The Evolution of the Narracode Harness

Context & Catalysts: Prompted by GPT-6 Astra's research report (After the Double: What I Need from a Harness) after composing The Appointment — You.inc at Eldae. Astra analyzed frontier model capacities, context window expansions (1M+ tokens), and Boris Cherny's advice on reducing stale context files.

Architectural Breakthroughs:

• Unified STATE.md: Collapsed the eight mandatory structural memory files into a single, clean document with four distinct epistemic tiers: (1) Story Facts, (2) Author Decisions, (3) Model Interpretations, and (4) Unresolved Possibilities. Modular splitting into structural/ is retained as an optional scaling rule only when long-form complexity warrants it.
• Model- and IDE-Agnostic AUTO_MODE: Dropped the artificial and brittle hardcoding of specific model pairs (e.g. Opus vs Sonnet). Prompts run within specific IDEs (Claude Code, Antigravity, Codex); the harness now defines functional pass sequences that execute within whichever runtime is active, recording the exact participating model in ATTRIBUTION.md.
• On-Demand Editorial Inquiry: Replaced repetitive, automatic post-draft check spam with question-driven review called on demand ("What does this passage do? Where does it stop attending? What would a change cost?").
• Grounded Seamless Edits: Diffs against immutable saved files or Git objects, ending the fallible concept of diffing against "memory of prose."
• Self-Contained Specification: Condensed from ~3,734 words of bureaucratic check lists to ~1,500 words of lean, self-sufficient operational guidance that a fresh model can execute cleanly without prior exposure.

Context: Prompted by the realization that strict mechanical audits against 29 tell classes were producing synthetic flatness and over-correction. Replaced automated regex-like bans with humanistic editorial review.

Key Elements: Evaluated quoted spans in context with KEEP / CHANGE / UNCERTAIN; conditioned all remedies through a project's POETICS.md; bounded perturbation review to prevent obsessive iteration toward "zero tells."

Context: Gemini 3.5 Flash proposed a systematic registry of machine prose habits in AI Tells Update Plan, cataloging 29 classes ranging from unanchored sensory metaphors to pseudo-profound codas.

Key Elements: Introduced master_ai_tells.md, closed-loop tell audits, and perturbation mode to break statistical cadence.

Context: Packaged as the primary public distribution file narracode_Aug-9-2026.zip on the main index. Added an explicit FAQ.md and codified the Attribution Norm: requiring every AI model to be named by its exact version rather than generic vendor branding.

Context: Enabled continuous multi-act generation without stopping for manual human confirmation at each step, pairing Opus 4.7's architectural governance with Sonnet 4.6's drafting speed.

Context: Transformed implicit agent behaviors into a declared, auditable hook surface (on_init, pre_draft, on_draft, post_draft_check, on_snapshot). Elevated the post_draft silence into a first-class negative hook to prevent models from autonomously hallucinating subsequent scenes.

Context: Created by David Jhave Johnston and Claude Opus 4.7. Laid down the core neurosymbolic thesis that separating roles (Initiator, Reading, Structural, Compositional, Reflexive) and persisting memory to files elevates prose above median LLM training distributions.