DRENDA Public Evaluation Lab

T06 · Does learning in text improve acquisition in numeric tasks?

DRENDA's bounded synthesizer builds M1, then reuses it in text capability M2. Validated reuse updates its search policy. On a separate numeric branch, it acquires N1 and attempts N2 under the same data and a 500-candidate search budget.

Compare with text experience → without it → remove only the learned policy → restore exactly → restart. The controls preserve the numeric parent and all observations. A larger 50,000-candidate diagnostic budget measures the additional search cost; it is explicitly not the matched-budget comparison.

Four composition forms were known during development. Each public trial creates new random study observations; hidden outputs are drawn after candidate freeze. The candidate sees no teacher formula or holdout labels. This measures transfer of a reuse policy between text and numeric search, not arbitrary cross-domain abstraction.

The search-order repair and protocol were externally engineered; the bounded DRENDA mechanism synthesizes the particular operators and updates its policy from validated experience. The earlier 16-example protocol failed and is preserved; the frozen public amendment uses 64 generic observations. No draw is retried to obtain PASS.

The first public R15 deployment trial was INVALID because the DAC-restricted supervisor could not read a disposable checkpoint under the service's 0077 file-creation mask. Its original signed INVALID is preserved. The transport repair changes permissions only inside a root-private disposable lease; the candidate, data, Judge and acceptance rules are unchanged.

A separate Judge replays artifacts without importing the candidate. Download raw synthetic observations, checkpoints, receipts and signed verdict; verify all controls offline. The private Resident is never accessed.

Select T06 when registered · Read T06's frozen protocol

T05 · Compose a learned capability into a structurally different task

Family A maps a sequence to a Boolean. Family B maps a sequence and an independent context input to a Boolean, using a learned A artifact as an executable parent. This changes the input structure and dependency graph, not just names. The preserved DRENDA mechanisms receive observations, not a solution program or hidden answers.

Compare B without A → B with learned A → remove only A → restore the same hash → restart. The baseline runs two existing acquisition routes on the same 62 B observations and frozen limits. All B conditions use the same 256 held-out inputs. The page separates wrong answers, abstentions, and execution/dependency errors. Removing A makes its executable dependency absent; this is not 256 wrong answers.

This is a bounded composition experiment in existing representation languages, not arbitrary or distant cross-family transfer. New server-side draws rename symbols/fields and reorder a fixed exhaustive support. The post-freeze draw changes holdout order, not its support. Seven fresh candidate processes have a maximum of 90 seconds each; one attempt is retained, including FAIL or INVALID.

T05 accepts only a private challenge label, never code, seeds or a problem description. It is available only when confirmed by the registry. Public verification replays generated data artifacts independently; the learner runs only on the private server. No private mechanism source is downloaded.

Select T05 only when registered · Read the frozen T05 protocol and limits

T04 · Create an executable tool from observations

The preserved DRENDA synthesis mechanism receives 31 observed input/output transitions, with opaque names and symbols. It constructs a data-only finite-state tool, freezes it, then a new process executes it on 64 longer sequences generated after the freeze. No target machine, hidden answers or program is supplied to the learner.

Seven fresh processes cover no-tool baseline → construction → restart/reuse → tool ablation → exact restore → constant-output replacement → restore again. The separate Judge reconstructs the synthetic experiment and interprets the generated artifact independently. All draws and outcomes are preserved, including failures.

The existing search algorithm and representation language are engineering; the particular tool is synthesized by the unchanged DRENDA mechanism. This tests bounded tool construction, not arbitrary software creation. The no-tool baseline and ablation have no executable predictor; the constant-output replacement provides a behavioral control. This holdout stays within the same sequence family.

Select T04 when confirmed by the live registry · Read the tool-creation protocol

T03 · Learn, reuse, ablate, restore

The existing DRENDA experiment-selection learner receives eight generated demonstrations and synthesizes a ranking program. Only after that artifact is frozen are 48 new tasks generated, with new names, values, order and numeric outcomes instead of categorical labels. Every task is retained, including errors.

Five fresh processes measure: before study → learning → restart/reuse → learned-artifact ablation → exact restoration. The separate Judge recomputes the training and holdout scores. The result page shows the measurements; the proof reveals the synthetic dataset seeds, learned program and all five receipts for offline audit.

This is supervised learning of a preference over already-described experiments, using four existing structural features. The demonstrations are supplied by the evaluator protocol. The DRENDA learner, unchanged from its installed version, synthesizes the program; it is not supplied the teacher's coefficients or the holdout answers. Your text labels the run; it is not the problem being solved.

Run T03 below · Inspect the frozen learning protocol

Published deployment check, 25 September 2026: 40/48 new-task decisions after learning and restart; 5/48 with the learned artifact removed; 40/48 after exact restoration, on the same 48 tasks. The separate source baseline was 2/8, not the same dataset. This is an operator-assisted public run, not an independent reviewer result. Inspect this recorded result · Download its proof.

Transfer here means new tasks within the same four-feature grammar, not transfer to a different problem family. The proof includes all 48 outcomes, including the eight errors.