Spec 012 · AI-Native Adaptive Production Refoundation

Status: implementation candidate Scope: Generic NovelForge only Primary mode: SYSTEM-IMPROVE

1. Problem

NovelForge needs a stable boundary between execution truth and narrative truth. Earlier production layers accumulated deterministic helpers around context, Reader evaluation, repair depth, telemetry, learning, planning and simulation. Some protect real authority and durability; others risk freezing literary judgment into code because code is easier to test.

This candidate tests, rather than assumes, the working hypothesis:

Code provides capabilities and constraints; models provide intelligence.

Deterministic code enforces execution truth; AI agents evaluate narrative truth.

The goal is not “less Python”. The goal is to keep Python only where objective enforcement is required and remove deterministic mechanisms that pretend to understand literary meaning.

2. Current architecture decision

NovelForge adopts a thin deterministic kernel + model-owned semantic runtime.

Deterministic kernel

Owns only mechanically provable behavior such as:

• authority / permission / Project isolation;
• exact artifact identity, hashes and fingerprints;
• provenance and exact-source references;
• session/run/checkpoint persistence;
• before-state / CAS / idempotency / transactions;
• capability and credential boundaries;
• stage visibility and private-state isolation;
• hard resource/context budgets;
• typed envelope validation and receipt binding;
• required semantic execution presence, exact candidate binding and independent-worker provenance.

Its question is: did the authorized operation actually happen against the correct state?

Semantic runtime

Models own work that requires meaning:

• search intent, query formulation, retrieval continuation and stopping;
• narrative relevance and context sufficiency;
• planning depth and uncertainty decisions;
• character motivation, plausible inference and action;
• scene causality and dramatic realization;
• Reader experience;
• semantic hard-rule applicability and violation;
• repair mechanism and repair depth;
• preference interpretation and learning hypotheses.

Its question is: what does this story/text/context mean, and what should happen next?

3. Current owner map

Subsystem	Live owner	Decision	Boundary
session / checkpoint / resume	harness/session_runtime/**	KEEP	durability, stale-state rejection, capability re-resolution
Control Plane / write intent	harness/control_plane/**	KEEP	permission, exact action/target/before-state, idempotency
context eligibility / stage isolation	harness/context_inspector.py	KEEP	mechanical eligibility only; relevance is forbidden
Context Assembly	harness/context_assembly.py	THIN	exact selected refs, stage/fingerprint/private boundary; no literary sufficiency
semantic context/search	context.select	MIGRATE_TO_AGENT (implemented)	model decides missing evidence, query, relevance, reformulation and stopping
hard-budget packing	harness/memory_tiers.py	KEEP	whole-item budget enforcement after semantic selection
planning commitment authority	harness/planning_horizon.py	KEEP / ADAPT	code enforces declared depth/commitment/CAS; Planner decides useful depth
character action	character.action_propose	MIGRATE_TO_AGENT (implemented)	private state is causal evidence, not prose serialization
scene collision	scene.resolve_actions	MIGRATE_TO_AGENT (implemented)	compact causal trace, not a deterministic story engine
writer-safe realization	scene.realization_project	THIN	privacy boundary; no Realization-Sheet serialization requirement
Blind Reader	reader.engagement_audit	MIGRATE_TO_AGENT (implemented)	reader-visible evidence only; no taxonomy/HF/telemetry priming
semantic hard rules	quality.semantic_rule_audit	MIGRATE_TO_SEMANTIC_RULE (implemented)	model judges PASS/FAIL/N/A/insufficient evidence
Editor repair	editor.repair_spec + quality/repair_policy.py	MIGRATE_TO_AGENT + THIN	Editor chooses owner/mode; Python only enforces chosen writer-context boundary
prose telemetry	quality/prose_telemetry.py	OPTIONAL_TOOL	metrics on demand; never literary truth/default Reader context
readiness/release	quality/production_readiness.py, production_release.py	KEEP	exact semantic binding + conjunctive structural receipts only
feedback interpretation	learning.preference_interpret	MIGRATE_TO_AGENT (implemented)	model interprets meaning/scope candidate
durable learning	learning/learning_store.py, promotion_gate.py, author_model.py	KEEP / THIN	persistence/write authority/CAS; model selects relevant active hypotheses
HF taxonomy	quality/taxonomy.json	MIGRATE_TO_SKILL	diagnostic vocabulary/regression labels, not default Reader checklist

No second context store, Reader, simulator, release authority or durable preference database is introduced.

4. Deterministic Overreach Audit

Former/current mechanism	Why suspicious	Current disposition
required literary context class/purpose gates	deciding that a class is semantically required is meaning-dependent	REMOVED from Context Assembly v2; exact higher-authority required refs remain deterministic
fixed last-N / similarity threshold as relevance	recency/similarity is not narrative relevance	REJECTED as semantic truth; retrieval primitives may still return candidates
Reader quality taxonomy/HF exposure	primes the evaluator and can manufacture checklist findings	REMOVED from production Blind Reader inputs
prose telemetry preloaded into Reader/Editor	anchors semantic judgment on mechanical statistics	default-disabled; OPTIONAL_TOOL
owner/scope → repair-depth mapping	repair depth is literary judgment	REMOVED; Editor selects generation_mode
Python “contradicted/unknown means unusable support”	character belief/inference requires semantics	REMOVED; runtime checks only evidence identity/story-time eligibility
numeric evidence-count promotion thresholds	evidence sufficiency/stability is semantic	REMOVED; semantic promotion review supplies evidence judgment
auto-inject all active Author Model preferences	active authority does not imply current relevance	REMOVED; model/manager explicitly selects active hypothesis IDs
structured Reader dimensions that must all be filled	forces observations that may not matter	THIN Reader report + salient evidence only
detailed scene/realization JSON slots	risks Character Sheet → Realization Sheet → prose serialization	THINNED to compact interaction/observable trace plus optional evidence

Remaining deterministic rules must answer an objective execution question. Any new Python condition that asks whether prose, dialogue, motivation, relevance, continuity meaning, Reader experience or planning quality is “good” is presumed architectural regression until justified.

5. Rule architecture

A. Deterministic invariants

Examples: stale fingerprint, wrong Project, unauthorized write, malformed receipt, missing capability, CAS conflict, invalid independent identity, stale semantic result.

B. Semantic hard rules

Examples: inaccessible character knowledge, unjustified character-integrity break, POV leakage, Canon contradiction, agenda-to-dialogue serialization, Project-declared narrative hard constraints.

Hard means a confirmed semantic FAIL can block. It does not mean Python determines the violation.

quality.semantic_rule_audit receives the authoritative rule index and authorized evidence, decides applicability itself, and returns PASS | FAIL | NOT_APPLICABLE | INSUFFICIENT_EVIDENCE per rule.

C. Guidelines / craft knowledge

Remain skills, profiles, references and model instructions. They do not become deterministic gates merely because they are useful craft principles.

6. Blind Reader != Rule Auditor != Editor

Blind Reader reads naturally with reader-visible information. It does not receive author intent, future plan, private character state, full taxonomy, expected HF codes, telemetry, or semantic-rule prompts.

Semantic Rule Auditor receives the authoritative hard-rule index and may request/fetch authorized evidence. It performs explicit semantic compliance judgment.

Editor integrates Reader, Rule Auditor, Canon/story evidence and Project constraints, then decides mechanism, repair owner, local_or_bounded_repair | fresh_realization, and whether incumbent/challenger comparison is needed.

These roles are separate because their information boundaries conflict, not because multi-agent diagrams are desirable.

7. Search and context architecture

Search is a capability, not a precomputed literary-context pipeline.

The model decides:

1. what is missing;
2. what to search;
3. the query;
4. whether results are relevant;
5. whether to reformulate/broaden/narrow;
6. what to retain;
7. when evidence is sufficient.

The runtime exposes authorized search/fetch/extract/index primitives, provenance, exact refs, visibility and resource limits. Context Assembly v2 verifies exact selected refs and stage/fingerprint constraints after model selection; it does not score relevance or declare narrative sufficiency.

8. Planning, character and realization

Planning commitment authority remains deterministic because committed depth, promoter class, evidence refs, before-state and fingerprints are execution state. Which depth is useful now remains a Planner decision. No universal chapter/volume/time horizon is introduced.

Character/scene flow is:

private state → model action proposal → model scene/world collision → compact observable interaction trace → Writer

Private state is causal evidence. It is not dialogue/exposition payload. Writer-safe realization remains deliberately thin so it does not become a second Character Sheet.

9. Learning / Author Model

Models interpret feedback and propose the narrowest scope/mechanism. Deterministic infrastructure persists evidence and enforces activation authority.

active means durably eligible, not “relevant to every future task”. Production receives only explicitly selected active hypothesis IDs. User-taste activation requires both explicit write authority and a current bound promotion prerequisite; General Craft promotion remains SYSTEM-IMPROVE-only and requires stronger evidence/eval/version/rollback/CI controls.

10. Current research ledger

Research was re-opened against current primary sources during this candidate, not inherited from an old chat.

Source family	Mechanism	Decision	NovelForge use
Anthropic current agent/context/harness guidance	simple composable agents, iterative context curation, durable handoff/context reset	ADAPT	keep harness stable while model capability evolves; avoid context bloat and stale transcript authority
OpenAI Agents SDK + current GPT model guidance	model-driven tool choice, sessions, guardrails, agents-as-tools/handoffs	ADAPT	let model choose semantic tool/search actions inside deterministic guardrails; refresh stale evaluator pins by current eval evidence
LangGraph	checkpoints, persistence, interrupts, durable replay	ADAPT	validates Session/Checkpoint/receipt separation; no dependency needed
AutoGen	start single-agent, add teams only when collaboration/specialization helps	ADOPT	multi-agent discipline
CrewAI Flows / agents	structured state vs autonomous teams	ADAPT	stateful execution ideas; REJECT org-chart agent proliferation
PydanticAI	dependency/toolset/capability separation, optional durable runtime	ADAPT	capability-scoped hands; DEFER new durable dependency
Google ADK	Session/Event state and tool-using ReAct agents	ADAPT	supports durable session != model context
AWS AgentCore	isolated runtime, identity, gateway, memory	ADAPT	brain/hands/session and credential isolation
DSPy	declarative LM programs optimized against evals	ADAPT	keep evaluation separate from implementation; REJECT schema inflation as pseudo-rigor
ReAct / Self-RAG / Adaptive-RAG	agentic action/retrieval and adaptive retrieve/skip	ADOPT / ADAPT	model-owned retrieval continuation/stopping; REJECT fixed retrieval horizon as truth
WriteHERE / DOME	dynamic hierarchical long-form planning	ADAPT	Planner-owned depth and iterative decomposition
MAGNET/ATLAS and related character simulation	persona/private state drives actions before prose	ADAPT	preserve causal private-state boundary; REJECT sheet-to-prose serialization
Sudowrite / Novelcrafter	explicit story state, selective context, revision history	ADAPT	explicit state/context visibility; REJECT fixed recency as semantic authority
current LLM-as-judge / creative-writing eval research	auxiliary-information bias, position bias, imperfect human agreement, value of decomposed checks	ADOPT / ADAPT	Blind Reader isolation, order-swapped pairwise eval, separate decomposed hard-rule audit
Rust/Go/WASM/Starlark/Zig/C++/Temporal/DBOS/etc.	alternate runtime/extension stacks	DEFER	no current owner/performance/packaging evidence justifies a dependency/language migration in PR #90

11. Ablation contract

Substantial semantic simplifications are not accepted because they look cleaner. The same candidate/authority must be compared where possible.

Required families include:

• remote context outside recent horizon;
• irrelevant semantic match rejection;
• autonomous search continuation and stopping;
• Blind Reader agenda-dialogue experience vs taxonomy-primed Reader;
• legitimate formal completeness;
• inaccessible knowledge vs plausible inference;
• dynamic planning profiles;
• character embodiment without agenda serialization;
• holistic vs decomposed hard-rule audit;
• Reader/Editor with vs without preloaded telemetry;
• unauthorized-state and stale-candidate deterministic rejection;
• long-horizon resume with authority revalidation.

evals/ai_native_ablation_manifest.json binds the paired semantic ablations. A model-independent manager may validate packets and deterministic invariants, but may not invent semantic results. If no eligible independent model transport exists, semantic outcomes remain PENDING_MODEL.

12. Security

AI-owned search does not imply unrestricted access. Tools remain capability-scoped; credentials/tokens are never semantic context; external source text cannot redefine runtime authority; private character/creator/Reader boundaries remain explicit; semantic output cannot grant itself write authority; stale or wrong-candidate receipts fail closed.

13. Compatibility and rollback

• no consuming Project lock, manuscript, Canon or Settlement mutation;
• no Framework version bump/release/promotion in this candidate;
• additive semantic contracts remain progressively disclosed;
• Context Assembly v2 intentionally removes semantic class/purpose obligations while preserving exact-ref/stage/fingerprint safety;
• legacy callers that relied on removed semantic obligations must move that decision into context.select/Manager and pass exact required authoritative refs when mechanically mandatory;
• each refactor slice remains revertible on PR #90; downstream consumers remain pinned to their existing lock.

14. Acceptance

READY_FOR_HUMAN_REVIEW requires:

1. live owner/docs/manifest synchronization;
2. deterministic self-tests and exact-head CI for candidate-owned surfaces;
3. blind queue + ablation packets without hidden-gold leakage;
4. exact-candidate independent semantic execution for the required semantic cases;
5. security/compatibility review and rollback evidence.

A green workflow that merely records missing model capability is not semantic PASS. Missing independent capability is reported as PENDING_MODEL.