Have the argument before you commit.

Disagreement is cheaper in private.

What happens in a debate.

1. You ask one clear question. The room argues exactly what you asked. Every voice is told to stay on the task: not to widen it, add requirements, or invent facts. A precise question gets a precise debate, so "Should we store customers' card details ourselves, or leave that to the payment provider?" works far better than "payments?"
2. The room is seated. You name the voices yourself, or Anvil picks a small room for your question. The Embers explains both.
3. The voices read up on your project. Before anyone speaks, Anvil looks up what your memory already holds about the question and gives every voice the same notes. They argue about your actual situation, not a generic one.
4. Opening positions. Each voice says where it stands, one at a time.
5. Cross-examination. In each round, every voice reads what the others said, goes after the weak points, concedes what's right and sharpens its own case. There are three rounds unless you choose otherwise. If every voice says it has nothing new to add, the debate stops early instead of going round in circles.
6. One answer. Either a neutral chair, who took no side, reads the whole argument and writes the strongest single answer, saying what each voice contributed and what genuinely remains unresolved. Or the voices must agree on an answer between themselves, and name plainly any question they couldn't settle. Every answer opens with a short Conclusion.

Every voice is built on purpose.

The voices in the Crucible are called Embers. An Ember isn't a model told to "act like an expert". It is a character written by hand: the stance it argues from is set down in so many words, so it pushes where it is meant to push and doesn't let things slide.

A stance, and optionally a skill

Build the room for the question

Your cast is final

Or let Anvil pick

Leave the voices blank and Anvil seats a room for your question. It starts small on purpose, with three debating voices:

• The Devil's Advocate, always. It runs the pre-mortem: assume this failed, now say exactly why.
• A First-Principles Thinker, always. It asks whether this is even the right question. No wording of your question can push these two out of the room.
• The specialist most relevant to your question, from a bench of about fifty working roles across a whole company. Each specialist argues from a temperament that suits the job: an engineer as a Systems Thinker, a finance voice as a Data-Driven Analyst, a security role as a Red Teamer.
• Regulation experts, when they genuinely apply. If your question touches rules such as GDPR, HIPAA or PCI-DSS, Anvil first checks which ones really apply and seats an expert for each, ahead of other specialists. The room grows to fit them, up to eight debating voices.
• The neutral chair, who writes the answer, unless you've chosen to have the voices settle it themselves.

The internet is not a nice place.

Most plans are reviewed as if every user will be patient, honest and paying attention. Real people are tired, take the default, follow whatever incentive is in front of them and find the shortcut you didn't intend. Some of them are actively looking for a way to abuse what you built. A review that only imagines good-faith users misses both.

The Crucible has voices for each of them.

A checklist no debate can skip

Every debate is also handed the same short list of questions, and the room must answer each one that applies:

• Does a law or regulation apply? Name each one, or say clearly that none does.
• How could this be attacked, abused or exploited, and by whom?
• Whose personal data is involved, and on what legal basis is it held?
• What legal liability, contract duty or intellectual-property exposure does this create?
• What does it cost to build and run, and is it worth it?
• Who could be harmed, and is this the right thing to do, not just the possible thing?

So a blind spot becomes a question the room has to answer, rather than something it catches only if someone happens to mention it.

An example

You're adding referral credits: invite a friend and you both get a month free. Seat a Behavioral Psychologist, Trust & Safety, the Threat-Actor Mindset and a CFO, and ask how it will be gamed. This is the room that raises people referring themselves from throwaway email addresses, organised credit farming, users pressuring their friends, and what all those free months cost if the scheme works too well.

Three ways to start a debate.

In the app

1. Click the ⚗ button at the top right. The Crucible opens on your past debates.
2. Click New debate and type your question under What should the room debate?
3. Choose a Mode. Auto lets Anvil pick the room. Manual lets you list the voices, separated by commas, for example devils-advocate, trust-safety, cfo. Names come from the list below.
4. Set Rounds, from 1 to 6. Three is the default.
5. Settle early on agreement is ticked to start with: the voices settle the answer between themselves. Untick it to have the neutral chair write the answer instead.
6. If you like, tick Cap the spend and set a limit. When the estimate is over it, Anvil holds the debate until you raise the limit or turn it off.
7. Click Preview plan. In Auto mode it shows the room Anvil would seat, how hard each voice will think, and a rough cost. Nothing runs and nothing is spent.
8. Click Run debate. It takes a few minutes, and the finished debate appears in the list.

From your assistant

From a terminal

See the room and a rough cost first. This spends nothing.

Run it with a room you chose, over two rounds

It runs on your own model.

Keep what you learn.

• Every debate is saved. It is stored encrypted on your computer and reads like a chat: each voice's statements in order, with times, and the conclusion at the end. In the app, the Crucible lists every debate and lets you search them by a word, a phrase or the name of a voice. In a terminal, use anvil crucible --history.
• Promote the answer to keep it in memory (--promote, or ask your assistant to). Only the conclusion becomes a memory, with a pointer back to the full debate. The transcript stays in your debate history, where you can read it, but it isn't pulled into your assistant's memory.
• Learn from it. --learn pulls out the general lessons, the rules of thumb rather than this one answer, so they can help in future work.
• Reuse a good room. Save the voices that worked and run them again by name.
• Clear what you don't need. Delete any single debate from the history with --forget.

Every voice you can name.

Use these names with --embers or in the app's Manual mode. Add +skill-name to fuse any of them with a skill.

Thinking styles

Working roles

Tell us.

In the app, open Settings › System and choose Report a bug, or run anvil report-bug "what happened". Reports include your account email if you are signed in to a paid plan, and never include your memories, prompts or code. For everything else Anvil does, see the Anvil docs; to install it, start with getting started.