# What transfers from LevchykUGS

Studied the active tree and recent history through
`d3b23e6fe0199533224bfec45c4fec7243dffdc6` (2026-09-09). Paths and content hashes are
frozen in [sources.json](levchyk-study/sources.json). The older September 2 worktree alone
missed the most relevant September 7–9 fixes.

| Upstream lesson | What changes here |
|---|---|
| `hookd/SCHEMA.md`: raw evidence → entities → callable meta-function → filled scenario | The v2 contract has a signature, preconditions, complete slot sheets, timestamped invariants and stage exit conditions. |
| Meta-function critiques exposed plot-specific “universal” invariants | Two unrelated calls must fill the same slots. Transfer the belief transition; keep the family quest as one adapter. |
| `entities/dopamine.md`: preserve the attraction machinery during localization | Track anticipation, renewal, micro-payoffs, contrasting worlds, sensory inserts, specific expressions and callbacks. No invented neurochemical/retention measurements. |
| `75e892ba` / later reverse-sequence revision: observation, transfer and production are separate | Source seconds remain evidence. Accepted new narration sets the new edit clock; no fixed source-sized remake. |
| `3d90bd1d`, `50bae42f`: product problem-state passports and actual pixel review | Distinct identity, concern markers, first use, intermediate weeks and completely resolved visual endpoint; every visible person has an assignment. |
| `50bae42f`: old week-four endpoint still showed the problem | A “happy” face or a mildly improved neck cannot pass resolved. Preserve the old rejected endpoint for comparison. |
| `129164f8`: accepted identity images can be reused | Hash-bound accepted references survive unchanged revisions; a different state gets a different reference, not a different person. |
| `facial_performance.py`: situation-specific setup → reaction → settle | Directions include trigger, subtext, brow, gaze/eyelids and mouth for listeners too. No generic “expressive faces” checkbox. |
| `437e6282`: stress instructions/marks did not repair an earlier take | Written text is only the request. Actual audio must be heard and linked to exact word occurrences. |
| September 9 pronunciation contract / U02 blind comparison | Later one-word U+0301 trial succeeded where prompt-only instruction failed. Use accents in actual submitted text, contextual evidence for every word, and full recording checks; do not generalize one success to all Ukrainian. |
| `8bae9613`: Pages rejects oversized single assets | Host small playback segments and a small progressive fallback; validate every published file against the size limit. |

The useful failure is visible below. These are **original upstream illustrative reference
images**, copied byte-for-byte. They establish a reusable target-state example, not the
identity of Natalia in the historical film and not evidence of cream efficacy.

| Baseline: concern clearly present | Rejected endpoint: concern persists | Resolved endpoint: concern absent |
|---|---|---|
| ![Baseline](levchyk-study/baseline.jpg) | ![Rejected endpoint](levchyk-study/previous-endpoint.jpg) | ![Resolved endpoint](levchyk-study/resolved-endpoint.jpg) |

We adopt the visual endpoint rule and selective invalidation, but do not import song
generation, melody/tempo quotas, or upstream provider requirements for songs. We also
do not transplant unverified clinical claims, the source's character de-aging, or its
claims that a particular mechanism caused audience dopamine or sales.