Skip to content

Getting Started (Science Journalists)

This page is a 20-minute read designed to bring a journalist with no prior exposure up to the point where they can hold a productive conversation with an FCC practitioner.

FCC in one paragraph

FCC stands for Find, Create, Critique. It is an open-source (MIT license) Python framework for building AI systems that do their work in three inspectable phases rather than one opaque call. Each step is handled by a "persona" — a role defined in a ten-field YAML specification called R.I.S.C.E.A.R. Deployments built on FCC produce structured logs (traces), risk-classified documentation (model cards), and regulator-oriented audit reports (compliance reports). The framework is owned and maintained by INFORMATION COLLECTIVE, LLC. It currently ships 164 personas and 10 specialized vertical vocabularies (constellation-verticals such as Ophiuchus for finance and Serpens for health).

The three phases, with analogy

  • Find: like a research assistant. Retrieves relevant context.
  • Create: like a draft writer. Produces an output from the context.
  • Critique: like an editor with veto power. Reviews the draft against rules and halts it if necessary.

The analogy matters because it gives readers a concrete way to reason about the system. It also highlights why FCC is auditable: the separation of duties is explicit, and the editor's halt is logged.

What R.I.S.C.E.A.R. stands for

Role, Input, Style, Constraints, Expected Output, Archetype, Responsibilities, Role Skills, Role Collaborators, Role Adoption Checklist. Ten fields. Every persona has all ten. For journalism purposes the most useful fields are:

  • Role (what the persona does)
  • Constraints (what it must not do)
  • Responsibilities (active duties during a run)

When you read a persona for a story, these three fields tell you what you need to quote.

Risk categories at a glance

FCC aligns with the EU AI Act's four-tier risk model: minimal, limited, high, unacceptable. Each persona carries one. A deployment's overall classification is the maximum of its personas' classifications. A system with one high-risk persona is a high-risk system, full stop.

This matters for copy because "high-risk AI system" is the exact language the regulation uses, and using it correctly is the simplest signal that your reporting is serious.

The artifact map

Five kinds of artifacts you may encounter and should know by name:

  1. Persona YAML — source file for a role.
  2. Scenario JSON — describes one run: which personas, what input, which workflow.
  3. Trace JSON — the log of what happened during a run.
  4. Model card — per-persona Markdown document modelled on Mitchell et al. 2019.
  5. Compliance report — JSON with findings against the EU AI Act (256+ requirements) and NIST AI RMF (29 subcategories).

Every one is machine-readable and shareable. Ask for them by name.

Sister-project landscape

FCC does not live alone. Its sister projects include:

  • POLARIS — policy ontology reasoning; supplies structured policy representations.
  • LYRA — a graph-backed UX layer.
  • Ten constellation-verticals — vocabulary and persona packs for specific domains (finance, health, law, legal records, insurance, metrology, logistics, energy, communications, environment).

When you cover a deployment, identify which sister projects are in play. A health-care deployment is probably FCC + Serpens (health vocabulary) + a specific hosted LLM.

What journalists often misunderstand

  • Version matters. FCC v1.0 and v1.7 are meaningfully different. The version appears in any artifact.
  • The framework vs. the deployment. You can criticize a deployment's choices without criticizing the framework, and vice versa.
  • Mock mode. FCC can run deterministically (no LLM called). Many student and research projects use mock mode. "The system ran" may not mean "an LLM was invoked."
  • Persona count is not system size. A deployment uses a small subset of the 164 personas. "164-persona system" is almost always wrong copy.

What to ask first

Three questions that get you 80% of the story:

  1. "Which FCC version are you on?"
  2. "Can I see the list of active personas, with their risk categories?"
  3. "When was the most recent ComplianceReport generated, and can I see it?"

If the answers are evasive, that is reportable.

Using this page in the field

Print or save this page. Bring it to interviews. It is under CC-equivalent license terms (MIT doc licensing) and you may quote any passage.

A one-hour learning plan

If you have an hour before an interview or a filing deadline:

Where next

Go to Reproducibility Primer to deepen your literacy in what "reproducible" means here, or to Citation Guide to get your bibliography right before filing.