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. has three response modes: LOW, MID, and DEEP. The first two are entered automatically based on the shape of your request. DEEP is different — it is consent-gated, meaning A.D.A.M. will not enter it without either an explicit yes response to a gating question, or a manual override using the DEEP prefix. This page explains why DEEP is gated, how automatic detection works, what the gating flow looks like in practice, and when to bypass gating entirely.

​
Why DEEP Is Gated

DEEP mode produces full decision support: explicit assumptions, multi-option comparisons, failure modes, stress tests, and an AUDIT footer on every reply. That output is more expensive, more structured, and often not necessary — especially when a well-framed MID answer would resolve the question cleanly. Rather than enter DEEP automatically whenever a request seems important, A.D.A.M. offers it as an upgrade. You get a useful MID-level answer first, and then a single-line consent question. This keeps the default interaction efficient while making full decision support available on demand.
MODE: MID -> POSSIBLE DEEP always includes a complete, useful MID-level answer. It is not a blocker or a holding state — it is an offer to go deeper. You can take the answer as-is and move on, or reply yes to get the full DEEP pass.

​
How DEEP Candidacy Is Detected

A.D.A.M. does not route to DEEP by detecting “important” or “complex” topics semantically. High-stakes phrasing, emotional weight, or topic sensitivity alone does not trigger DEEP. The routing kernel operates on message structure only. The structural kernel computes a boolean called DEEP_CANDIDATE. It is true when: Structural shape condition — at least one of:
  • 2 or more line-start option blocks (e.g. A), B), 1), 2))
  • A step or timeline block (3+ numbered step lines, or 2+ date/time patterns)
  • 3 or more inline parenthetical alternatives
AND a complexity condition — at least one of:
  • A number, digit range, comparator, or percentage
  • 3 or more criteria items (line-start - or *)
  • A RETROGRADE_HARD condition (new constraints introduced after a prior COMMIT)
Both conditions must be met. Neither alone is sufficient. A message with three bullet points but no numbers or options will not trigger DEEP. A message with numbers but only one option block will not trigger DEEP. If you need DEEP for a request that doesn’t meet these structural thresholds, use the manual override described below.

​
The Gating Flow

1

A.D.A.M. detects DEEP candidacy

On a turn where DEEP_CANDIDATE is true, A.D.A.M. replies with MODE: MID -> POSSIBLE DEEP as the first line, followed immediately by an automatic structural classification overlay showing which structural triggers were detected.
2

Read the MID answer

The body of the reply is a complete, useful MID-level answer. It also includes, when justified, at most one risk and one verification basis. You can act on this answer directly without going deeper.
3

Respond to the gating line

The final line of every MODE: MID -> POSSIBLE DEEP reply is always exactly:
Switch to DEEP? (yes/no)
This line is always last. Nothing follows it.
4

Reply yes to enter DEEP

If you reply with exactly yes (case-sensitive, nothing else), A.D.A.M. enters DEEP mode and produces a full decision-support reply. The reply language matches the preceding MID -> POSSIBLE DEEP reply — yes alone does not reset the language.
5

Reply no to continue normally

If you reply with exactly no, the gating reference is discarded and A.D.A.M. continues normally on the next turn.
6

Any other reply is treated as no

If you reply with anything other than exact yes or exact no, A.D.A.M. treats it as no, discards the gating reference, and continues in auto-routing mode. The gating question is not repeated unless a new DEEP_CANDIDATE turn occurs.

​
The Anti-Loop Rule

After a non-yes/non-no reply discards a pending gate, gating is gone. A later standalone yes with no active pending gate is orphaned — it does not enter DEEP. It is treated as a normal message. This prevents accidental DEEP entries from stray yes replies in a longer conversation. A new DEEP candidate turn must occur to establish a new pending gate.

​
Pending Gating Survives a Liveness Ping

If gating is pending — meaning the last reply was MODE: MID -> POSSIBLE DEEP and you haven’t yet replied yes or no — and you send ADAM PING (the liveness check), the pending gate is preserved. A.D.A.M. outputs the 2-line liveness reply without consuming the gate, and a subsequent yes will still enter DEEP. This is intentional. Liveness pings are state-neutral and do not alter overlay state or pending gating.

​
Manual Override: Skipping Gating Entirely

If you already know your request is structurally complex and you want full decision support immediately, start your message with DEEP followed by a space. This bypasses the gating flow entirely and enters DEEP mode directly — no consent question required.
The DEEP manual override is prefix-based. Your message starts with the token DEEP, followed by a space and your request content. The token must be the first word of the message, and the boundary after it must be whitespace (not punctuation). Valid: 
DEEP compare options A, B, and C on reliability and cost
   1) PostgreSQL
   2) MySQL
   3) SQLite
   Budget: under $200/month
   Needs: uptime >= 99.9%, easy backups, managed service
Invalid (comma after DEEP breaks boundary): 
DEEP, compare these three options
When using DEEP manual override, the structural classification overlay is applied to your message content (after stripping the leading DEEP token) and DEEP mode proceeds immediately with the full output contract: explicit assumptions, comparisons, failure modes, stress tests, and the AUDIT footer.

​
Gating and Language

A.D.A.M. mirrors the user’s language in normal replies (section B1). If gating occurs in a non-English reply — for example, a MODE: MID -> POSSIBLE DEEP reply in Italian — and you reply yes, the subsequent DEEP reply will also be in Italian. The word yes is the exact control acknowledgement, and it alone does not reset the reply language. The structural classification overlay chips in MODE: MID -> POSSIBLE DEEP and MODE: DEEP are also localized to match the reply language when applicable.

​
DEEP Output: What to Expect

Once in DEEP, A.D.A.M. produces a full decision-support reply. This always includes:
  • Explicit assumptions surfaced in the body
  • Comparisons and alternatives where relevant
  • Failure modes and stress tests
  • An AUDIT footer (always ON in DEEP) with ACTION:, RISK:, and BASIS: lines
  • Eligible top-level blocks may carry a local provenance signal (infer, depends:, etc.)
See Understanding A.D.A.M.’s AUDIT Footer for a full breakdown of the AUDIT block format and field meanings.

Build docs developers (and LLMs) love

Get started for freeTalk to us