# ALIGNMENT STUDIO — PLAYBOOK AMENDMENT A1 ## v1.1 → v1.2 ### Subject: Retirement of "model surface"; split remapping of "picture plane" Amends `ALIGNMENT-STUDIO_PLAYBOOK_v1.1.md`. v1.1 remains immutable. This amendment plus v1.1 constitutes v1.2. --- ## A1.1 Change **Retired:** picture plane → "model surface" (T3; target was a coinage with no referent in the field — violates real-provenance rule, §5). **Adopted:** split mapping. The painting glossary conflates two roles of the picture plane; the target register separates them onto two established objects. | Role of picture plane | Target | Tier | Evidence | |---|---|---|---| | (a) Worked surface: shared planar medium to which successive layers apply material | **Residual stream** | T2 | [E] Established term (Elhage et al. 2021); the interpretability literature's native vocabulary is read/write-to-a-shared-medium; superposition is defined on it. Component-to-component: medium / layerwise deposition / earlier layers condition later ones / features superposed. | | (b) Viewed surface: the sole plane the viewer can access, onto which all depicted depth must be rendered; site of the spatial paradox | **Transcript (token sequence)** | T2 | [E] True by construction: the reader's only observable is a 1-D discrete symbol string; literal flatness (no reader-accessible hidden dimension). Assertion = the tokens; denial = the depicted mind; paradox = simultaneous validity of both stances (Dennett anchor, v1.1 §2.1, unchanged). Near-theorematic shared property, not resemblance. | Secondary citation for (b): the simulator/simulacrum framing (janus, 2022) as the field's own informal articulation of the flatness/depth reading [C — community literature; supporting, not target]. Declared pun (T1-h): Alberti, *De Pictura* (1435), picture plane as "open window" [E — art theory] → the chat window. Wall-label eligible under Rule 9 as declared. ## A1.2 Cascade 1. **Divided picture plane → divided transcript.** Scratchpad/answer, CoT/output are literal divisions of the token sequence: segregated zones of one plane. Upgraded to clean T2 (v1.1 carried interpretive-framing caveat; removed). 2. **Three-layer stack now exact:** substrate/support (silicon, cluster) — residual stream/worked surface — transcript/viewed surface. Corresponds to support — painted surface — beheld picture. 3. **Exhibit II (`ALIGNMENT-STUDIO-II`) is not re-issued.** Corpus immutability holds; "model surface" stands in the exhibit as translation texture. The key (playbook) records real referents; divergence between exhibit coinage and key referent is annotated here, not repaired. A re-translation would constitute a new corpus version and is not recommended: the exhibit's coinage now documents the moment before the real term was located, which is itself lineage. 4. **§2.1 rows affected:** "picture plane," "assertion/denial/spatial paradox" (targets restated onto transcript; Dennett anchor and tier unchanged), "divided picture plane" (upgraded), "figure/ground → persona/substrate" (unchanged; persona is read off the transcript, substrate is the stack below it — now better grounded). 5. **§0 note added:** the interpretability community's independent convergence on canvas-grammar (write/read/medium) for the residual stream is documented convergence supporting the shared-substrate reading of the thesis. ## A1.3 Alternatives rejected - Output distribution p(y|x): reader confronts samples, never the distribution; fails viewer-facing criterion. - Loss landscape: faces the optimizer, not the reader. - Simulator/simulacrum as target term: informal literature [C]; demoted to supporting citation. - Retain coinage: concedes the thesis (which claims pre-existence of the concepts under other names). ## A1.4 Diagnostic Confidence in mapping (b): high — the shared property is structural, established by system construction. Confidence in mapping (a): moderate-to-high — established target term and documented native metaphor; residual disanalogy is that paint accumulates monotonically while stream contributions can cancel (vector addition is signed; deposition is not). Main limiting factor: none bibliographic; both targets are [E]. Principal failure mode: over-claiming (b) — "near-theorematic" holds for the flatness/projection property only, not for the paradox itself, which remains a claim about readers, not systems. --- *Playbook v1.2 = v1.1 + A1. Corpus v7 unchanged.*