[enaboapps]

Closes #29

Stacked on #28 (base claude/incremental-learning).

Learned counts were added to a copy of the baseline at 5×. That is negligible beside a baseline of millions of words, so a name or care term typed a few times hardly ever surfaced. A first version of this PR raised the fixed multiplier to 500. Review found real problems with that: one misspelling learned once became the top suggestion, heavy use would outgrow the baseline, and crafted counts could overflow. This version fixes all three:

• Learned counts are kept separate and combined when predicting. The arithmetic is floating point, so stored counts can't overflow when weighted (huge_stored_personal_counts_do_not_overflow).
• Bounded weight. One learned occurrence counts as 500 / (1 + learned words / 2000). Learned counts approach a bounded share of the baseline however much is learned.
• New and rare words need two uses. Learned counts of a word are used once it has been learned twice, unless the baseline has it at least three times. The production baseline holds over 9,000 words seen only once or twice, many of them misspellings, so a single misspelling is never boosted: learning "hav" once doesn't suggest it, and twice does. Words the baseline knows well count from their first use.
• Less memory. The baseline is no longer copied, so peak predictor memory falls from 362 MiB to 134 MiB, with or without a personal database.
• Speed. What prediction uses of the learned counts (usable counts, per-context totals and the weight) is updated in place when a sentence is learned, and rebuilt only on reload (updating_in_place_matches_rebuilding). Prediction and learning costs match Make personal learning incremental #28 within noise on the production baseline, measured by the ignored test prediction_cost_on_production_baseline and independently against Make personal learning incremental #28 by review:
  • Without learning: about 5 ms for an empty sentence, about 1 ms after a letter.
  • 2,000 new words learned: 9–15 ms per prediction, 17–23 ms per learn.
  • 40,000 new words (an extreme case): 30–40 ms per prediction, about 190 ms per learn, including a learn that makes a new word usable.
• Unchanged without learning. With no or an empty personal database, predictions are bit-identical (an_empty_personal_database_predicts_exactly_like_none). The schema and API are unchanged, and existing personal databases keep working. Make personal learning incremental #28's incremental learning now merges into the separate counts, and a reload reloads only the personal counts.

Results

These were measured on a synthetic person, neural/fixtures/personal.json, which is not user data. They learned 44 sentences, four of them containing one misspelling each. The 25 held-out sentences reuse their vocabulary: 445 queries, 78 of them on personal words. Measured on this branch:

Weighting	Top six	Personal words	Misspelling suggested
no learning	50.3%	1.3%	0%
500×, new words at once	61.8%	53.8%	0.2%
this PR	60.4%	46.2%	0%

General English was checked with that person's learning present, on 600 frozen-workload queries: 70.3% with no learning, 70.7% light, 70.2% heavy (×10). The differences are within noise.

Offering learned words to the neural generator as candidates was also tried. It added under half a point and nearly doubled generation latency, so it is left out. Details are in docs/personal-learning.md.

Validation

• cargo fmt --check, cargo clippy --locked --all-targets -D warnings and cargo test --locked pass.
• New tests cover: a misspelling learned once not being suggested while a new word used twice is; a well-known word counting at once while a rare one waits; an empty personal database predicting exactly like none; scores bit-identical to the interpolation formula without learning; huge stored counts; and the weight halving at saturation. The ignored prediction_cost_on_production_baseline test gives the timings above. The incremental-versus-reload, trigger, schema and concurrency tests from Make personal learning incremental #28 still pass.
• python -m unittest passes, and scripts/personal_evaluate.py runs end to end.

🤖 Generated with Claude Code