Title-Boundary Heuristic Experiments

This note records deterministic Rust heuristic changes and Pi RPC reviewer evidence for DMHY template title_spans generation. The reviewer output is evidence only; recipes and training JSONL are changed only by Rust code.

Baseline

Input recipe evidence:

reports/dmhy_template_recipes.full_top5000.title_boundary_trial.seed.jsonl
total title-boundary decisions: 5631

Decision distribution:

decision	count
`merge_dissimilar_title_continuation`	4661
`keep_similar_duplicate_or_alias`	463
`keep_structural_boundary`	502
`keep_path_boundary`	5

Pi RPC stratified sample, --threads 4:

stratum	rows	model result	note
ambiguous merge, similarity `0.25..0.85`	20	20 merge	Good. Reviewer cited concrete title-word pairings.
high-sim `keep_similar_duplicate_or_alias`, similarity `0.85..1.0`	20	19 merge, 1 parse-error	Bad baseline heuristic. Examples such as `Asobi ni Iku yo!`, `Code Geass ~Akito the Exiled~`, `Yagate Kimi ni Naru`, and `Big Order` were wrongly kept separate by the heuristic.
structural keep	20	18 keep, 1 merge, 1 skip	Mostly good. Covers category+title, multilingual aliases, and distinct fields.
path keep	5	5 keep	Good. Covers path duplicate/category boundaries.

Conclusion: keep_similar_duplicate_or_alias was too broad because it used max literal similarity. Short connector repeats (no, wa, ni, the, of) and repeated title words inflated similarity to 1.0 even for ordinary contiguous title phrases.

Experiment 1: strict substantive duplicate keep

Change tried:

Keep adjacent TITLE SEP TITLE only when top literals include a substantive duplicate/near-duplicate pair.
Ignore short connectors and literals with normalized length < 6.

Generated recipe output:

reports/dmhy_template_recipes.full_top5000.strictdup.seed.jsonl

Decision distribution:

decision	count
`merge_dissimilar_title_continuation`	4661
`merge_adjacent_title_continuation`	444
`keep_similar_duplicate_or_alias`	19
`keep_structural_boundary`	502
`keep_path_boundary`	5

Pi RPC review of all 19 remaining keep_similar_duplicate_or_alias decisions:

rows	model result
19	19 merge

Examples the reviewer judged as merge:

Air In Summer, Assault Lily BOUQUET
Aoki Hagane no Arpeggio Ars Nova
Shoujo Kageki Revue Starlight
Motto! Ojamajo Doremi
Fate stay night Heaven's Feel
Arafou Otoko no Isekai Tsuuhan Seikatsu
Watashi, Nouryoku wa Heikinchi de tte Itta yo ne!

Conclusion: strict duplicate evidence still produced false keeps. The repeated literal was often a real repeated word inside one title phrase or a top-literal artifact, not a duplicate/alias boundary.

Experiment 2: merge all adjacent separator-only title slots

Final change:

TITLE PATH TITLE remains keep_path_boundary.
Non-separator structural boundaries remain keep_structural_boundary.
Every separator-only adjacent title boundary, TITLE SEP TITLE, now merges.
Low-similarity merges keep the old decision name merge_dissimilar_title_continuation.
High-similarity merges use merge_adjacent_title_continuation.
keep_similar_duplicate_or_alias is no longer emitted by the deterministic heuristic.

Generated recipe output:

reports/dmhy_template_recipes.full_top5000.mergeall.seed.jsonl

Decision distribution:

decision	count
`merge_dissimilar_title_continuation`	4661
`merge_adjacent_title_continuation`	463
`keep_structural_boundary`	502
`keep_path_boundary`	5

Pi RPC sample of high-similarity merge_adjacent_title_continuation decisions:

rows	model result
30	29 merge, 1 parse-error

The reviewer cited concrete title continuations such as:

Asobi ni Iku yo!
Code Geass ~Akito the Exiled~
Kami nomi zo Shiru Sekai
Yagate Kimi ni Naru
Big Order
Gate Jieitai Kanochi nite Kaku Tatakaeri
JoJo's Bizarre Adventure - Stardust Crusaders

Conclusion: merge-all adjacent separator-only title slots is the best of the tried heuristics. It matches Pi RPC evidence and simplifies downstream synthetic title-span replacement.

Synthetic generation smoke

Generated schema v2 synthetic augmentation with merge-all recipes:

data/schema_v2_synthetic_aug.mergeall.jsonl
data/schema_v2_synthetic_aug.mergeall.manifest.json

Validator passed. Manifest summary:

metric	value
`generated_rows`	49998
`numeric_title_rows`	30000
`path_rows`	19998
`path_series_rows`	15000
`path_movie_rows`	1666
`path_special_rows`	1666
`path_confuser_rows`	1666
`dropped_media_kind_mismatch`	19611

Commands

Pi RPC high-sim adjacent merge sample:

cargo run --release --manifest-path tools\rust_dmhy_template_apply\Cargo.toml -- `
  --review-title-boundaries-pi-rpc `
  --recipes reports\dmhy_template_recipes.full_top5000.mergeall.seed.jsonl `
  --title-boundary-decisions-output reports\dmhy_title_boundary_pi_rpc_mergeall.adjacent_highsim.jsonl `
  --title-boundary-heuristic-contains merge_adjacent_title_continuation `
  --title-boundary-min-similarity 0.85 `
  --title-boundary-max-similarity 1.0 `
  --threads 4 `
  --limit 30

Synthetic smoke:

cargo run --release --manifest-path tools\schema_v2_synthetic_augment\Cargo.toml --bin schema_v2_synthetic_augment -- `
  --recipes reports\dmhy_template_recipes.full_top5000.mergeall.seed.jsonl `
  --label-schema-file label_schema.json `
  --numeric-title-seeds data\synthetic_numeric_titles.txt `
  --path-prefix-seeds data\synthetic_path_prefixes.txt `
  --limit-templates 3000 `
  --max-rows 50000 `
  --output data\schema_v2_synthetic_aug.mergeall.jsonl `
  --manifest-output data\schema_v2_synthetic_aug.mergeall.manifest.json

cargo run --release --manifest-path tools\schema_v2_synthetic_augment\Cargo.toml --bin validate_synthetic_aug_jsonl -- `
  --input data\schema_v2_synthetic_aug.mergeall.jsonl `
  --manifest data\schema_v2_synthetic_aug.mergeall.manifest.json