Critique results and the loop

An agent runs a critique by calling validate with its finished output and one rule or rule set. You get a pass or fail, the passage that broke any failed rule, and what is left to spend. If a rule fails, the agent rewrites that passage and calls again.

The call

validate takes the output and exactly one target.

validate inputs
Input	Required	What it does
`output`	Yes	The text to check. Send the whole revised output on every attempt.
`rule_id` or `ruleset_id`	One of the two	What to check against.
`version_id`	No	Which version to run. Without it, the active version runs. Your own company can name any version of its own rules, including one not yet active, to test it. Other companies get the active version.
`session_id`	No	Names the loop, so attempts are counted together. Without it, the MCP connection is the session.
`max_attempts`	No	Lowers the attempt cap for this session. It can never raise it.
`request_id`	No	Makes a retry safe. A repeat of a call that already returned a result, with the same output, returns that result and is not charged again.
`agent_id`	No	Identifies the agent in reports.
`tags`	No	Up to 10 tags of up to 50 characters, for comparing runs in reports, such as instructions:v2.
`agent_model`	No	The model the agent runs on, as it reports it, up to 100 characters. Reports group by it. Vindex does not check it.
`agent_label`	No	A name for the agent, up to 100 characters.
`output_ref`	No	Your own reference for the output, such as a ticket or document ID, up to 200 characters. It is how you match a result to an output later.

Values over these limits are refused before anything is charged, never cut short.

What comes back

A single rule

result: pass or fail.
On a fail, the passage that broke the rule, quoted. A rule that fails because something is missing has no passage to quote, and the result still names the rule that failed.
rule_id and version_id: the exact version that ran.
attempt and attempts_left: where the loop stands.
credits_left, and low_balance when the balance is low.
validation_id: quote it to support if you dispute a result.

A rule set

result for the whole set. The set passes only if every rule in it passes.
failed_rules: each failed rule with its rule_id, name, and version_id.
rules_run: every rule version that was judged, each with its own result.
ruleset_id and version_id, plus the same loop and credit fields as above.

A result is a pass or a fail. There is no score. The judgement is probabilistic, so the same output can occasionally get a different result on another run.

The loop

The agent calls validate with its draft.
If a rule failed, it rewrites the quoted passage and calls validate again, with the whole revised output, in the same session.
It stops when every rule passes, when it reaches the attempt cap, or when credits run out, and then tells the person what is still failing.

The agent should not resend unchanged output. It is charged like any other attempt.
An attempt counts calls in the same session against the same rule or rule set, starting from 1. A call that Vindex released because of an error is not an attempt.
The cap applies to each session and target. A call over the cap is refused before anything is held, so it costs nothing, and the agent is told the cap and the rules that failed last time.
When the agent stops, the person decides what to do: edit the output by hand, or change the rule.

What a critique costs

$0.03 per call, which is one critique credit. A rule set costs the same as a single rule.
Each attempt is a separate call and is charged.
The credit is held before the engine runs and charged when a result comes back. If Vindex fails after the hold, the hold is released and you are not charged.
Refusals happen before anything is held, so they cost nothing.
Listing and reading rules, and reading your balance, are free.
Credits belong to the company. Every member and agent spends the same balance.

Refusals

Vindex checks the call before it holds any credit and returns the first failure. Each refusal carries a code and names what was wrong. Something you cannot see returns NOT_FOUND, never “forbidden”.

Refusal codes
Code	Meaning
`NO_CREDENTIAL`	Not signed in. The result tells the agent how to sign in.
`MEMBER_INACTIVE`	The member is deactivated.
`NOT_FOUND`	The rule or rule set does not exist, or you cannot see it. This includes another company's rule that is in test.
`VERSION_NOT_AVAILABLE`	You named a version of another company's rule that is not its active version.
`RULE_NOT_READY`	The rule is a draft. Set it to test first.
`RULE_UNAVAILABLE`	The rule is disabled or archived, and the refusal says which. For a rule set that holds one, it names the rule.
`RULESET_ARCHIVED`	The rule set is archived. The refusal names its replacement, if one is set.
`NOT_APPROVED`	Your company only allows approved items, and this one is not on the approved list.
`OUTPUT_EMPTY`	There is no output to check.
`OUTPUT_TOO_LARGE`	The output is over the size limit.
`TAGS_INVALID`	More than 10 tags, or a tag over 50 characters.
`FIELD_TOO_LONG`	A field is over its limit.
`ATTEMPT_CAP_REACHED`	The session has used its attempts. Nothing is charged.
`TOO_MANY_IN_FLIGHT`	Your company has too many critiques running at once. Try again shortly.
`INSUFFICIENT_CREDITS`	The balance cannot cover a critique. The refusal gives the balance, the price, the names of your Company admins, and a link where an admin can buy credits. The MCP server never takes payment.
`REQUEST_ID_REUSED`	A request_id was reused with different output.

Two further outcomes happen after the hold: ENGINE_FAILED and TIMEOUT. Both say nothing was charged and whether to retry. A call returns within the MCP server’s 70-second limit.

If a result looks wrong

Contact support with the validation_id from the result and the output. Vindex finds the call, the exact rule versions that ran, the engine version, and the charge, and can re-run the output against the same versions. If the engine judged wrongly, support may return the credits.

See also rule sets, reporting, and the first-run guide.