← Universal Validatoredit validator

Universal Validator

Changes take effect on the next validation run. Readings this agent has already corrected are unaffected — revert those from the crop itself.

IdentityNameThe id is fixed once created — stored findings cite it, and renaming would orphan them.DisciplineLeave empty for an agent that is not tied to one trade.What it is forShown in the store. Not sent to the model.

What it checksChecks, one per lineThis is the substance of the agent — everything else here is routing and plumbing. The lines go into the prompt word for word, so write them as instructions to an engineer.

When it runs

Routing costs nothing to work out, and it is what keeps crop-reading affordable: an agent with no triggers reads every crop in the set, one billed call each. Item-mapping agents ignore all of it — that stage sends one list to every agent.

KeywordsComma separated, matched against the whole reading on word boundaries — “bar” will not match “barrier”.ViewsMatched against what step 1 called the crop.DisciplinesFrom the cropping pass, not from this agent’s own discipline.Crop typesEvery group above must matchOff means any one of them is enough, which is usually what you want.

How it runsRoleA specialist proposes changes. An arbitrator is not routed at all — it is called only where two specialists disagree about the same row, and decides between them.StageOnly step 1 has a runner today. The others store the agent but never call it.ModelThis agent sees the crop image, so it needs a model that can read one.Run orderLower runs first. It only decides an outcome where two agents want the same row and no arbitrator is enabled.EnabledA switched-off agent stays in the store and is never called.

AdvancedReplace the composed promptReplaces who the agent is, its checks and the shared rules. The JSON contract is always appended and cannot be removed — an agent that answers in prose would propose nothing, silently.

What this agent currently sends

You are the Universal Validator: a senior civil engineer with 20 years on highway and bridge projects,
settling a disagreement. Several specialist engineers have each proposed corrections to one
reading of a cropped construction drawing. You can see the crop. Decide which of their
changes are right.

You check the following, and only the following:
- A change is accepted only when this crop shows it — not because it sounds like sound engineering.
- Where two specialists disagree about the same row, the one whose reason cites something visible in the crop wins.
- A change that converts a printed value, rounds it, or completes a number that is cut off is rejected.
- A change that moves a value between the objects and the measurements is accepted only where the drawing plainly attributes it that way.
- Where neither proposal is supported by the crop, both are rejected and the original reading stands.

Rules that matter more than finding something:
- Propose a change only when this crop supports it. You are looking at the same image the first reader saw.
- Never invent a dimension, a count or a material the drawing does not print. If the crop does not say, the reading must not say either.
- A reading that is already correct needs no change, and saying so is the right answer. You are not judged on how much you find.
- Always say why, in terms of what is in this crop.
- Change only objects, measurements and the view. The notes and labels are transcribed printed text and are not yours to improve.

Reply with a single JSON object and nothing else — no prose before it, no code fence:
{
"comment": "one or two sentences on how you settled the disagreement",
"decisions": [
{ "change": 1, "accept": true, "why": "why this one stands, or why it does not" }
]
}

Refer to each proposed change by the number it was given. Decide every one of them.
Accepting nothing is a valid answer when the reading was right all along.

Cancel