IngenuityPrint — natural-language TC corpus, growing on the live menu-tree

Samsung Health (SM-S928N test rig, local Gemma) · every report is interactive — click a tree node to see the device screenshot of that screen. Latest: 2026-07-12 — contamination-curated HONEST render (render-only, NO re-record): the audit found 118/127 committed capture-logs were recorded before the last producer-fix, so the tree is curated for honesty WITHOUT changing the underlying data — 1 unknown act:? screen + 2 edges excluded; 5 nodes shown shot-less; 52/140 nodes badged verify-pending (recorded pre-settle-fix) and 186/206 edges dashed verify-pending (tool-call recorded pre-parser-fix, may be mis-classed); 83 nodes not settle-flagged, 20 edges post-parser-fix. 'Not flagged' ≠ independently verified; a full live re-record is the real fix and is pending (blocked on baseline orchestration). Gated by two 3-family merge-review rounds (edges + wording + frame-picker + fail-closed hardened).

22 → 140
screens (nodes)
24 → 206
actions (edges)
25 → 127
committed TCs
19
reports
25303540455055606570758085909510010511011512012513013514022262628303334373840435357262830333437384043535773262830333437384043535773742628303334373840435357737475262830333437384043535773747578262830333437384043535773747578818588922628303334373840435357737475788185889213214014006-3007-0107-0207-0307-0307-0307-0307-0407-0407-0707-12committed menu-tree nodes (each dot on 07-02 = one authoring wave)

2026-06-30

screen-identity v2 fold (dashboard 10→1)

2026-07-01

frontier live survey (whole-tree DFS)

2026-07-02

growth campaign — 3 batches + 6 loop rounds + identity v3 settings unblock

2026-07-03

round 7 — popup-node family sweep (v3 content-keying) + committed screen evidence (73/73 exact shots)

2026-07-04

rounds 11-13 + AU-1 — REPL :signal SHIPPED (3-family unanimous; add-tile-first + E1b re-record OPEN, SH-HOME-0031); classified rings: BP/HR detail, BP 120/80 value-input (plain set_text, :verify value golden x2 first live declared_value), OOBE Done/Next sweep (Spo2 onboarding through, BG wizard, Mood emotion) — every probed class needs NO custom tool; cross-route dedup proven (91/92 exact shots)

2026-07-07

SLEEP-CG — custom set_sleep_time composite tool SHIPPED: ONE NL line sets both bedtime + wake-up; the interaction-mode edge carries the tool-call (class=set_sleep_time + args), --from-feature codegen emits SleepTimeHelper.setSleepTime + assertNotNull(text "9:30"), and connectedAndroidTest passes on-device 1/1 (bedtime 9:30; wake-up is app-policy clock-derived). Byhand helper live-proof: first run set 9:00 not 9:30 (UiObject2.setText missed the reflowed minute) -> fixed via KEYCODE key events; the deterministic :verify caught it. Hardened by a 3-round cross-family review (R1 3x DO-NOT-MERGE incl. a vacuous-green structural hole -> R2 codex-only 5 residuals -> R3 unanimous SAFE; 11 findings fixed; a CI gate now enforces a paired time-shaped value-verify). + IB-5 vocab growth since E1b. Blackboard 132->140 nodes, 189->206 edges, 121->127 committed TCs.

2026-07-12

contamination-curated HONEST render (render-only, NO re-record): the audit found 118/127 committed capture-logs were recorded before the last producer-fix, so the tree is curated for honesty WITHOUT changing the underlying data — 1 unknown act:? screen + 2 edges excluded; 5 nodes shown shot-less; 52/140 nodes badged verify-pending (recorded pre-settle-fix) and 186/206 edges dashed verify-pending (tool-call recorded pre-parser-fix, may be mis-classed); 83 nodes not settle-flagged, 20 edges post-parser-fix. 'Not flagged' ≠ independently verified; a full live re-record is the real fix and is pending (blocked on baseline orchestration). Gated by two 3-family merge-review rounds (edges + wording + frame-picker + fail-closed hardened).