| name | final-bundle-quality |
| description | Evaluate and replay final_bundle PDF/LaTeX quality for Research Harness project-level self-evolution. Trigger on phrases like "final bundle quality", "PDF quality replay", "LaTeX quality gate", "AAAI PDF QA", "final_bundle_quality", or equivalent requests to scan publication bundles, classify layout/encoding/recompile defects, and decide whether a skill can be promoted. |
| category | research |
| version | 0.1.2 |
| public_suite | benchmark |
| task_type | evaluation |
| data_access_level | local-files |
| inputs | topic_id, final_bundle artifact, TeX/PDF/log paths, target profile |
| outputs | final_bundle_quality_report, replay trajectory, promotion decision |
| gates | zero blocking and zero recompile-repairable issues before promotion claims |
| artifact_schemas | final_bundle_quality_report, runtime_trace_grade |
| estimated_runtime | 5-30 min |
Final Bundle Quality
Use this skill when a Research Harness topic has a final_bundle artifact and the goal is to verify PDF/LaTeX quality, compare quality trajectory, or evaluate whether a project-level repair policy can be promoted.
Inputs
- Source topic id and active
final_bundle artifact id.
- Target profile, for example
AAAI-style-science-discovery.
- TeX, PDF, log, and optional bibliography paths from the artifact payload.
- Lifecycle pack and trajectory gate artifacts when evaluating promotion.
- Held-out
final_bundle artifacts from another topic.
Workflow
- Scan the source bundle with
scan_pdf_quality_bundle.
- Classify issues into blocking, recompile-repairable, layout-only, and encoding groups.
- Build or load a path-safe
pdf_quality_replay_trajectory; do not infer historical quality by rescanning shared mutable paths.
- Run the trajectory gate before any held-out replay decision.
- Collect held-out replay signals from active non-source
final_bundle artifacts.
- Evaluate lifecycle promotion with
evaluate_skill_lifecycle_pack.
- If a
multi_route_review_report exists, parse it as strong readiness guidance before reporting significant PDF quality improvement.
- If the blocker is figure readability, placeholder diagrams, or weak visual evidence, trigger
figure-quality before reporting publication-readiness improvement.
- Record gate, discovery, and management artifacts before changing global skill behavior.
Promotion Rules
- Source trajectory must pass final PDF quality thresholds.
- Source bundle and held-out bundle must have zero blocking and zero recompile-repairable issues.
- Layout-only warnings must include line ranges, source excerpts, or accepted thresholds.
- Real held-out replay must pass page, citation, anonymization, and bibliography gates.
- Controlled or synthetic backfills can exercise replay plumbing but must not count as promotion evidence.
- Exact target-profile replay is preferred. Compatible target profiles are allowed only when explicitly supplied through
compatible_target_profiles, and the artifact metadata must preserve target_profile_match=compatible.
- Promotion metadata must distinguish
promotion_strength=exact_profile_held_out from promotion_strength=compatible_profile_held_out; do not report compatible-profile evidence as exact-profile validation.
- A clean deterministic PDF scan does not override a multi-route consensus reject; reviewer consensus blocks publication-readiness handoff until project-level review gates record and address the blockers.
- Novelty, scope, and evidence objections from multi-route review are submission-readiness blockers only; this skill must not mutate topic-level research strategy or topic self-evolution policy.
- Figure/table quality objections should route to
figure-quality; deterministic PDF cleanliness does not make placeholder figures publication-ready.
Current Evidence
- Current strongest promotion evidence is exact-profile held-out replay:
project_artifacts:771, topic 5 ResearchFlowBench recompiled through the RH AAAI-style compiler path.
- Current lifecycle gate:
project_artifacts:772, action promote, promotion_strength=exact_profile_held_out, target_profile_match=exact.
- Current strongest reviewer-readiness evidence is
project_artifacts:778: all completed host-agent routes rejected layout-v8, with minimum scores aaai_readiness=2, evidence=2, visuals=3, format=2.
- Earlier compatible-profile replay remains useful historical evidence:
project_artifacts:757, target_profile_match=compatible, promotion_strength=compatible_profile_held_out.
- Controlled replay
project_artifacts:752 remains plumbing-only evidence and must not count as promotion evidence.
Memory Policy
- Transferable memory: issue category, severity, target profile, gate outcome, and repair policy.
- Topic-local memory: topic id, artifact id, file paths, source paths, and one-off bundle layout details.
- Do not hard-code topic 8 paths, artifact ids, filenames, or AAAI-specific thresholds into project-level behavior.
Known Lessons
- Tab-prefixed malformed LaTeX commands and invalid UTF-8 are recompile-repairable defects.
- Underfull hbox warnings are layout-only diagnostics and should not trigger infinite repair loops.
- Overfull hbox warnings require line-level diagnostics and should not mask citation or venue gates.
- A clean source topic is insufficient for project-level promotion without held-out replay.
- A compatible-profile held-out replay can promote this skill only when explicitly declared and recorded.
- When recompiling single-column source material into AAAI-style two-column drafts, infer
tabularx/array dependencies and promote tabularx{\textwidth} tables to table* floats to avoid high-severity overfull boxes.
- Multi-route review consensus can be stricter than deterministic PDF scanning: official venue style, evidence-bearing tables/visuals, reproducibility artifacts, and reviewer score minima should guide project-level gates before any human handoff claim.
