NVFLARE Auto-FL Report

Purpose

Turn the recorded evidence from a stopped NVFLARE Auto-FL campaign into a reproducible Markdown report, machine-readable JSON summary, and refreshed progress plot without changing the campaign or its results.

Use When

Use this skill after an NVFLARE Auto-FL campaign has stopped, reached an explicit cap, hit a hard blocker, or was manually interrupted. Use it when the user asks for the final report, achieved improvement, literature findings, failed ideas, reproduction details, or a refreshed progress plot.

Do Not Use When

Do not use this skill to start or continue optimization, invent missing results, or finalize a campaign that is still running. Use nvflare-autofl for the active candidate loop. Do not stop an active campaign merely because the user asks for a status snapshot.

Available Scripts

Run this bundled CLI directly with Python. This skill has no NVFLARE or agent run_script() helper; do not invent or call one. Resolve the script from the directory containing this SKILL.md, as shown in the workflow.

Workflow

1. Locate the job directory containing results.tsv, autofl.yaml,

.nvflare/autofl/campaign_state.json, and candidate manifests.

1. Confirm execution has stopped. Prefer campaign state with

finalresponseallowed=true. When a process was abruptly interrupted and state is stale, independently confirm no campaign or job process remains, then use --confirm-interrupted. Before finalizing, confirm that campaign state, results.tsv, and available candidate manifests contain no pending candidate. Finalize or abandon pending work through nvflare-autofl first.

1. Generate the deterministic report artifacts:

``bash python "$REPORTER" <job-dir> ``

Resolve REPORTER once to the absolute path of scripts/generatereport.py relative to this SKILL.md. Do not assume a provider-specific skill installation directory or $CODEXHOME.

1. Read both autoflfinalreport.md and autoflreportsummary.json. Check

warnings about metric use, executed budget changes, campaign-state/ledger disagreement, missing provenance, and incomplete interruption state.

1. Give the user the baseline, best score, delta, strongest candidate lineage,

concise "what helped" and "what did not help" findings, literature ideas that helped or failed, selection rationale, reliability caveats, and absolute artifact paths.

The helper attempts to refresh progress.png by reusing the product Auto-FL plotter. Plotting is optional evidence: if plotting dependencies are missing or the artifact is not a valid PNG, the helper preserves the artifact, records a warning and artifacts.progressplotavailable=false, and still writes the Markdown and JSON reports without embedding the broken image. It does not modify source, candidate manifests, results.tsv, or campaign state. All relative helper path options, including --plotter, resolve from the campaign job directory rather than the shell's current working directory. The helper holds the campaign lifecycle lock from evidence loading through artifact writes. It refuses a busy campaign and rejects writable output paths that alias campaign evidence, job.py, trust-contract source paths, or another output, including filesystem aliases. Outputs must not match the trust contract's allowed source-creation patterns. A read-only campaign archive is reportable only when its persisted campaign.lock already exists and writable output paths are supplied; the persisted lock target alone does not imply live contention.

Interrupted Campaigns

The report helper must refuse state with finalresponseallowed=false unless the human has said the campaign was stopped/interrupted and the agent confirms that execution is no longer active. Only then run:

python "$REPORTER" <job-dir> --confirm-interrupted

This records a reporting-time interruption assertion; it does not rewrite the campaign state or pretend the runner finalized cleanly. It bypasses only stale stop state. A candidate ledger row or pending-candidate state always blocks finalization. Any available candidate manifest whose status is not a recognized terminal value (keep, discard, crash, or abandoned) also blocks. Missing, unknown, or unreadable status is unfinished evidence because completion cannot be established. The agent must finalize or abandon that candidate first.

Troubleshooting

Report Contract

The final report must include:

• campaign termination reason, objective, metric source, direction,

environment, cap, abandoned-candidate count, and declared fixed budget;

• baseline, best retained result, score delta, runtime, failures, and status

counts;

• selected-candidate rationale, strict retained improvements, representative

non-improvements, grouped failures, and outcomes by recorded algorithm family;

• running-best trajectory selected by first, final, and largest objective improvements,

plus a refreshed progress.png when plotting is available, with explicit plot availability in the JSON summary otherwise;

• best-candidate manifest, patch hash, base-candidate lineage, inherited code

changes, artifacts, and exact baseline/best commands;

• every recorded literature checkpoint, its event ID and source markers,

explicitly linked candidates, and whether measured evidence helped, matched, failed, or did not confirm the idea;

• discarded/crashed ideas and deterministic comparability warnings;
• optional agent model, reasoning effort, cost, or tooling notes when supplied;
• absolute paths to autofl.yaml, results.tsv, campaign state,

progress.png, autoflreportsummary.json, and autoflfinalreport.md.

The report must distinguish imported/declared budget from executed command arguments. It must warn when the best candidate changed training compute or comparison population, when authoritative state disagrees with ledger-derived accounting, or when repeated selection used a test-like metric. Product Auto-FL campaigns retain the imported min or max direction; the report rejects legacy minimization evidence without direction provenance. It must not add PR-specific sections such as "Product Findings" unless the user explicitly requests them. best means a scored retained baseline or keep row; an unretained scored discard may appear only as best_observed. Candidate and crash rows never become retained best results, milestones, or literature improvements. Baseline identity is determined strictly by status=baseline, matching the campaign guard. The report preserves per-run metric name, extraction source, artifact, candidate kind, algorithm family, and literature event linkage. It does not infer algorithm families or mechanisms from candidate names.

Read [report-contract.md](references/report-contract.md) when interpreting lineage, literature outcomes, budget warnings, or interrupted state.

Limitations

• The report is only as complete as the persisted ledger, campaign state, and

candidate manifests; it does not validate or reconstruct unrecorded claims.

• Results from a single campaign do not establish robustness or

generalization.

• The helper does not start, stop, resume, or resubmit campaign jobs.
• Progress plotting remains optional, so Markdown and JSON may be produced

without an embeddable PNG.

• Copied campaign archives may retain only partial provenance when recorded

absolute manifest paths are no longer available.

Requirements

• Treat results.tsv as recorded evidence; never repair scores by guessing.
• Work without Git. Do not commit or push unless the user separately asks.
• Preserve the campaign and job sources exactly as found.
• Never bypass campaign-lock contention or output-path collision checks.
• Use candidate manifests when available, but still report partial provenance

when copied artifacts make old absolute manifest paths unavailable.

• Keep conclusions proportional to the evidence. A single run is a candidate,

not a robustness claim.

• For POC/production, report standard NVFLARE job IDs and downloaded artifacts

already present in the ledger; do not resubmit jobs during reporting.