Skip to content

Autonomous research system for measurable, computer-executable research.

Python
1.7 KB
Raw Blame

Praxist Task Template

This is a scaffold for creating an external Praxist task project. It is not a scientific benchmark and should not be used for real research until the task owner replaces the placeholder domain material.

Required Replacements

Before running a real experiment, replace:

  1. task.yaml task identity, metrics, budget defaults, workflow refs, roles, audits, evaluations, tools, and output policy.
  2. description.md and prompt_task.jinja2 with the task's actual research brief, constraints, and agent instructions.
  3. roles/ with task-owned role skills.
  4. audit_rules/ with task-owned scope, claim, and agenda criteria.
  5. evaluations/pareto_tiered/run.py with the task's single public evaluation entrypoint.
  6. assets/ with the task's harness, baselines, fixtures, dataset metadata, literature packs, and optional reference implementations.

Boundary

The task project is an explicit Praxist input selected with --task-path. Task facts, benchmark rules, and domain-specific prompts belong here, not in praxist/core or generic system plugins.

Default Research Loop Policy

This scaffold declares gen0 DIG, independently switchable initial/later QD, and a continuous-evolution Gems policy in task.yaml so new tasks start from the standard research-loop shape:

  • pre-code DIG planning is enabled;
  • gen0 cohort-level QD allocation is enabled; later QD is explicitly disabled for this one-generation smoke fixture;
  • periodic Gems reset is disabled by default.

Replace the placeholder diversity dimensions with task-specific axes before a real run. Enable periodic Gems reset later only if the operator requests it or a diagnostic pass identifies a performance ceiling and recommends a reset cadence.