Submitting A Claim Bundle

This page defines a practical, reproducible submission workflow.

Required files

• protocol.yaml
• hypothesis.jsonl
• evaluation_result.json
• manifest.json

Optional:

• cross_model_results.json

Submission lifecycle

1. Build bundle artifacts from discovery outputs.
2. Validate schema and hash integrity.
3. Run evaluator and report generation.
4. Run deterministic submission review.
5. Resolve failed gate clusters.
6. Resubmit with updated evidence.

Mandatory validation commands

python -m automechinterp_evaluator.cli evaluate \
  --bundle /path/to/bundle \
  --output /path/to/bundle/result.json

python -m automechinterp_evaluator.cli report \
  --bundle /path/to/bundle \
  --output /path/to/bundle/stage_gate_report.md

python -m automechinterp_evaluator.cli submission-review \
  --bundle /path/to/bundle \
  --reruns 3 \
  --output-json /path/to/bundle/submission_review.json \
  --output-md /path/to/bundle/submission_review.md

What makes a submission strong

• complete confirmatory and exploratory slices where required
• coherent control evidence
• robust behavior across perturbations
• statistical checks that pass under declared policy
• deterministic rerun agreement

What usually blocks acceptance

Failure family	Typical symptom	Suggested follow-up
confirmatory completeness	confirmatory_present fails	regenerate confirmatory slice
method sensitivity	unstable outcomes across methods	tighten intervention protocol consistency
robustness	effect collapses under perturbations	add robustness-focused follow-up experiments
controls leakage	strong control effects	redesign controls and baseline comparisons
statistical rigor	CI/power/multiplicity failures	increase evidence quality and sample adequacy

Submission package recommendation

When sharing externally, include:

• full bundle directory
• submission_review.json
• submission_review.md
• protocol hash/version
• environment manifest path