Skip to content
NovelForge DocsReference

Corpus Policy · Govern evidence without turning source access into story or style authority

NovelForge uses corpus material to study mechanisms, test preference/craft hypotheses, build eval evidence and identify counterexamples. Corpus is an evidence domain. It is never Project Canon, character knowledge, a hid

NovelForge uses corpus material to study mechanisms, test preference/craft hypotheses, build eval evidence and identify counterexamples. Corpus is an evidence domain. It is never Project Canon, character knowledge, a hidden imitation prompt, or automatic Framework guidance.

Core invariant ✦ Access, rights, storage, analysis, learning and promotion are separate gates. Passing one gate never implies that the next gate has passed.


Corpus evidence may support:

  • project-specific craft analysis;
  • user-taste hypothesis testing;
  • cross-work General Craft research;
  • capability/regression eval design;
  • mechanism benchmarks;
  • counterexample/profile-boundary discovery;
  • external evidence for a bounded research question.

Corpus evidence may not by itself:

  • create or modify Project Canon;
  • prove that a character knows something;
  • settle relationship/resource/information state;
  • override explicit user/project authority;
  • activate durable user taste;
  • promote Framework behavior;
  • create a reusable named-author imitation profile.

02 · Rights class and storage intent are different fields

Section titled “02 · Rights class and storage intent are different fields”

Every source candidate that may enter durable Corpus handling needs a declared rights_class and storage_intent.

Reference rights classes:

Full-text storage may be permitted when there is a documented basis such as public-domain status, a compatible open license, explicit permission, or user-owned/user-authored material with permission to store.

A non-empty rights basis and provenance are required.

The material may be accessed/analyzed under the declared basis, but NovelForge must not store the full text as Corpus data.

Allowed storage may include:

  • source metadata;
  • derived observations/metrics;
  • mechanism analysis;
  • summaries;
  • a short excerpt only when genuinely required for the declared analysis/eval purpose.

Rights have not been established well enough for content storage. Only metadata-level storage is permitted until the evidence changes.

A private repository does not turn unknown rights into redistributable rights.


Section titled “03 · Deterministic Rights Gate is not legal analysis”

rights_gate.py validates whether declared metadata and requested storage intent are internally consistent with NovelForge policy.

It enforces, for example:

unknown + anything beyond metadata_only → reject
analysis_only + full_text → reject
short_excerpt without excerpt_purpose → reject
redistributable without rights_basis → reject

The validator does not infer copyright status from a URL, title, creator name or repository visibility. legal_analysis_performed = false is deliberate.

Rights/source status must be established from real evidence through the authorized research process.


A durable Corpus record should be able to answer:

corpus_id:
source_title:
creator:
source_url_or_ref:
source_type:
language:
publication_date:
rights_class:
rights_basis:
storage_intent:
accessed_at:
content_fingerprint:
analysis_scope:
research_question:
source_tool_or_capability:

Not every field must apply to every source class, but substantive claims must remain traceable to a real source/ref and the capability that retrieved it.

If provenance cannot be established, lower confidence, keep only safe metadata, or block the step. Never fabricate a quotation or source access event to complete a pipeline.


A discovery result says a candidate source was found. It does not mean NovelForge may copy or persist the source content.

discovery
→ verify source identity + provenance
→ establish declared rights basis
→ choose storage intent
→ deterministic rights gate
→ bounded ingestion / observation

The Corpus Scout and discovery runtime may prepare and normalize candidate evidence, but they cannot manufacture authorization that the host/source did not provide.


Do not analyze an entire work merely because it is available.

Start from a concrete research question, such as:

  • how does a fast scene create pressure without pseudo-speed fragmentation?
  • how is exposition made causal through an active task or conflict?
  • how does a supporting character maintain an independent agenda inside a protagonist-centered scene?
  • how does a chapter create forward pull without narrator advertising?

Select the minimum sufficient range/evidence required to answer the question.

Bounded analysis reduces context waste, imitation pressure, accidental source leakage and confirmation bias.


07 · Counterexamples are required for generalization

Section titled “07 · Counterexamples are required for generalization”

Corpus research should actively seek:

  • evidence supporting the candidate mechanism;
  • successful examples that violate the superficial pattern;
  • examples where the mechanism fails;
  • genre/platform/profile exceptions;
  • alternative explanations for the same observed effect.

A search that only retrieves examples agreeing with the current hypothesis is not strong General Craft evidence.

One work may create an observation. It does not create a universal rule.


NovelForge may analyze broad, transferable craft mechanisms such as:

  • scene causality and pressure sequencing;
  • information timing;
  • paragraph function;
  • dialogue embodiment;
  • character-agenda independence;
  • setup/payoff management;
  • broad genre/platform conventions.

It must not turn modern/living authors into reusable imitation fingerprints.

Do not create Framework behavior whose goal is:

  • “write exactly like Author X”;
  • reproduce signature phrases/cadence from copyrighted work;
  • preserve extensive copyrighted passages as style prompts;
  • optimize writer context for source imitation rather than mechanism understanding.

User-owned evidence and public-domain material remain subject to their actual rights/provenance and the same authority boundaries.


Raw Writer context should normally receive task-relevant mechanism/profile guidance, not bulk Corpus text.

Preferred path:

source evidence
→ rights-safe bounded observation
→ per-work mechanism analysis
→ counterexample / cross-work synthesis
→ benchmark / eval calibration
→ minimal relevant guidance
→ Writer

Regression bad examples and hidden eval answers stay outside Writer pre-draft context. Corpus/learning memory defaults to post-generation use unless a higher-level contract explicitly makes a particular bounded item writer-safe.


10 · Learning and promotion remain separate gates

Section titled “10 · Learning and promotion remain separate gates”

Corpus observations can support project, user_taste or general_craft learning scopes, but the Corpus layer cannot activate them.

General Craft normally requires:

  • multiple independent cross-work refs;
  • counterexample/profile-boundary evidence;
  • capability + regression evals;
  • provenance;
  • version/rollback evidence;
  • green Framework CI;
  • authorized promotion after prerequisites pass.

See the Self-Improvement Protocol.


If rights, provenance or analysis evidence later proves invalid:

mark source/item invalid
→ remove storage that is no longer permitted
→ identify dependent analyses / benchmarks / evals
→ invalidate or rebuild derived evidence
→ narrow / contest / deprecate dependent learning hypotheses
→ roll back affected promoted behavior when required
→ preserve correction provenance

Derived evidence must remain traceable enough for this dependency repair to be possible.


Corpus automation may:

  • detect an evidence gap;
  • prepare a discovery request;
  • normalize real returned source metadata;
  • run deterministic rights/storage checks;
  • package bounded evidence for semantic analysis;
  • record missing capability or blocked rights state.

It may not:

  • pretend Web/GitHub/MCP retrieval occurred without an eligible authorized capability;
  • infer legal rights from weak metadata;
  • fabricate quotations;
  • promote Corpus observations into Canon, user taste or Framework behavior.

A useful Corpus system remembers enough to support evidence and rollback, but never so much that source possession becomes a shortcut around rights, authority, or craft reasoning.