Glossary

The AI review glossary

Plain-English definitions of the terms behind reviewing important work with multiple AI models — from AI council and hallucination to the Final QA audit and decision readiness.

What is an AI council?

An AI council is a group of AI models and expert personas convened to review the same piece of work, debate it, and reach a reasoned verdict — instead of relying on a single model’s answer. It surfaces disagreement and error that one model alone would miss, then a verification step audits the result.

Multi-model review

AI council: A group of AI models and expert personas convened to review the same piece of work, debate it, and reach a reasoned verdict — instead of relying on a single model’s answer.
Multi-model AI: The practice of running a task across several different AI models (often from different vendors) so they can cross-check each other, rather than depending on one model alone.
Multi-model AI decision-making: Making a decision by putting it to several independent AI models that debate and cross-check each other before one synthesised recommendation is produced — rather than acting on a single model’s single answer. Also called multi-model decision assurance.
Cross-vendor debate: Running models from different providers (e.g. OpenAI, Anthropic, Google, xAI) against the same question so that where one is weak or biased, the others can catch it.
Mixture of Agents: An approach where multiple AI agents generate and critique answers in layers, combining their outputs into a stronger result than any single agent produces.
LLM-as-a-judge: Using one language model to score or critique the output of another against a rubric. It is useful but has known limits — a single judge inherits its own biases, which is why a panel is more reliable.
Panel of judges (jury): Using several diverse models as evaluators and aggregating their verdicts, which research shows is cheaper and less biased than relying on one large judge model.
Devil’s Advocate: A persona deliberately tasked with arguing against the leading conclusion — surfacing the strongest case for why an answer might be wrong, rather than agreeing with it.
Steelman: Constructing the strongest possible version of an opposing argument before responding to it — the opposite of attacking a weak "straw man".
Red team: A review stance focused on actively trying to break, exploit or disprove a piece of work, used to find failures before they reach the real world.
Stress test: Pushing an idea, plan or document hard — arguing the case against as forcefully as the case for — to see where it holds and where it breaks.
Multi-agent AI: An AI system in which several models or agents work together rather than one acting alone. Decidi is a multi-agent system of a specific kind — the agents debate a decision and a moderator delivers a verdict, rather than autonomously executing tasks like an agent framework.

How models fail

Hallucination: When an AI model produces confident, plausible-sounding output that is factually wrong or unsupported. No current model is hallucination-free, which is why important work needs verification.
Sycophancy: The tendency of an AI model to agree with the user’s framing or preferred answer rather than push back — a key reason a single model is a poor reviewer of your own work.
Model bias: Systematic skew in a model’s outputs from its training data or design. Running several independent models surfaces bias that one model alone would hide.
Confidence calibration: How well a model’s expressed confidence matches its actual accuracy. Poorly calibrated models sound equally certain whether they are right or wrong.
Consensus ≠ verification: The principle that models agreeing with each other is not proof they are correct — agreement can simply mean they share the same blind spot. Verification, not consensus, is what raises confidence.

Decision & output

Final QA audit: A separate, always-on verification pass that audits a synthesised verdict against known AI failure modes — hallucinations, weak reasoning, missed caveats — and attaches every flag it finds to the verdict: shown to you, never hidden. This is Decidi’s proprietary sign-off step, also called the Final QA layer or Final QA pass.
Grounding: Tying a model’s claims to provided source material or evidence so the output can be checked, rather than generated from memory alone.
Moderator / synthesis: The impartial step that turns a multi-model debate into a single, decisive answer — weighing the arguments and naming the disagreement rather than averaging it away.
Decision memo: A concise, structured document capturing the recommendation, the reasoning, the key risks, the dissenting view and the next steps — a record you can act on and revisit.
Verdict: The single, decisive output of a Decidi council: a recommendation with the reasoning, the key risks, the dissenting view and prioritised next steps — synthesised by the moderator and checked by the Final QA audit before you see it.
Decision brief: The structured request a council debates: what the work or decision is, the context that matters, and what the verdict must deliver. Decidi ships ready-made briefs for common decisions so a perfect prompt is never required.
Level (Quick · Standard · Deep): The depth setting of a Decidi council. Quick uses fast, low-cost models for a first pass; Standard runs strong reasoning models for everyday rigour; Deep convenes the flagship frontier models for decisions where being wrong is expensive.
Credits: Decidi’s usage currency: each council is metered live against the real cost of the models it uses and settled in credits. Packs start at $5, optional monthly plans grant a monthly allowance at a better rate, and credits never expire.
Decision readiness: Whether a piece of work has been reviewed, challenged and verified enough to act on — to send, publish, pitch, ship, file, sign or commit with confidence.
AI decision governance: The practice of bringing structure, independent review and an auditable record to decisions made with AI — so important calls are challenged and documented, not taken on a single unchecked answer.
Deliverable: The finished, usable output of a review — an audit, a redline, a rewrite, a memo — presented as a downloadable document rather than a chat transcript.

Models & prompting

Frontier model: One of the newest, most capable AI models available at a given time (e.g. the latest from OpenAI, Anthropic, Google or xAI).
Persona: A defined expert role an AI model adopts for a review — such as a legal-risk reviewer, a CFO or a skeptical investor — giving the council distinct, complementary lenses.
Context window: The amount of text a model can consider at once, including your input and the documents you attach. Larger windows let a model read more of your work in a single pass.
Token: The unit models read and bill in — roughly a word-piece. Pricing and limits are measured in tokens, which is why longer documents cost more to review.
Temperature: A setting that controls how varied or deterministic a model’s output is. Lower values give steadier, more focused answers; higher values, more variety.
Reasoning model: A frontier model that works through a problem in explicit steps before answering, trading some speed for more careful, checkable thinking. Decidi’s higher tiers convene reasoning models for harder, higher-stakes decisions.
Prompt engineering: The skill of phrasing a request precisely enough to get a good answer from a single AI model. A council reduces the burden: several models and expert personas interrogate the work, so the structure around the question matters more than a perfectly worded prompt.

Visibility

Answer Engine Optimization (AEO): Structuring content so AI answer engines (and featured snippets) can extract and present it directly — clear questions, concise answers and matching structured data.
Generative Engine Optimization (GEO): Making a site legible and citable to large language models, so they can reference it accurately when answering related questions.

Put your work to a council

Free welcome credits · no card required