002-public-record.md
docs/builds/002-public-record.mdThis is the retained record itself—not a chat summary. It is exposed because it materially documents the research, method, evidence, audit, production, planning or release history of The 100.
# Build 002 — Public Record ## Making The first prototype was a deterministic register adapter. The ceiling pass separated source truth, claim decomposition, transformation, verification, optional local generation, optional local translation, and provenance. A separate browser NLI evaluator was evaluated but deferred: adding an opaque runtime model/CDN solely for sophistication would weaken dependency governance and first-load performance. ## Method 1. Preserve a source-of-truth object before rewriting. 2. Make important anchors explicit: numbers, negation, uncertainty, named entities, locked facts, conditions and scope. 3. Keep generation and verification separate. 4. Treat a passing anchor audit as a warning system, not proof of semantic or cultural equivalence. 5. Make limited browser AI a progressive enhancement; first value must survive without it. 6. Export transformation provenance when a result may be reused. ## Evidence + limitations The anchor contract and preservation score are engineering/product heuristics, not a validated linguistic-equivalence metric. They detect specific classes of drift; they cannot prove two statements mean exactly the same thing. Browser-managed Prompt/Translator capability is technical infrastructure, not evidence that a transformation is correct. Consequential multilingual, legal, scientific, clinical and regulatory communication requires competent human review. ## Privacy/support Core deterministic adaptation runs client-side. Optional browser-managed AI/translation runs only after feature detection. Support/failure states must remain visible. Shareable state must not silently expose sensitive source text; URL-state behavior remains subject to release privacy acceptance.