`task_evolve` — SETA curriculum variants¶

seta_evol reads a complete parent Harbor task and applies an explicit evolution strategy before rebuilding its fixtures, instruction, reference and tests. Parent files remain unchanged. Child metadata records the parent bundle hash, strategy and variant number. The workflow follows SETA's evolution pipeline.

Pipeline, step by step¶

flowchart TD
  S["Integrity-checked parent Harbor task"] --> A["Load bounded text assets and parent hash"]
  A --> ST["Select one of six explicit strategies"]
  ST --> P1["P1 · Design a coherent child task"]
  P1 -->|"filtered_reason"| X["Retain reason; skip variant"]
  P1 --> P2["P2 · Build child fixtures, tests and reference"]
  P2 --> Q["Q1 · Optional draft consistency review"]
  Q --> R["Fresh remote baseline + reference"]
  Q -->|"Blocking issue within max_repairs"| P2
  R -->|"Failure feedback within max_repairs"| P2
  R -->|"0 / 1"| H["Export child with parent hash and strategy"]

P1, P2, … identify actual model calls. Unlabelled stages are code or remote execution.

Read the complete parent. The author sees the parent instruction, environment, private tests and reference as private source evidence. The parent bundle hash must still match.

Select and apply a strategy. Variants select strategies round-robin: increase, decrease, change context, combined increase/context change, slight increase or slight decrease. There is no model-based difficulty calibration here.

Rebuild the child. The builder generates a standalone child using the evolution design, then the common terminal runner validates and executes it. Parent files are not edited.

Every prompt and its data¶

Two authoring calls on the first successful attempt: evolution design, then builder. Strategy selection is deterministic. When review_drafts: true, Q1 reviews each complete draft before execution; a blocking issue returns to the bounded builder loop.

Call	System prompt composition	User / input material	Output	Retry or branch
P1 · Evolution	evolution_prompt.md + selected strategies/*_adapter.md + owned adaptation	Complete parent files, strategy, parent hash and ordinal variant.	EvolutionDesign: core_capabilities, draft_spec, filtered_reason	A reason filters the variant without building it.
P2 · Child builder	builder_prompt.md + shared materialization instructions	Child design and execution/schema feedback.	TerminalDraft	Default initial attempt plus two repairs.

Read the complete seta_evol prompt reference for every retained template, appended instruction, substitution, example and output schema. The shared prompt guide explains how to inspect the fully resolved request from a real run.

Follow one task¶

Illustration: evolve a log-aggregation task by adding a time-window requirement. The child must change the requested behavior and tests coherently; changing only filenames is not a valid intended transformation.

What repeats, what is checked¶

The strategy prompt is applied once per variant. Build retries repair the selected design rather than silently choosing a different strategy. Difficulty labels remain design intentions until later solver measurements.

An exported bundle is a generation result. Independent leakage review, shortcut probes and blind solver traces belong to the later quality campaign.

Implementation map¶

Run and supported profile¶

Use source.kind: task with either an owned task directory or a directory of owned tasks. The current profile requires integrity-checked text assets totalling at most 150 kB per parent. Binary-heavy and legacy unhashed tasks need a dedicated input adapter; they are not silently accepted as equivalent inputs.

repo2rlenv generate --config examples/owned-seta-evol.yaml

The shared terminal generation options apply, plus:

Option	Default	Meaning
`variants_per_parent`	1	Distinct variant slots per parent, at most 20
`strategies`	increase, context change, decrease	Round-robin strategy selection

Exact strategy identifiers are increase_difficulty, decrease_difficulty, change_context, increase_difficulty_and_change_context, slight_increase and slight_decrease. The slight strategies are design directions; their effects on model success rates are unmeasured until the later rollout campaign.

The owned recipe retains the upstream evolution, strategy and builder prompts, with Apache-2.0 notices. Typed responses, offline execution and owned Harbor/JUnit materialization replace the upstream agent filesystem interface and legacy task templates. This is a workflow adaptation, not byte-identical reproduction.

The released collection contains 100 children: 20 retained tasks and 80 new exports. All 80 new exports have matching fresh baseline reward 0 and reference reward 1 evidence. The six strategy counts are 13 increase, 21 decrease, 15 context change, 20 combined increase/context change, 14 slight increase and 17 slight decrease. These describe requested transformations, not measured difficulty.

The expansion recorded $44.72: $30.75 in model estimates across 392 calls and $13.97 in estimated worker/build costs, or **$0.56 per new export** including unsuccessful attempts. One timed-out model request retains a separate $1.25 uncertain reservation; it is not silently reported as either billed or free. All expansion workers are terminated. These costs exclude the 20 retained tasks and subsequent independent quality evaluation. See the generation economics and published Harbor dataset.

Generation does not imply difficulty calibration or independent quality acceptance. See RFC 0014 and the packaged recipe's provenance.md.

task_evolve — SETA curriculum variants¶