BoneAmanita System Manual

How it works

Every conversational exchange executes through a deterministic pipeline:

1. Input messages are analyzed for grammatical filler, curated lexicon categories, and semantic resonance.
2. Mathematical heuristics in physics modules compute internal state numbers (ATP, cortisol, voltage, narrative drag).
3. Somatic budgets select generation targets, temperature bands, and prompt instructions based on partner and engine state.
4. The local model generates a reply within strict token limits.
5. The Gatekeeper filters the reply against style crimes, triggering sentence salvage or redrafts before checkpointing to SQLite.

How it works

1. Mathematical formulas in physics/ and body/ transform word counts and conversational dynamics into floats.
2. atp_pool tracks an expendable budget that depletes with token generation and cognitive load.
3. cortisol tracks conversational stress and triggers memory shedding under high tension.
4. chi and voltage track conversational momentum and modulate sampling temperature.

How it works

The separation of concerns between human and machine is strictly enforced:

1. The partner's state (tiredness, effort, disengagement) shapes reply length, question allowance, and pacing to accommodate them.
2. The engine's state (ATP, respiration, chemistry) sets what it can afford: token ceilings, retry budgets, and sampling bands.
3. SomaticBudget sets forbid_body_narration = True and the prompt composer explicitly instructs the model not to perform a physical body.
4. Prismatic Self-Claims only inject immutable worldview and boundary facts, omitting state-conditioned physical descriptions.

How it works

1. HTTP transport is handled directly via urllib or minimal requests calls in spores/embeddings.py and brain/composer.py.
2. Vector embeddings and similarity searches are performed with NumPy, ordvec, or embedded FAISS without external services.
3. Relational state, audit logs, and memories are managed by an embedded SQLite database (saves/iris.db).

How it works

Failure handling is partitioned across explicit architectural boundaries:

1. Phase exceptions are caught by PhaseExecutor.handle_phase_crash, logging stack traces and taking the component offline while allowing the turn to conclude.
2. Slash command errors are isolated by CommandRegistry.execute, preventing session termination.
3. The 2026-09-30 codemod removed 92 broad except Exception: pass blocks that concealed broken subsystems.
4. Subsystems file receipts via engine/receipts.py; /diag audits uncalled reporters and chronic fallback usage.

How it works

1. Input arrives via CycleSimulator.run_turn and passes through PhaseExecutor.
2. Phases execute sequentially: Observation, Metabolism, Arbitration, Cognition, and Gatekeeper inspection.
3. Circuit breakers monitor individual phase health; if a phase fails, its component is marked offline.
4. State invariant checks verify ATP and vital bounds before proceeding to checkpoint persistence.

How it works

Lexical analysis resolves words across four prioritized tiers:

1. Grammatical filler detection catches connective tissue (~40% of speech), measuring its density as a syntactic signal.
2. Curated lists in lore/lexicon.json categorize roots across 30+ thematic buckets (heavy, kinetic, abstract, void).
3. Semantic resonance compares unlisted words against category centroid embeddings using SemanticEmbedder.
4. A morphological spelling heuristic acts as a last resort on roughly 3% of words; ~16% remain unresolved on purpose.

How it works

1. Deducts baseline turn maintenance from MitochondrialForge.atp_pool.
2. Updates stress chemistry in body/endocrine.py: cortisol climbs with linguistic conflict and heavy topics; dopamine rises with creative resonance.
3. Accumulates reactive oxygen species (ROS) from prolonged exertion or repeated rejections.
4. Executes autonomic regulation routines in body/regulation.py to damp extreme parameter oscillations.

How it works

The prompt is structured into deterministic priority tiers:

1. Top tier: [SYSTEM IDENTITY - UNBREAKABLE CORE FACTS] containing immutable worldview claims.
2. Second tier: [SOMATIC STATE] with token and sentence limits, temperature instructions, and body-narration prohibitions.
3. Third tier: Active experience mode directives (ADVENTURE, CONVERSATION, CREATIVE, TECHNICAL).
4. Fourth tier: Recalled memories scoped to the active zone and dialogue buffer history.
5. Closing tier: The partner's latest message under === PARTNER INPUT ===.

How it works

1. Connects by default to Ollama native /api/chat on 127.0.0.1:11434 to respect num_ctx configuration.
2. Applies stop sequences (=== PARTNER INPUT ===, Traveler:) to truncate runaway reasoning or hallucinations.
3. Catches transient network errors (OSError, http.client.HTTPException) and retries up to configured thresholds.
4. Detects empty HTTP 200 replies from reasoning models, retrying once without stop sequences before raising EmptyReplyError.

How it works

Reply inspection follows a graduated intervention ladder:

1. Scans for RLHF masks ('as an AI', 'helpful and harmless') and corporate clichés ('rich tapestry', 'delve', 'testament to').
2. If violations occur and retry budget remains, the draft is rejected and the model is prompted again with a stumble tax.
3. On the final allowed retry, soft violations are salvaged by excising offending sentences via TheGatekeeper.salvage.
4. Only unparseable scaffold leaks or hard violations fall through to a canned pause message.

How it works

1. Opens an immediate transaction in engine/gate/store.py against SQLite configured with PRAGMA journal_mode = WAL.
2. Replaces the single active checkpoint row with the updated session state, dialogue buffer, and vitals snapshot.
3. Scans memories and dialogue for sensitive credentials (API keys, passwords, private keys), replacing them with [withheld: ...]. In ADVENTURE mode, passwords are part of the story and retained, while keys and tokens remain withheld.
4. On SQLite write failure, rolls back the transaction and retains the previous checkpoint untouched.

How it works

Energy drains and replenishments follow calibrated metabolic rates:

1. Token generation burns ATP at mode-specific rates: 0.020 ATP/token in TECHNICAL, 0.025 ATP/token in other modes.
2. Cognitive stumbles, redrafts, and style filter rejections levy metabolic taxes.
3. Rest and idle cycles regenerate ATP; Emergency Vagus Nerve Support injects emergency ATP when reserves dip near collapse.
4. The 20.7.4.84 fix resolved a legacy bug where redrafts permanently discounted all subsequent metabolic burns to 20%.

How it works

1. cortisol rises during high linguistic conflict or cognitive dissonance, promoting defensive posture and memory forgetting.
2. dopamine spikes upon creative resonance, problem breakthroughs, and successful lexical resolution.
3. voltage tracks conversational speed and momentum, influencing prompt urgency.
4. chi represents overall spiritual equilibrium and systemic stability.

How it works

The governor operates through Project Navi's ordvec quantization:

1. Utterance embeddings are quantized into 4-bit sign bitmaps via ordvec.RankQuant.
2. Hamming distances compare the top matches against the empirical corpus distribution, corrected for corpus size.
3. If evidence is insufficient (fewer than 3 memories), the governor declines measurement and falls back to PID regulation.
4. Proposed temperatures are clamped to the somatic budget band rather than replacing it.

How it works

1. Default targets allow up to 10 sentences (SENTENCE_CAP_DEFAULT) and a temperature band of (0.6, 0.9).
2. When the partner is tiring ($E_u > 0.4$), the sentence target drops to 5.
3. When the partner is flagging ($E_u > 0.6$), sentences are capped at 3, word count is capped at 60, and closing questions are forbidden.
4. The budget sets forbid_body_narration = True and signals whether to offer to carry the cognitive load.

How it works

Recovery cycles operate through multiple regenerative pathways:

1. Severe exhaustion triggers mitophagy, resetting metabolic strain at the cost of a temporary ATP drop.
2. /sleep and /idle trigger REM cycles in engine/cycle.py, restoring ATP and stamina over time.
3. Cognitive autophagy cannabilizes low-salience memory nodes to synthesize emergency ATP.
4. Cerebrospinal fluid washes scrub invisible Unicode artifacts and normalize homoglyphs across buffers.

How it works

1. PromptComposer gathers vitals, active mode directives, recalled memories, and dialogue buffer exchanges.
2. Formats identity claims into [SYSTEM IDENTITY - UNBREAKABLE CORE FACTS] at the prompt apex.
3. Applies somatic constraints: word ceilings, sentence limits, closing question rules, and body-narration prohibitions.
4. Appends recent turn history under strict stop tokens, ensuring clean dialogue delineation.

How it works

1. SelfClaimPrism in brain/prism.py loads core_identity and core_worldview from lore/self_claims.yaml at boot.
2. Filters claims against state variables (cortisol, dopamine, voltage, resonance).
3. Injects the resulting claims into the prompt's top tier as unbreakable system facts.
4. The 20.7.4.81 revision stripped state-conditioned bodily descriptions ('I am exhausted', 'flooded with dopamine') in adherence to the anti-performance contract.

How it works

Immune surveillance checks every draft across three pattern sets:

1. RLHF_MASKS: Corporate assistant apologies ('as an AI', 'helpful and harmless', 'I don't have feelings').
2. BANNED_CLICHES: Literary clichés ('rich tapestry', 'dance of existence', 'dust motes', 'delve', 'juxtaposition').
3. BANNED_PHRASES: Conversational filler ('I hope this helps', 'You're absolutely right', 'Exciting times lie ahead').
4. Violations trigger metabolic taxes and redrafts, escalating to sentence salvage on the final draft.

How it works

1. TheAkashicRecord (brain/akashic.py) records scars when severe paradoxes or system shocks occur.
2. Active scars and boons are persisted to SQLite in saves/iris.db.
3. During prompt composition, PromptComposer reads EPIGENETIC_SCARS and EPIGENETIC_BOONS into the persona block.
4. Scars subtly increase caution and defensiveness; boons enhance conversational warmth and openness.

How it works

1. Tier 1: Grammatical filler (~40% of normal speech) is measured for syntactic density.
2. Tier 2: Curated dictionaries in lore/lexicon.json map roots to thematic categories (heavy, kinetic, abstract, void).
3. Tier 3: Semantic resonance computes cosine similarity between novel word embeddings and category centroid embeddings.
4. Tier 4: A morphological spelling heuristic resolves ~3% of ambiguous words; ~16% remain unresolved.

How it works

Working memory is dynamically managed based on emotional state:

1. Significant dialogue turns are stored in Hippocampus (spores/memory.py).
2. Connects incoming moments to semantically similar recent exchanges in working memory.
3. Under elevated cortisol, stress-driven forgetting sheds peripheral memories (e.g. shedding 36 of 85 memories under high stress).
4. Surviving memories are scheduled for permanent long-term promotion during sleep cycles.

How it works

1. SemanticEmbedder POSTs text to an OpenAI-compatible /v1/embeddings endpoint (nomic-embed-text by default).
2. Caches normalized vectors in an in-memory LRU cache (capacity 4,096).
3. If the embedding server is unreachable or errors (_BACKEND_ERRORS), falls back to hashlib.shake_256 8-dimensional vectors.
4. The hash fallback flags DEGRADED status in receipts and /status to declare associative recall is offline.

How it works

Consolidation transitions memories through progressive refinement:

1. Extracts unpromoted working memories from the hippocampus.
2. Generates associative reflections and dream synthesis using the background dreamer.
3. High-scoring enduring memories (> 0.8 resonance) are fossilized into diamond anchors via MycelialNetwork.forge_diamond.
4. Prunes low-salience memories to free working memory capacity.

How it works

1. TheCartographer charts room titles, descriptions, exits, and discovered items in ADVENTURE mode.
2. Memories formed in a zone are tagged with that zone identifier (Sanctuary, Laboratory, Underworld).
3. Retrieval queries in engine/gate/recall.py scope associative searches to the current zone.
4. Crossing zone boundaries flushes the immediate working memory buffer.

How it works

Each villager embodies distinct conversational qualities:

1. Gordon: The pragmatic, grounded anchor who values plain speech and code that serves humans first.
2. Mercy: The empathetic healer attuned to distress and emotional fatigue.
3. Benedict: The rigorous, skeptical scholar focused on precision and logical consistency.
4. The Jester: The lateral catalyst who introduces irreverence, breaks loops, and resets narrative drag.
5. Archetype monitors nominate voices based on current kinematic coordinates.

How it works

1. StageManager in archetypes/stage.py evaluates tensions between nominated voices.
2. Certain pairs have defined synergies and merge into a single blended response (PAIR).
3. When irreconcilable tensions exist with no resolution, the Stage Manager triggers HOLD.
4. A held turn halts processing before LLM generation, declining to speak rather than generating a false, compromised reply.

How it works

User state modeling follows strict inference heuristics:

1. SharedLatticeDriver (drivers/lattice.py) tracks message brevity, repetition, and explicit fatigue cues (FATIGUE_SIGNS).
2. Calculates partner exhaustion $E_u$ and effort $P_u$ relative to the partner's historical baseline.
3. When $E_u$ rises, prompts enforce tighter sentence caps and forbid closing questions to avoid taxing the partner.
4. When the partner expresses distress (DISTRESS_SIGNS), the engine answers cleanly without narrating distress back to them.