beat it where local constraints are expensive to rediscover, feedback is sparse, failures recur, and retained reasons identify what changed. Those are candidate favorable conditions, not proof. An honest test must give the rival the same tools and source facts, rather than withholding basic project information and calling the resulting deficit a learning advantage.

The decision changed by this rival: do not invest first in a general extraction pipeline. First identify a later decision that an earlier episode should improve and compare retained-method use with inexpensive reconstruction. A null result should narrow or stop investment in that method, not automatically trigger a more elaborate memory system.

7. A bounded comparison worth considering, not an adopted protocol

The decision is whether retained methodology deserves further investment for one recurring kind of knowledge work. A small future comparison could give fresh workers the same model, tools and project facts, varying only access to the earlier method and its evidence. Compare that with access to ordinary episode records and task-local reconstruction. Charge preparation as well as execution; neither historical abstraction work nor reconstruction is free.

Include a new case that preserves the proposed mechanism, and a case where a named applicability condition fails. Ask whether workers reach a warranted result, correctly decline the old method where needed, and require different operator repair. Distinguish faster execution from better adaptation. If later method development is the claimed return, a second development episode—not just a second task execution—is needed.

This is a discriminating shape, not a request to build a benchmark or launch more agents. Repeat counts, tasks, costs and acceptance criteria remain undecided. Ordinary use observations can justify cheap reversible trials before controlled comparisons; the neighboring records workshop correctly warns against requiring a full experiment for every useful scrap of evidence.

8. Unknowns, parked ideas and evidence limits

• Unknown: the distribution of future work, recurring failure mechanisms, operator repair burden and maintenance cost. Without those, no general investment optimum follows.
• Unknown: whether Commonplace's retained methodological explanations improve transfer relative to fresh reconstruction. Its instrument reference explicitly says disciplined transfer has not been demonstrated.
• Parked: task-order perturbations as a test of path-dependent learning. I located and read the ingest kb/sources/on-the-fragility-of-self-improving-agents.ingest.md, but its Quotes section contains no retained source passages. I do not rely on its empirical claims here. Grounding would require a separately authorized operation; no source mutation was attempted.
• Parked: whether preserving rejected candidates gives more future value than preserving the selected method alone; whether learned evaluators can replace some operator judgment without reproducing its blind spots.
• Parked: choosing real work partly for the information it offers about method limits. This belongs with later turns on productive drift and investment, not a new controller commissioned here.

Existing notes above supply arguments, not independent confirmations. External findings are limited to the retained quoted source material identified above and were not reproduced. The WikiSkill synthesis was also read as an architectural comparison, but no implementation or performance claim from it is needed for this conclusion. No new capability improvement is demonstrated by this exploration itself.

Costs and handoff

Available activity accounting: one explorer output, zero subagents, zero external fetches or ingests, zero executed learning experiments, and zero changes outside the owned file. No nested assignments or session paths exist. Token usage, monetary cost and elapsed runtime are unavailable to this worker and are unknown, not zero.

The coordinator retains integration and scheduling. The most consequential proposed revision is to separate useful execution, later transfer, adaptation and improved method development before deciding what should grow. This is within the current purpose; adopting a different objective or a general learning architecture would require an operator decision. Broader graph connection work, if desired after integration, belongs to cp-skill-connect under a separate authorization.