AutoResearch Readiness

Use this module for deterministic planning and readiness validation. It adapts reviewed AutoResearchClaw design ideas as file-backed template controls, not as an autonomous research agent.

AutoResearch CLI-style measurement ideas are treated the same way: adopt exact metric extraction, review-status vocabulary, baseline/noise/confidence disclosure, and append-only evidence discipline; do not add lifecycle hooks, git commit/revert ownership, or no-human autonomous loops by default.

Commands

bash
uv run python -m infrastructure.autoresearch.cli validate --project templates/template_code_project --fail-on-issues

Public API

python
from infrastructure.autoresearch import (
    AutoResearchConfig,
    AutoResearchIssue,
    AutoResearchPlan,
    AutoResearchReport,
    mad_confidence,
    metric_unit_from_name,
    build_autoresearch_plan,
    load_autoresearch_config,
    parse_metric_lines,
    parse_string_sequence,
    validate_autoresearch_plan,
    write_autoresearch_report,
)

validate_autoresearch_plan(..., phase="intrinsic"|"extrinsic"|"all") splits pre-write structure checks from post-write artifact checks.

Configuration

Project-local autoresearch.yaml supports:

• enabled
• strict
• topic
• quality_checks
• stage_gates
• required_artifacts
• security_profile (mapping: enabled, mode, integrity_algorithm,

network_policy, external_signing, threat_model_frameworks)

• source_manifests (list of source-manifest artifact paths)

stage_gates must use exact stage names from pipeline.yaml. The full accepted key set is defined by _CONFIG_KEYS in infrastructure/autoresearch/config.py.

Guardrails

Keep v1 deterministic: do not add network calls, LLM calls, generated-code execution, or autonomous loops here. Delegate execution and validation to the existing pipeline, project, validation, and reporting modules.

Use parse_metric_lines() only for output already produced by a trusted local command. It accepts exact METRIC name=value lines and rejects ambiguous or invalid metric evidence. Use mad_confidence() as a disclosure helper for baseline/best/noise comparisons, not as an automatic publication decision.