validation-growth: 42 blind-written Validation items and scenes (35 per language)

weksa commented

2026-11-13 05:22:36 +00:00

Member

Grows data/original Validation from 14 to 35 per language: 42 blind-written items (item:vgrow:1..21, item:vgrow-ja:1..21), each its own new group whose name was chosen and bucketed (SplitScheme::V1, 40..=79) before any text existed; 42 keeper-to-unnamed scenes at inventory rev 4 by a separate annotator. Live Train+Validation 56 to 98; Train 14 and Test 8 per language unchanged.

Writer: six blind Claude writer agents (Anthropic recorded). No cross-family generator was reachable, so the self-preference effect (Panickssery et al. 2024) was NOT measured. Blindness is by brief content, not sandbox.

Changed: data/original/corpus.cc, data/original/speaker-state.cc, crates/weksa/tests/original_corpus.rs (pins, shared file-level leaks fn, new test grown_items_hold_no_test_text_and_every_group_buckets_validation with DIGEST_OF_THE_42), the map's Validation growth section. Nothing in src, entity, inventory, ontology.

Expect encode refusals on new items (written free of the inventory; question validation-growth-vocabulary stays open). Every earlier calibration number is for the 14-item corpus; recalibration is a separate step and is not run here. Not merged.

Generated with Claude Code

Grows data/original Validation from 14 to 35 per language: 42 blind-written items (item:vgrow:1..21, item:vgrow-ja:1..21), each its own new group whose name was chosen and bucketed (SplitScheme::V1, 40..=79) before any text existed; 42 keeper-to-unnamed scenes at inventory rev 4 by a separate annotator. Live Train+Validation 56 to 98; Train 14 and Test 8 per language unchanged. Writer: six blind Claude writer agents (Anthropic recorded). No cross-family generator was reachable, so the self-preference effect (Panickssery et al. 2024) was NOT measured. Blindness is by brief content, not sandbox. Changed: data/original/corpus.cc, data/original/speaker-state.cc, crates/weksa/tests/original_corpus.rs (pins, shared file-level leaks fn, new test grown_items_hold_no_test_text_and_every_group_buckets_validation with DIGEST_OF_THE_42), the map's Validation growth section. Nothing in src, entity, inventory, ontology. Expect encode refusals on new items (written free of the inventory; question validation-growth-vocabulary stays open). Every earlier calibration number is for the 14-item corpus; recalibration is a separate step and is not run here. Not merged. Generated with [Claude Code](https://claude.com/claude-code)

weksa added 4 commits

2026-11-13 05:22:36 +00:00

corpus: 42 blind-written Validation items, vgrow and vgrow-ja, 21 per language 05ef79af6a

Names bucketed before text (SplitScheme::V1, 40..=79); written by six blind
Claude writer agents (Anthropic fallback; no cross-family writer reachable).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

speaker-state: 42 model-annotated keeper scenes for the vgrow groups at inventory rev 4 0d2a51014c

Annotator A committed; annotator B only measured agreement.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

original_corpus tests: pins for 114 items, 98 live; shared leaks fn; grown-items test with DIGEST_OF_THE_42 891e4073e1

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

map: validation growth is built; recalibration is a separate step fb6f21c84f

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

weksa merged commit 44b66f94dc into main

2026-11-13 05:29:02 +00:00

weksa referenced this pull request from a commit

2026-11-13 05:29:03 +00:00

validation-growth: 42 blind-written Validation items and scenes (35 per language) (#75)

Rows
Columns

validation-growth: 42 blind-written Validation items and scenes (35 per language) #75