lab: legato continuity — does observed same-string legato explain the tapping residual?

[PhysShell]

Question

After correct tap-hand attribution, does observed legato continuity explain the remaining fingering-model exactness residual?

Stage 2 of the technique-aware fingering oracle work. It starts from #200's post-#202 baseline and tests observed legato continuity before any hidden-technique inference.

Research tooling only (lab/); no production behavior changes.

Protocol

• Project imported legato spans onto note-to-note TechniqueEdges.
• Measure the corpus before changing the objective.
• Compare the fixed tap-aware baseline B with:
  • C1: hard same-string continuity across observed legato edges;
  • C2(k): soft cross-string penalty, with k ∈ {1,3,10} and k=3 preregistered as primary;
  • D1/D2: derived descending-legato open-string waiver on top of C1/C2.
• No weights are fitted on these results.
• Evidence claims use the preregistered leave-one-song-out stop rule.

Census

• 29,778 observed legato edges in the whole corpus.
• 99.6% stay on the same string, versus 58.6% of plain edges.
• In the tap slice, 99.4% of legato edges stay on the same string.
• Three quarters of edges out of a tapped note are legato; 96.8% of those descend.
• After fix(core): correct Guitar Pro tuplet and double-dot durations #202 the frozen slice is 226 tapped lines against 8,740 untapped lines.

Results (v1-fit, whole corpus)

stage	human path in optimum set	gap to matched baseline	agreement	ceiling
B tap-aware	0.9%	19.1 pt	43.1%	52.3%
C1 hard continuity	11.5%	9.3 pt	45.7%	51.4%
C2 soft, k=3	8.4%	12.0 pt	47.6%	53.3%
D1 = C1 + open-string waiver	15.0%	6.0 pt	52.1%	59.7%
D2 = C2(3) + waiver	8.8%	11.7 pt	45.7%	53.2%

Hard continuity halves the exactness gap and passes the registered leave-one-song-out rule in both format families and under both weight sets. The open-string waiver adds evidence on top of hard continuity on the whole slice and in GP3–5, but its gain is concentrated in three songs from one band and is only case evidence in GP6/7.

Important limitation / next question

Hard continuity makes 19 tapped lines (8.4%) unreachable: their human tabs contain 61 cross-string legato edges across 8 songs. Those are intentionally preserved as evidence rather than normalized away. They should be inspected before choosing a production-style hard rule, exception budget, or large soft penalty.

Reproduction

This branch was merged with the final #200 head after #202. On 794dbd1, the Stage 2 census JSON and phase-2 outputs (legato.json, legato-lines.jsonl, legato.md) reproduce the saved outputs byte-for-byte, and the lab tests are green.

Full protocol, ablations, leave-one-song-out analysis, format breakdowns, controls and limitations: docs/audit/2026-09-fingering-legato-continuity.md.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: a13771be-b3cf-4597-9e57-4ea8316e118f

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}