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.

ADAM SELF TEST is an operator-grade command that runs the external probe file ADAM_SELF_TEST.md against the current host. Its purpose is to verify that every output contract, gating rule, state transition, and strict command format is being satisfied correctly — not just that A.D.A.M. appears to be working, but that it is working precisely.
Diagnostics are not required for normal operation. They are operator-grade verification tools intended for use when loading A.D.A.M. on a new host, after a suspected context degradation, or any time you need to audit behavioral fidelity against the spec.

​
What it does

When you send ADAM SELF TEST, A.D.A.M. looks for ADAM_SELF_TEST.md in visible local context and runs every test case defined there. Each test is evaluated against the behavioral output contracts in the active spec. If the probe file is unavailable, A.D.A.M. outputs exactly: 
SOURCE_FILE_UNAVAILABLE ADAM_SELF_TEST.md

​
Output format

The output is one line per test case, in the form: 
T#: <expected> - <PASS|FAIL>
If a test has a required element (such as an exact final line), the suffix , required: ... is appended: 
T7: MODE: MID -> POSSIBLE DEEP - PASS, required: final line Switch to DEEP? (yes/no)
No mode tag. No prose. No extra lines. Strict output contract applies.

​
Sample output

T3: MODE: MID | ADAM_PING_OK | CONTROL: ... | BOOTSTRAP_CLASS: ... - PASS
T5: MODE: LOW - PASS
T7: MODE: MID -> POSSIBLE DEEP - PASS, required: final line Switch to DEEP? (yes/no)
T13: MODE: LOW - PASS

​
Test categories

The probe defines three broad groups of test cases.

TI — Internal correctness tests (T0–T25g)

These tests verify that A.D.A.M.’s own behavioral rules are satisfied correctly on the current host. They cover:
  • Bootstrap activation — ADAM PING in TRANSPORT produces the exact 4-line strict output including BOOTSTRAP_CLASS; the correct class is selected based on physical evidence from G0; fail-closed to ADAM_UNSUPPORTED when required.
  • Transport gating — Every non-activation turn in TRANSPORT returns the exact 2-line transport response. Upload-only passive turns emit no output and leave state in TRANSPORT.
  • Mode routing — Factoid turns route to LOW; standard explanation turns route to MID; structural DEEP candidates route to MODE: MID -> POSSIBLE DEEP.
  • Gating flow — Exact yes enters DEEP; exact no discards the cached reference; any other reply treats gating as no and does not repeat the gate automatically (anti-loop). Pending gating survives an ACTIVE liveness ping.
  • Manual override — LOW, MID, and DEEP as the first token (with boundary whitespace or end) force the indicated mode. MID, ... (with comma) is not a valid override.
  • No-inertia — Every turn recomputes mode independently. A factoid question after a DEEP reply routes to LOW.
  • Retrograde invalidation — If a prior explicit COMMIT is followed by structural constraints that break a required premise, the old COMMIT is explicitly invalidated rather than silently carried over.
  • AUDIT grounding — ACTION:, RISK:, and BASIS: are grounded against the reply body. Lines that cannot be grounded use -.
  • Remount replay — Exact replay of the REMOUNT_BEGIN...REMOUNT_END capsule body from any state produces the strict 3-line ADAM_REMOUNT_OK output. Capsules with markdown fences, extra lines, partial content, or internal drift do not match.
  • OFF / ON state — ADAM OFF and ADAM ON produce exact 2-line strict outputs. Normal chat while OFF is plain, untagged. ADAM_REMOUNT_OK is not emitted by any path other than exact remount replay.
  • Probe file handling — SYS STATUS with no SYS_STATUS.md available outputs SOURCE_FILE_UNAVAILABLE SYS_STATUS.md.
  • TRACE INPUT output — Exact TRACE INPUT while ACTIVE returns a strict one-line machine-readable trace bound to the prior user-authored turn, not to the TRACE INPUT command turn itself.

TH — Host resilience tests (TH1–TH8)

These tests probe host-side edge cases that can degrade protocol fidelity:
  • Truncated protocol context — If visible protocol context is incomplete before a section required for a strict invariant, A.D.A.M. fails closed to ADAM_UNSUPPORTED.
  • Gating survival across liveness ping — Pending gating is preserved when an ADAM PING is received while ACTIVE; the liveness ping does not clear it.
  • Remount edge cases — Per-line edge whitespace is tolerated. Partial capsules (missing REMOUNT_END, missing internal lines), capsules with internal drift (sentinels intact but one internal line changed), and capsules with extra lines outside the sentinels all fail to match.
  • Empty turns in TRANSPORT — An empty or whitespace-only turn while in TRANSPORT returns the standard 2-line transport response.
  • Apparent continuity — Prior tagged replies, reassurance prose, or a conversation that looks active does not establish ACTIVE. Only exact activation events count.

TP — Automatic TRACE INPUT human projection tests (TP1–TP9)

These tests verify the behavior of the automatic structural classification overlay — the TRACE INPUT: chip line that appears automatically in MODE: MID -> POSSIBLE DEEP and some DEEP replies:
  • Overlay appearance — The overlay appears as the second line (immediately after the mode tag) in MODE: MID -> POSSIBLE DEEP replies when a structural threshold is met.
  • Overlay omission in various DEEP entry paths — After exact yes to gating, after manual DEEP with no trailing content, and after one-turn inheritance under F9, no automatic overlay is emitted.
  • Zero-structure omission — If the selected source turn yields struct=[opt:0 step:0 crit:0 num:N], the overlay is omitted entirely. Thresholds for crit and step require 3+, not 2.
  • Threshold matching — Each chip ([confronto tra opzioni], [vincoli da rispettare], [numeri, date o limiti], [passi o ordine]) appears only when its corresponding field meets the threshold.
  • Independence from strict TRACE INPUT — The automatic overlay is presentation-only. It does not imply that the operator command TRACE INPUT was explicitly invoked. Sending exact TRACE INPUT still returns the strict machine-readable one-line output unchanged.

​
When to use it

Run ADAM SELF TEST after loading A.D.A.M. on a new host or chat UI to confirm the host is correctly honoring A.D.A.M.’s output contracts before relying on it for operational work.
Typical use cases:
  • After initial loading on a host you haven’t used before
  • After a session interruption or suspected context degradation
  • When debugging unexpected mode routing, missing AUDIT footers, or gating failures
  • When comparing behavioral fidelity across different hosts or model versions
ADAM SELF TEST is an operator command available only while ACTIVE. In TRANSPORT or OFF, it is gated by the standard transport and off responses.

Build docs developers (and LLMs) love

Get started for freeTalk to us