The Atelier
Pendraic's typed, composable, validated, reversible execution runtime for LLM-assisted writing. Every AI operation in the product flows through it.
The Atelier is not a ninth PENDRAIC layer. The acronym is load-bearing, Promise · Engine · Nexus · Dynamics · Registry · Anchors · Iterations · Closure, and stays at eight. The Atelier sits orthogonal as shared infrastructure. The eight layers hold truth; the Atelier is how truth gets created, refined, and verified.
The through-line
The AI is a proposer. The human is the actor. Nothing is real until ratified.
Operations produce Change proposals. Pipelines batch them. The Review Queue is where humans ratify. Autopilot Policy ratifies trivially-safe Changes automatically under user-configured rules. Every accepted Change is reversible for 30 days. This stance is non-negotiable for creative writing where the writer stakes their name on the output.
The six primitives
1 · Operation
A single LLM call with a typed contract: declared reads from the lattice, declared writes to the lattice, a parameterized prompt template, a model selection, a set of Wards, and optional Overlays. The atomic unit of LLM-assisted work.
2 · Pipeline
A directed acyclic graph of Operations. Pipelines fan out across inputs, fan in across candidates, pause at human gates, and can call other Pipelines.
3 · Binder
A typed I/O adapter between Operations and the lattice. Read Binders pull from addresses like Registry.StoryIndex.Character[elena].traits; write Binders target addresses like Registry.WorldIndex.Faiths[aethon].doctrine. Binders are what make Operations portable across formats, the same Operation works for epic fantasy, cozy mystery, or biography because the Binder resolves against the active format preset.
4 · Ward
A validator that runs between an Operation's LLM call and its commit. Wards check structural validity (output matches schema), semantic validity (no contradiction with existing Registry entries), and stylistic validity (prose passes Anchors / Voice Print conformance). Failed Wards trigger retry, escalation to user review, or outright rejection depending on policy.
5 · Overlay
A composable prompt modifier that wraps a base Operation. “Drunk Critic” persona is an Overlay. “Match the voice from Anchors” is an Overlay. “Ask clarifying questions before producing output” is an Overlay. Overlays stack: an Operation with voice + critique overlays is a different beast than the base, without duplicating the base prompt. Improving the Drunk Critic Overlay propagates to every Operation that uses it.
6 · Change
The proposal emitted by a write Binder. Not a direct commit, a queued record with before/after state, origin, Ward results, risk tier, and status. The single primitive shared by every consumer of the Atelier: Story Index bidirectional sync, Smart Import, Iterations passes, Penny tool calls, World Index auto-discovery. All emit Changes. All flow through the same Review Queue. All share one audit log. All respect the same autopilot policy. One queue, one audit log, one set of rules across every surface that touches your manuscript.
What the Atelier adds vs. doing it by hand
The closest comparable is Novelcrafter's Paint-by-Numbers community guide, a book of prompts that demonstrably produce professional-grade planning, drafting, and revision output. Its ceiling is that the user is the runtime: every prompt run by hand, every output unvalidated, every transition a copy-paste, every prompt improvement re-propagated manually.
| Capability | Manual prompts | Atelier |
|---|---|---|
| Typed I/O | free text | declared reads / writes with schemas |
| Validation | visual inspection | Wards run automatically, gate commit |
| Format portability | re-author per format | Binders resolve against active preset |
| Prompt variants | copy-paste + diverge | Overlays stack onto base Operations |
| Multi-model kitbashing | manual copy-paste | Ensembles run N-way; Merge synthesizes |
| Commit safety | paste into codex | Changes queue, Review Queue gates |
| Reversibility | undo-in-head | 30-day audit log, one-click revert |
| Observability | your memory | queryable telemetry on every run |
Who consumes the Atelier
The runtime is shared. Multiple subsystems run on top of it:
- Iterations passes, the pass catalog (Voice Pass, Continuity Pass, Line Pass, …) is a set of Atelier Pipelines.
- Story Index bidirectional sync, when a character's description changes mid-prose, cross-references update via Atelier Operations.
- Smart Import, pulling a character roster from existing prose runs as a family of Atelier Pipelines.
- Penny, every dispatcher tool Penny invokes is an Atelier Operation. There is no parallel AI backend.
- World Index auto-discovery, proposing entries from prose, promoting tags to constraints, all Atelier-mediated.
One Change schema, one Review Queue, one audit log, one autopilot policy engine across all consumers.
What you get today
- The six primitives as typed contracts, persisted to your project so Operations are repeatable and Changes are auditable across sessions.
- A workflow runtime that handles long, multi-step LLM chains with retries and checkpointing.
- The Change schema and Review Queue UI.
- An Autopilot Policy engine with hard rails. Promise, Engine, Nexus, and Anchors never auto-accept; nothing destructive to prose auto-accepts.
- Pendraic-authored Operations powering Story Index sync, Smart Import, and the Iterations pass catalog.
- Observability instrumentation so every Operation, Ward outcome, and Change acceptance is queryable.
The prose drafting loop bypasses the Change gate by design: when you ask the editor to continue a scene, output streams straight into the manuscript with normal undo, the same way any word processor works. The Change / Review machinery is for passes, audits, and bulk operations, not for the writer's own line-by-line drafting cadence.