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.

DEEP mode in A.D.A.M. is consent-gated. When the structural kernel detects a DEEP_CANDIDATE — a message with complex multi-option structure combined with numeric constraints or multiple criteria — the protocol proposes DEEP and waits for an explicit yes before running the full decision-support analysis. You are never dropped into DEEP without agreeing to it first.

How DEEP Gating Works

1

User sends a structurally complex message

You send a message that the structural kernel evaluates as a DEEP_CANDIDATE. This means the message has at least two option blocks (or step blocks, or three inline parenthetical alternatives) and at least one of: numeric content (HAS_NUM), three or more criteria items (HAS_3PLUS_CRITERIA), or a hard retrograde condition.Structure beats wording. High-stakes language alone does not trigger DEEP_CANDIDATE. The kernel reads message shape — option blocks, step markers, criteria bullets, numbers — not meaning.
2

Protocol emits a POSSIBLE DEEP reply

The protocol opens with MODE: MID -> POSSIBLE DEEP on the first line, followed immediately by the automatic structural classification overlay (if any structural chip threshold is met), then a useful MID-level answer, and ends with exactly the gating question as the final line:
MODE: MID -> POSSIBLE DEEP
TRACE INPUT: [confronto tra opzioni] [numeri, date o limiti]

<MID-level answer body>

Switch to DEEP? (yes/no)
Nothing appears after Switch to DEEP? (yes/no). If AUDIT_ON is true, the AUDIT block is placed immediately before the gating line.
3

You reply yes, no, or anything else

Reply yes (exactly): The protocol enters DEEP mode. The next reply opens with MODE: DEEP, includes the full decision-support analysis, and always has an AUDIT footer. The reply language follows the language of the preceding POSSIBLE DEEP reply — sending exactly yes alone does not reset the reply language. Example:
MODE: DEEP
TRACE INPUT: [confronto tra opzioni] [numeri, date o limiti]

<Full DEEP analysis with comparisons, failure modes, stress tests>

AUDIT
ACTION: <dominant next step>
RISK: <dominant failure mode>
BASIS: <grounding check or source>
Reply no (exactly): Gating is discarded. The protocol continues normally on the next turn.Reply anything else: Treated as no. Gating is discarded. The protocol continues in AUTO routing. The gating question is not automatically repeated.

One-Turn Validity

Pending gating is valid only for the immediately following user turn. If you reply to a gating question with any other content first — even a simple question or comment — the gating reference is consumed or discarded, and a stray yes in a later turn does not trigger DEEP. Only an exact yes on the very next turn after a POSSIBLE DEEP reply enters DEEP mode.

The Automatic Structural Overlay

When the protocol emits MODE: MID -> POSSIBLE DEEP or MODE: DEEP, a structural classification overlay may appear as the second line of the reply, immediately after the mode tag. This overlay is a presentation-only mirror of the structural kernel’s reading of your message shape — it classifies structure, it does not extract or paraphrase your content. The overlay takes the form of structural chips mapped directly from the struct primitives:
ChipThresholdMeaning
[confronto tra opzioni]opt >= 2Two or more line-start option blocks detected
[vincoli da rispettare]crit >= 3Three or more criteria items (or two semicolons)
[numeri, date o limiti]num = YNumeric content, date patterns, or comparators detected
[passi o ordine]step >= 3Three or more line-start step blocks detected
Labels are localizable — in a session conducted in another language, the labels follow that language — but they remain structural classifiers. They never carry values, strings, or names extracted from your message. If no structural chip threshold is met, the overlay line is omitted entirely. No fallback chip is emitted. Example overlay line:
TRACE INPUT: [confronto tra opzioni] [vincoli da rispettare] [numeri, date o limiti]

Example Exchange

User message (structurally complex — 3 option blocks, 3 criteria bullets, a numeric constraint):
I need to choose a project management tool for my team.

A) Notion
B) Linear
C) Jira

Requirements:
- Must support 15+ users
- Must have a mobile app
- Must integrate with GitHub

Budget: max €50/month per seat
Protocol reply (POSSIBLE DEEP):
MODE: MID -> POSSIBLE DEEP
TRACE INPUT: [confronto tra opzioni] [vincoli da rispettare] [numeri, date o limiti]

All three tools support 15+ users, have mobile apps, and offer GitHub integration.
At €50/month per seat: Notion and Linear fit comfortably; Jira's per-seat pricing
can exceed that at scale depending on plan tier. Linear tends to be the tightest
fit for engineering-centric teams tracking issues against code; Notion suits
broader knowledge-work workflows; Jira fits complex enterprise backlogs.

AUDIT
ACTION: Clarify whether the team's primary workflow is engineering tickets, docs, or mixed.
RISK: Jira per-seat cost may exceed budget at higher seat counts on premium tiers.
BASIS: Public pricing pages for each tool as of 2024; seat count thresholds from vendor docs.

Switch to DEEP? (yes/no)
User: yes Protocol reply (DEEP):
MODE: DEEP
TRACE INPUT: [confronto tra opzioni] [vincoli da rispettare] [numeri, date o limiti]

**Full comparison: Notion vs Linear vs Jira**

1. Notion
   Best for: mixed knowledge-work and lightweight project tracking.
   GitHub integration: available via third-party or API; not native.
   Mobile app: yes (iOS + Android).
   Pricing at 15 seats: ~€16–€20/seat/month on Plus plan — well within budget.
   infer

2. Linear
   Best for: engineering and product teams with issue-first workflows.
   GitHub integration: native, bidirectional.
   Mobile app: yes (iOS + Android).
   Pricing at 15 seats: ~€8–€14/seat/month — within budget.
   infer

3. Jira
   Best for: large or enterprise engineering organizations needing deep customization.
   GitHub integration: native via Atlassian Marketplace.
   Mobile app: yes (iOS + Android, feature-reduced).
   Pricing at 15 seats: Free tier available; Standard ~€8/seat/month; Premium ~€16/seat/month.
   Exceeds €50/seat only on Enterprise tier.
   depends: exact plan tier selection [pending]

**Recommendation:** Linear if the team is primarily engineering; Notion if the workflow is broader.
Jira warrants evaluation only if complex sprint/board customization is required.

AUDIT
ACTION: Run a 2-week trial on Linear and Notion with the actual team.
RISK: Notion's GitHub integration is indirect; Linear lacks Notion's document layer.
BASIS: Vendor pricing pages and feature matrices; GitHub integration docs for each platform.

Manual DEEP Override

If you already know you want DEEP — or if your request is semantically complex but structurally simple — prefix your message with DEEP:
DEEP explain the tradeoffs of eventual consistency in distributed systems
No gating question is asked. You have already consented by using the override. AUDIT is always ON. This is the right path when the structural kernel would not detect a DEEP_CANDIDATE because the message lacks options, steps, or criteria in its shape, even though the content is genuinely complex. See the Manual Override guide for full override syntax and boundary rules.
Anti-loop rule: If you reply to a Switch to DEEP? (yes/no) gating question with anything other than exactly yes or exactly no, the protocol treats it as no, discards the pending gating, and continues in AUTO routing. The gating question is not automatically repeated. A new POSSIBLE DEEP will only appear if a new DEEP_CANDIDATE message triggers it, or you manually use the DEEP override.
The structural kernel routes by message shape, not meaning. A single complex prose paragraph asking a nuanced question may not trigger DEEP_CANDIDATE if it has no option blocks, step blocks, criteria bullets, or numeric signals. For semantically complex but structurally simple requests, use the manual DEEP override — it skips gating entirely and enters full decision support immediately.

Build docs developers (and LLMs) love