RAPP Gyre

1. The question

Frame Chains studied structural continuity: content-addressed history, recovery, and orchestration. This follow-up asks about behavioral continuity: can useful changes to a program be carried in a complete frame, inherited by later candidates, and retained under a fixed task contract?

A source frame contains the literal program, not a claim that a capability exists. A mutation lens proposes a complete replacement. An external evaluator observes what the candidate does. Those roles must not be confused.

Here, speculative or hallucination-driven generation means proposing counterfactual program variants. It does not turn a generated success story into evidence. The language model's pretrained programming knowledge remains a substantial input.

2. What the pilot actually did

The seed implemented flat lists of signed integers. The fixed extension added inclusive ranges and nested repeat groups while preserving the original interface and rejection behavior. The two conditions used Opus 5, the same contract, the same interpreter and cases, the same number of candidate requests, and the same parent-selection rule.

The Brainstem kernel stayed fixed. Program source changed; model weights did not. Candidate code ran in a bounded Python/data interpreter hosted in separate virtual environments. A venv is not itself a security sandbox.

Parent choices were fixed before opening the reserved audit. Both conditions reached the frozen-case target in generation one. Generation two retained those scores and selection favored smaller syntax trees under the predeclared tie-break. A smaller syntax tree does not establish better semantics or lower execution cost.

The original behavioral gates did not discriminate among offspring: every candidate passed them, including candidates later shown to fail additional permitted inputs. This run therefore did not establish that eliminating failed offspring, adding generations, or using a speculative lens caused an advantage over direct repair.

Post-hoc counterexamples: the frozen scores stay unchanged

An automated adversarial review found a real regression. The program speculative-1/g1 rejects unsigned numbers longer than six digits, although the contract allows leading zeros. It was not selected as a parent. A subsequent reproduction applied the same three probes to the seed and every frozen candidate.

Post-hoc probe outcomes are reported separately. The original 109 selection-stage and 119 reserved-audit cases were not changed or regraded. These targeted probes do not establish an unbiased arm comparison. They do refute complete contract-preservation claims for the affected programs under the stated execution conditions.

A perfectly valid frame can carry a defective program. Integrity checks and behavioral acceptance are different obligations.

3. Inspect and reproduce the evidence

The public package is an allowlisted projection of a synthetic experiment. It is not an export of the author's estate, authentication material, machine logs, or unrelated conversations. Program frames retain their original contents and ancestry; any sanitized call metadata is labeled as a projection.

The legacy result field original_preserved means that the original controller's frozen original-case set passed. It is not a guarantee over all permitted inputs. The field is preserved as recorded rather than renamed inside immutable frames.

git clone https://github.com/kody-w/frame-chains.git
cd frame-chains
python3 gyre/verify.py
python3 gyre/verify.py --self-test
python3 gyre/privacy.py

No model access is needed to replay the published programs and outcomes. Replaying stored artifacts is different from independently regenerating those artifacts through a stochastic model. The package must not claim to authenticate the original provider responses or reproduce hidden provider internals.

The released audit cases were withheld from candidate generation and parent selection in the original run. They are public now. A new confirmatory experiment must use newly prepared evaluation instances rather than treating this published set as still hidden.

RAPP/1 content addresses make records checkable relative to a retained head. They do not, by themselves, prove authorship, trusted timestamps, or semantic truth. These are unsigned research frames, not the predecessor paper's signed authority mechanism.

4. What would justify a stronger claim?

In his 1974 Caltech address, Richard Feynman argued for reporting what could invalidate an interpretation, including competing explanations and contrary results. Applying that standard is more useful than guessing what would personally have persuaded him.

At a fixed resource budget and with a fixed external model, does inherited program-and-lens adaptation produce more retained capability on independently prepared tasks than independent best-of-N generation or fixed-lens search?

That is a falsifiable question. An impressive isolated offspring, a large population, or a growing frame archive is not sufficient to answer it.

What would count against the hypothesis?

If independent best-of-N generation performs just as well at the same budget, the population may be buying attempts rather than cumulative adaptation. If resetting inherited state does not hurt, inheritance has not shown a causal benefit on that distribution. If evolving lenses only consumes extra inference, the meta-level has not earned its cost. If success disappears under an independent grader, the system may have exploited the evaluation rather than acquired the claimed capability.

The current direct-repair tie is part of that record, not something to explain away.

5. The proposed Gyre, and the experiment it still needs

RAPP Gyre is a proposed role-pivoted system: program source, lens source, and eventually the literal Brainstem program can all be represented as frames. A pivot assigns which frame is the subject, which is the transformer, which runtime materializes the candidate, and which dimension is being explored.

Parallel workers would explore separate branches from pinned ancestors. Completed observations could guide later pivots. The externally fixed original-use-case contract would determine promotion; a mutable candidate could not weaken its own mandatory evaluator. The full co-evolving architecture was not run in this pilot.

The term quantum in dimension-mining descriptions is a metaphor for concurrent alternatives and measured selection. There is no claim of quantum computation or access to physical alternative universes. The existing QQDRILL protocol also separates local discovery from compatibility-gated assimilation. The proposed generative Gyre adds speculative candidate creation; it does not silently redefine a read-only lookup as a successful program mutation.

Prospective research specification records the controls, disconfirming outcomes, and prerequisites. It is a draft design, not a claim that a confirmatory trial has already been registered or run. Task generators, budgets, endpoints, exclusions, sample size, and stop rules must be frozen and publicly timestamped before that trial's primary data are collected.

A large number of case rows is not a large number of independent experiments. The unit of replication must be the independently sampled task/run, not correlated siblings or many inputs to one parser. Both successful and unsuccessful runs must remain in the denominator.

6. Prior art and position

The RAPP/1 substrate and the predecessor's orchestration work are not claimed as new here. Nor is the general idea of using language models to propose programs and an evaluator to select them. Any distinctive contribution must be located in the implemented composition, its operational guarantees, or a measured result that survives comparison with existing methods.

This overlap is encouraging for plausibility and raises the novelty bar. A possible RAPP contribution is an interoperable representation, reconstructable execution, compatibility contracts, and inspectable provenance across these roles. It is not evidence of a new physical principle or an automatically better search algorithm.

Reading locations: FunSearch accepted manuscript methods and supplementary ablations (PDF positions 3-6, 17-18, 28-29); AlphaEvolve pp. 5, 16, 18 and Figure 8; Promptbreeder pp. 1-3; STOP original version pp. 1-5; Darwin Gödel Machine sections 3, 4.3, and Appendix F. Adaptive use of evaluation data is also discussed by Dwork et al., Generalization in Adaptive Data Analysis and Holdout Reuse.

7. Limitations and publication status

One task family; one model; five siblings per condition per generation; provider-default sampling; no pure-random mutation or fresh-seed-each-generation control; a restricted execution language; no causal advantage observed over direct repair; no rejected offspring on the primary behavioral gates despite later counterexamples; and no demonstrated full-runtime or lens co-evolution. The external model supplied pretrained programming knowledge.

The original interpreter is not a general CPython-equivalent implementation. Review found that its list += handling does not preserve Python's in-place alias semantics. None of the twenty candidates uses augmented assignment, so that mismatch does not invalidate their frozen scores. It remains a reason not to generalize the interpreter's claims beyond the stated study.

Automated adversarial review is not independent human replication or peer review. A public replay can reproduce artifacts and computations, but cannot retroactively authenticate a private model invocation. No confidence interval over a population of task families is inferred from this single-family pilot.

The release implementation review also found two structured-JSON privacy gaps. Decoded-key inspection and strict duplicate-member rejection were added and independently rechecked before release. No original program or experimental observation was changed to close those findings.

Corrections belong in the public record. Original observations must not be edited into a more favorable history.