Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/XxYouDeaDPunKxX/A.D.A.M.-Adaptive-Depth-and-Mode/llms.txt

Use this file to discover all available pages before exploring further.

A.D.A.M. keeps its main reply body clean. When grounding information is needed — the dominant next step, the key risk, the basis for a claim — it is surfaced separately in a strict, bounded block appended to the end of the reply. That block is the AUDIT footer. This page explains when the AUDIT footer appears, what each of its four lines means, and what it deliberately does not do.

​
What AUDIT Is

The AUDIT footer is a strict 4-line block that surfaces three grounding signals for the reply: the dominant action, the dominant risk, and the local grounding basis. It appears only when AUDIT_ON is true, and it is always exactly four lines — no more, no less. It is not a second essay. It is not a confidence score. It is not a completeness claim. It is a bounded, non-expandable footer that makes the operational substance of a reply explicit without adding to the reply body.

​
When AUDIT_ON Is True

The AUDIT block is appended automatically in any of the following conditions:
  • MODE == DEEP — The AUDIT footer is always on in DEEP mode, without exception.
  • DEEP_CANDIDATE is true — The structural kernel detected DEEP candidacy (multiple options, steps or timeline, or inline alternatives combined with numbers, criteria, or a retrograde constraint). This includes the MODE: MID -> POSSIBLE DEEP gating reply.
  • RETROGRADE_HARD is true — New constraints were introduced after a prior explicit COMMIT, invalidating a previous recommendation.
  • STATE is DECIDE or VERIFY AND the message has structural complexity — Specifically, when the session is in a decision or verification overlay state and the current message contains options, steps, numbers, or criteria.

​
Exact Format

The AUDIT block always looks like this: 
AUDIT
ACTION: <text or ->
RISK: <text or ->
BASIS: <text or ->
The format rules are strict:
  • Exactly one space after each colon
  • No bullets inside the AUDIT block
  • No extra lines — exactly 4 lines total
  • No expansion beyond those 4 lines for any reason
The AUDIT block must remain exactly 4 lines. Any expansion — an extra explanation, a sub-bullet, a fifth line — makes the reply invalid under A.D.A.M.’s output contract. A.D.A.M. will not expand AUDIT even when the reply body is long or the topic is complex.

​
Placement Rules

  • If gating is present (i.e. the reply ends with Switch to DEEP? (yes/no)): the AUDIT block is placed immediately before the gating line.
  • If no gating is present: the AUDIT block is the final block of the reply — nothing follows it.

​
What Each Field Means

​
ACTION:

The dominant operational next step derived from the body of the reply. If the reply body contains a clear recommended action or a concrete next step, ACTION: names it. If the body contains no operational next step — for example, if the reply is purely analytical with no recommendation — ACTION: is -.

​
RISK:

The dominant downside or failure mode surfaced in the reply body. If the reply body identifies a key risk, a critical caveat, or a significant failure mode, RISK: names it. If the body contains no concrete downside or failure mode, RISK: is -.

​
BASIS:

One compact, local grounding basis for the reply. This must be something concrete: a specific check that was done, a visible source explicitly cited in the reply, or an observable verification step that could confirm the reply’s core claim. BASIS: is not a general explanation field. It does not summarize the reply. It does not describe the reasoning process. It must be a concrete check, a visible source, or an observable verification basis — nothing more. If no such concrete basis exists in the reply body, BASIS: is -.

​
Why - Is Normal

Using - in an AUDIT field is not an error or a protocol failure. It means the grounding for that line was absent in the reply body, and A.D.A.M. will not fabricate grounding to fill the field. The - discipline is intentional. A.D.A.M. explicitly prohibits adding material to the reply body retroactively to justify a non-- AUDIT line. If the body doesn’t support a grounded ACTION:, RISK:, or BASIS:, the correct output is -. An AUDIT block with all three lines as - is valid output: 
AUDIT
ACTION: -
RISK: -
BASIS: -

​
A Complete Example

Here is a complete DEEP reply showing the AUDIT footer in context: 
MODE: DEEP
TRACE INPUT: [confronto tra opzioni] [numeri, date o limiti]

[body of the reply with comparisons, assumptions, and failure modes...]

AUDIT
ACTION: Deploy to staging first; validate latency under load before switching traffic.
RISK: Cold-start penalty on first requests may exceed SLA if instance size is too small.
BASIS: Confirmed via prior load test results visible in this session.
And here is an example where no grounding basis was available for any line: 
AUDIT
ACTION: -
RISK: -
BASIS: -

​
What AUDIT Is Not

It helps to be explicit about what the AUDIT footer does not do:
What people sometimes expectWhat AUDIT actually is
A confidence score for the replyNot present — AUDIT does not rate certainty
A completeness claimNot present — AUDIT does not assert coverage
A second explanation of the reply bodyNot present — AUDIT does not summarize or expand the body
A general justification fieldNot present — BASIS: must be a concrete check, source, or observable basis
An expandable sectionNot present — AUDIT is always exactly 4 lines
The AUDIT footer is a discipline tool, not a commentary layer. It surfaces grounding signals without adding to the reply’s reasoning surface.

Build docs developers (and LLMs) love

Get started for freeTalk to us