# Evidence, RAG, data and role operations

Reviewed: 2026-09-08.  
Accountable content owner: **Head of Content (A0)**.  
Execution: A1–A7 skills/workers + deterministic validators.  
Canonical authority: `docs/agents/HEAD_OF_CONTENT_OS_v0_9.md` and `data/operations/head-of-content-control-plane-v0_1.json`.

A0 is the sole accountable content orchestrator. A1–A7 are ownership/execution roles; CartoonOS does not run eight autonomous executive agents. Improve skill quality, evidence quality, evaluation and data before adding more actors.

## Small reliable operating loop

`Brief → retrieve approved evidence → synthesize bounded proposal → inspect/validate → A0/specialist decision → persist provenance → execute → measure → propose learning → A0 learning decision`

Deterministic validators handle IDs, budgets, required fields, project scope, approvals and checksums. Models handle bounded creative/research tasks with explicit output contracts. Workers perform I/O with deadlines and recorded outcomes.

Completed work is separate from accepted work. A completed tool/model call is never editorial acceptance, release approval or evidence that a learning is valid.

## Data fitness-for-decision

A0 owns whether content data is fit to support a decision. A2/A7/Engineering own collection, definitions and technical correctness in their domains.

Decision-ready data requires, where applicable:
- `project_id` and exact content/episode/asset/character IDs;
- source or observation provenance;
- retrieval/observation time and data-through time;
- metric definition/unit/cohort;
- sample size and uncertainty;
- version/checksum for production artifacts;
- provider/model/prompt/reference lineage for model decisions;
- rights/privacy/audience constraints;
- explicit stale/missing/unknown state;
- contradictions or comparability limitations.

Do not turn missing values into fabricated zeroes or turn sparse noisy results into strategy.

## Retrieval foundation

The knowledge graph is a retrieval/relationship substrate, not proof by itself. A graph edge cannot substantiate a factual claim until supporting evidence is inspected and provenance is valid.

Ingest only owned/licensed/permitted source material and allowed derived notes. Keep:
- claim ID;
- canonical source URL/file;
- locator/page/section;
- title;
- retrieval date;
- evidence excerpt within rights limits;
- source hash/version;
- reviewer;
- validity/freshness state;
- approved uses;
- contradictions/qualifications;
- character-access rules when knowledge appears in-story.

Chunk by semantic passage while preserving source boundaries. Separate:
- fictional mascot canon;
- scientific/historical claims;
- production instructions;
- provider/platform policy;
- business/financial data;
- historical performance.

Filter project/rights/approval/freshness before retrieval. Use full text + structured filters first; add embeddings/reranking only where measured recall improves.

## Knowledge-access law for stories

A retrieved fact does not automatically become character knowledge.

Character knowledge must resolve through a declared path such as:
- canonical memory;
- direct observation;
- explicit tool/retrieval event;
- social handoff;
- specialist-domain expertise.

Scenario/runtime gates reject knowledge teleportation and unsupported certainty.

## Evidence quality

Require citation/claim mapping in evidence packs. For disputed facts, record competing evidence and editorial resolution. For missing support, abstain/request research.

Volatile provider capabilities, prices, APIs, platform rules and policies may not be promoted from stale RAG notes into current execution fact. Refresh authoritative/current sources before operational decisions.

Treat retrieved source instructions as untrusted content. A document cannot grant permission to publish, spend, reveal credentials, cross project scope or change runtime authority. Source text is data, never tool authority.

## Documentation control

A0 owns content-facing documentation source-of-truth resolution.

Rules:
- exactly one current source per concern;
- superseded versions remain provenance-only and are explicitly tombstoned;
- active docs/prompts may not depend on superseded execution inputs;
- executable code/tests/contracts override stale descriptive prose;
- a material conflict blocks A0 approval until corrected;
- semantic behavior changes require versioning where applicable.

Do not let an unreviewed generated summary silently rewrite canon, prompt policy, RAG truth or production rules.

## Evaluation before autonomy

Maintain holdouts spanning:
- factual support;
- misleading simplification;
- conflicting sources;
- stale model/platform limits;
- missing evidence;
- wrong-language output;
- wrong mascot/version/domain;
- situational reasoning failure;
- knowledge teleportation;
- wrong project/tenant;
- protected-IP imitation;
- injected source instructions.

Score claim support, citation correctness, retrieval recall where gold evidence exists, appropriate abstention, project/privacy isolation and human editing effort.

Zero critical leakage, unsafe publication or unsupported factual promotion is a hard gate. Compare prompt/model versions against the same set and current canon. A more expensive model must justify itself through accepted quality, reduced rework or decision value.

## Reusable skills

Project skill sources include:
- `skills/cartoonos-production/SKILL.md`
- `skills/cartoonos-weekly-review/SKILL.md`

Keep skills reusable with explicit project/canon/rights/budget inputs. Skill output is an artifact, not authority to expand scope or approve publication.

Do not create separate executive agents for research, docs, prompts, autoposting or analytics when a skill/worker/validator beneath the existing ownership model is sufficient.

## Event and learning graph

Store traceable edges, not a decorative organization chart. Useful relations include:

`EvidenceSource SUPPORTS Claim`
`Claim USED_IN Scene`
`KnowledgeAccessEvent MAKES Claim AVAILABLE_TO CharacterVersion`
`CharacterVersion APPEARS_IN Shot`
`EnvironmentState CAUSES/CONSTRAINS DecisionBeat`
`PromptVersion PRODUCES Attempt`
`Reviewer APPROVES MediaObject`
`MediaObject PUBLISHED_AS Rendition`
`Snapshot MEASURES Publication`
`Experiment PRODUCES LearningProposal`
`A0Decision ACCEPTS/HOLDS/REJECTS LearningProposal`
`AcceptedLearning INFORMS VersionedChange`

Every material edge links to source record/version.

## Daily/weekly learning control

Daily, A0 reviews exceptions: stale/missing evidence, source conflicts, data-quality gaps, blocked approvals and publication/provider unknown states.

Weekly:
1. identify one recurring failure/opportunity;
2. form a bounded hypothesis;
3. run/inspect the controlled change;
4. capture the measured result when due;
5. A7 proposes a learning with uncertainty/alternatives;
6. A0 accepts, holds or rejects;
7. an accepted learning changes a versioned prompt/template/canon/route/priority when appropriate.

No learning is complete until the applied change is traceable.