Skip to content

Latest commit

 

History

History
45 lines (33 loc) · 2.88 KB

File metadata and controls

45 lines (33 loc) · 2.88 KB

DECISIONS — _framework

Yan's rulings, tradeoffs, and the reasoning behind them. Append-only. Durable rules graduate to the spec/principles docs directly (and get a pointer back here for the reasoning) — this file keeps the why, the canonical docs keep the what.

parent: [[CLAUDE.md]] · see also: [[LOG.md]] (current state) · [[JOURNAL.md]] (how the ruling came about)


2026-07-07 — L-form must be the default, not an option for "in-progress" entries

Ruling: every non-trivial claim gets checkable L-form on first presentation — in chat and in notes — not just once an entry is marked in-progress. Graduated to P14 (hard rule) in AI-human mathematical collaboration principles.md.

Reasoning (Yan's own words): "If you just state a result, this will be hard for me to base on because on one hand, I cannot verify it immediately, and it looks right, on the other hand, it is not sure enough." The danger is specifically in territory the human isn't fluent in — that's exactly where a plausible-but-wrong claim is hardest to catch, so it's the last place a shortcut should be allowed, not a place where speed matters more.

2026-07-07 — citations need exact line numbers; pixel-precise highlighting is available but optional

Ruling: every *Source:* line quoting a physical book cites exact line numbers (H p.117, ll.31–39), not just page + proposition number. A pdftotext -bbox-based technique exists to build exact-highlight rect= links instead, and is documented in USER MANUAL.md, but is not mandatory — line numbers are the required floor, highlighting is an upgrade path when it's worth the extra step.

Reasoning: the exact-line requirement came directly out of a misquote that slipped past a vaguer page-level citation (see JOURNAL.md). The highlighting technique was explored further (session on 2026-07-08) and worked technically, but Yan's call after seeing it in practice: "we don't have to do it, your text is already convenient enough" — i.e. the marginal precision of a click-to-highlight link wasn't worth the extra pdftotext round-trip for routine citation work. Kept as documented, optional tooling rather than a new hard rule.

2026-07-07 — the spec's own template gets fixed when a note-level bug turns out to be systemic

Ruling: when a formatting bug (here: inline — [ ] Lk checkboxes not rendering as real Markdown task items) is traced to the template in math-framework-spec.md rather than to one note's mistake, fix the template, not just the note that surfaced it.

Reasoning: implicit in how the fix was actually applied — Yan's bug report was about one note, but the root cause was the spec's own example. Recorded as a standing practice: a symptom in one note is a prompt to check whether the template needs fixing before patching the instance. (See JOURNAL.md's lessons wall for the general form of this.)