dev(ml): Adisankar's follow-up answers (F1–F3) — archive, dispositions, verification

[mmcky]

Adisankar returned the follow-up round (F1–F3, the re-posed A5–A7 case-grammar questions) on 2026-08-06 — 3 of 3 answered. This archives the reply byte-exact with parsed JSON and appends the F-round section to DISPOSITIONS.md: provenance, dispositions, verification against live output, and a watch list. With this, every question either round asked is resolved; Phase 2 waits solely on his inline review of lecture-python-programming.ml PRs 1–5.

F1 (was A5) — rule change + live defect. Both earlier renderings were wrong: "the cell with the flashing cursor" identifies which cell, and the correct form is the relative participle flashing cursor ഉള്ള cell-ൽ. The post-packet regeneration still renders accompaniment (cursor-നോടൊപ്പം) under the current rules, so this is a reproducible gap, not a historical one. His full sentence joins the example bank and the Phase 3 judge ground truth; the fix rides v0.3 with the PR-review flags — no standalone release.

F2 (was A6) — accepted as-is + narrow rule. -നെക്കാൾ is a valid colloquial variant of the reference's -നെ അപേക്ഷിച്ച്; the one hard instruction is that കുറച്ചു drops alongside it. All five delivered PR diffs scanned at paragraph level: zero instances of the banned stack, so the batch under review is already compliant. The കൂടുതൽ-as-quantifier structure (present in PR 4) is outside his ruling and deliberately watch-listed for the organic PR round rather than ruled on.

F3 (was A7) — accepted as-is. Both forms validated (-ൽ നിന്നും for "selecting from", list-ന് താഴെ for "below"), and the regeneration renders the sentence near-verbatim as his second natural form — confirming the packet had paired two different grammatical slots. No tool defect ever existed here.

Provenance is clean (valid UTF-8, ZWJ/ZWNJ 0/0, +198 Malayalam codepoints, 3/3 boxes) with one recorded oddity: a line of our covering prose came back reworded outside the answer boxes, so "returned exactly as sent" is now checked per round rather than assumed.

Refs #189

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Archives Adisankar’s Malayalam follow-up round (F1–F3) in experiments/ml-benchmark/, including the returned Markdown, a parsed JSON extraction, and an appended dispositions/verification section so the ml benchmark tracking record is complete and auditable.

Changes:

• Add the byte-preserved follow-up Q/A Markdown file for F1–F3.
• Add a parsed JSON representation of the three answers plus zero-width character stats.
• Extend DISPOSITIONS.md with provenance, dispositions, verification against regenerated output, and a small watch list.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
experiments/ml-benchmark/malayalam-review-followup-answers-adisankar.md	Adds the returned follow-up file with filled F1–F3 answer boxes (archive source of truth).
experiments/ml-benchmark/malayalam-review-followup-answers-adisankar.json	Adds machine-readable extraction of F1–F3 answers and integrity metadata for downstream checks.
experiments/ml-benchmark/DISPOSITIONS.md	Documents follow-up round status, provenance checks, rulings → dispositions, and verification outcomes.

Suppressed comments (1)

experiments/ml-benchmark/malayalam-review-followup-answers-adisankar.md:110

• Trailing whitespace at end of the F3 answer block line is likely unintentional and can cause noisy diffs; please trim it.

Both are acceptable. 

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.