evaluate-and-optimize

1# SPDX-License-Identifier: Apache-2.02# yaml-language-server: $schema=https://nika.sh/spec/v1/workflow.schema.json3#4# TEMPLATE · evaluate-and-optimize · draft → two fixed critique/revision rounds.5#6# The bound is structural: exactly two revision rounds are unrolled in the DAG,7# followed by one score-only measurement of the final result. The model cannot8# add another turn, and every infer has its own token ceiling.9#10# Needs · nothing. Runs offline in an empty directory with mock/echo.11# Run · nika run <file> --model mock/echo12nika: evaluate-and-optimize13​14model: mock/echo                     # SLOT: replace with a local or hosted seat15​16const:17  brief: "Explain why idempotent retries matter in two sentences."   # SLOT18​19⛨permits: {}20​21tasks:22  draft:23    ◇infer:24      max_tokens: 30025      prompt: "Draft this deliverable: ${{ const.brief }}"26      schema:27        type: object28        additionalProperties: false29        required: [text]30        properties:31          text: { type: string, minLength: 1 }32​33  evaluate_1:34    with:35      candidate: ${{ tasks.draft.output.text }}36    ◇infer:37      max_tokens: 20038      prompt: |39        Critique this candidate against the brief. Name one concrete repair.40        Brief: ${{ const.brief }}41        Candidate: ${{ with.candidate }}42      schema:43        type: object44        additionalProperties: false45        required: [score, repair]46        properties:47          score: { type: integer, minimum: 0, maximum: 100 }48          repair: { type: string, minLength: 1 }49​50  optimize_1:51    with:52      candidate: ${{ tasks.draft.output.text }}53      repair: ${{ tasks.evaluate_1.output.repair }}54    ◇infer:55      max_tokens: 30056      prompt: |57        Revise the candidate using the repair, without changing the brief.58        Candidate: ${{ with.candidate }}59        Repair: ${{ with.repair }}60      schema:61        type: object62        additionalProperties: false63        required: [text]64        properties:65          text: { type: string, minLength: 1 }66​67  evaluate_2:68    with:69      candidate: ${{ tasks.optimize_1.output.text }}70    ◇infer:71      max_tokens: 20072      prompt: |73        Critique this revision against the same brief. Name one final repair.74        Brief: ${{ const.brief }}75        Candidate: ${{ with.candidate }}76      schema:77        type: object78        additionalProperties: false79        required: [score, repair]80        properties:81          score: { type: integer, minimum: 0, maximum: 100 }82          repair: { type: string, minLength: 1 }83​84  optimize_2:85    with:86      candidate: ${{ tasks.optimize_1.output.text }}87      repair: ${{ tasks.evaluate_2.output.repair }}88    ◇infer:89      max_tokens: 30090      prompt: |91        Produce the final revision using the last repair.92        Candidate: ${{ with.candidate }}93        Repair: ${{ with.repair }}94      schema:95        type: object96        additionalProperties: false97        required: [text]98        properties:99          text: { type: string, minLength: 1 }100​101  final_evaluate:102    with:103      candidate: ${{ tasks.optimize_2.output.text }}104    ◇infer:105      max_tokens: 200106      prompt: |107        Score the final candidate against the brief. Evaluate the candidate108        exactly as written without proposing another revision.109        Brief: ${{ const.brief }}110        Candidate: ${{ with.candidate }}111      schema:112        type: object113        additionalProperties: false114        required: [score, reason]115        properties:116          score: { type: integer, minimum: 0, maximum: 100 }117          reason: { type: string, minLength: 1 }118​119outputs:120  result: ${{ tasks.optimize_2.output.text }}121  final_score: ${{ tasks.final_evaluate.output.score }}122​