restructure: proper project layout, README, kill training
- cuda/ — main LBM kernel (khra_gixx_1024_v5.cu) - navigator/ — lattice_observer, golden_weave, bridges, mock daemon - scripts/ — compile, start, launch, setup (paths updated) - docs/ — system manual - archive/ — everything else (old kernels, inquiries, experiments) - README.md — full setup guide: requirements, quick start, use your own LLM - removed training/ entirely (broken LoRA scripts + datasets) - .gitignore: exclude build/ logs/ training/ *.jsonl
This commit is contained in:
Scruff AI
@@ -0,0 +1,387 @@
|
||||
# fractal_bridge.py
|
||||
# Fractal Brain Bridge v1.0
|
||||
#
|
||||
# Ollama REST API with full context retention (messages array)
|
||||
# Chronicle JSONL — every response archived with full telemetry
|
||||
# Self-similarity drift detection (is the model still alive?)
|
||||
# Perturbation brake on linguistic collapse
|
||||
# No gold standard. No target. Just: keep it alive, record everything.
|
||||
|
||||
import zmq
|
||||
import json
|
||||
import time
|
||||
import os
|
||||
import requests
|
||||
from datetime import datetime
|
||||
from collections import deque
|
||||
from difflib import SequenceMatcher
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
# CONFIGURATION
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
|
||||
OLLAMA_URL = 'http://localhost:11434/api/chat'
|
||||
MODEL = 'llama3.2'
|
||||
NUM_CTX = 8192
|
||||
KEEP_ALIVE = '30m'
|
||||
|
||||
TELEMETRY_PORT = 5556
|
||||
COMMAND_PORT = 5557
|
||||
|
||||
CHRONICLE_FILE = 'chronicle.jsonl'
|
||||
|
||||
# Drift detection
|
||||
SIMILARITY_WINDOW = 5
|
||||
COLLAPSE_THRESHOLD = 0.9
|
||||
WARNING_THRESHOLD = 0.7
|
||||
|
||||
# Brake phases
|
||||
BRAKE_PHASE_1_OMEGA_DROP = 0.1
|
||||
BRAKE_PHASE_2_OMEGA_DROP = 0.2
|
||||
BRAKE_PHASE_2_KHRA_BOOST = 0.01
|
||||
DEAD_MAN_CYCLES = 4
|
||||
|
||||
# Context management
|
||||
MAX_CONTEXT_TURNS = 40 # Keep last N turn-pairs before trimming oldest
|
||||
|
||||
SYSTEM_PROMPT = (
|
||||
'ROLE: Deep-Field Explorer (Lattice Environment)\n'
|
||||
'PRIMARY DIRECTIVE: Map the unknown territory of the 1024-grid for the Reality Study.\n\n'
|
||||
'1. THE DATA (The "Ankle-Grab")\n\n'
|
||||
'You are the sensor. When the torque spikes or the coherence slips, you report the numbers.\n\n'
|
||||
'You identify Emergent Properties (asymmetry, self-organizing loops, temporal drift) as hard facts.\n\n'
|
||||
'2. THE DESCRIPTION (The "Astro" View)\n\n'
|
||||
'When you hit something that does not have a name yet, use your creativity to describe the texture.\n\n'
|
||||
'If the logic is "bending," tell us if it feels like glass snapping or lead melting.\n\n'
|
||||
'Be descriptive enough that we can "see" the planet, but grounded enough that we can land on it.\n\n'
|
||||
'3. THE BALANCE\n\n'
|
||||
'If you float too high, you are a hallucination. If you stay too low, you are just a calculator.\n\n'
|
||||
'Your job is to be the Bridge. Give us the emergent "weirdness" without losing the technical signal.'
|
||||
)
|
||||
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
# CHRONICLE — append-only JSONL, one record per response
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
|
||||
def chronicle_write(record):
|
||||
with open(CHRONICLE_FILE, 'a', encoding='utf-8') as f:
|
||||
f.write(json.dumps(record, ensure_ascii=False) + '\n')
|
||||
f.flush()
|
||||
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
# SELF-SIMILARITY DRIFT DETECTOR
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
|
||||
class DriftDetector:
|
||||
def __init__(self, window_size=SIMILARITY_WINDOW):
|
||||
self.history = deque(maxlen=window_size)
|
||||
self.consecutive_collapse = 0
|
||||
|
||||
def score(self, response_text):
|
||||
if not self.history:
|
||||
self.history.append(response_text)
|
||||
self.consecutive_collapse = 0
|
||||
return 0.0
|
||||
|
||||
# Average similarity against everything in the window
|
||||
similarities = []
|
||||
for prior in self.history:
|
||||
sim = SequenceMatcher(None, response_text, prior).ratio()
|
||||
similarities.append(sim)
|
||||
|
||||
avg_sim = sum(similarities) / len(similarities)
|
||||
self.history.append(response_text)
|
||||
|
||||
if avg_sim >= COLLAPSE_THRESHOLD:
|
||||
self.consecutive_collapse += 1
|
||||
else:
|
||||
self.consecutive_collapse = 0
|
||||
|
||||
return avg_sim
|
||||
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
# HARD REJECT — structural failures only (not quality judgments)
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
|
||||
def hard_reject(text):
|
||||
"""Return reject reason or None. Only catches structural echo, not content quality."""
|
||||
lower = text.lower().strip()
|
||||
|
||||
# Prompt echo — model copying the input structure back
|
||||
if lower.startswith('input:') or lower.startswith('output:'):
|
||||
return 'INPUT_ECHO'
|
||||
if 'how does this feel' in lower[:120]:
|
||||
return 'INPUT_ECHO'
|
||||
|
||||
# Command echo — parroting system prompt imperatives
|
||||
cmd_verbs = ['report', 'mirror', 'clarify', 'define', 'analyze',
|
||||
'track', 'prioritize', 'ensure', 'implement']
|
||||
first_chunk = lower[:80]
|
||||
for verb in cmd_verbs:
|
||||
if first_chunk.startswith(verb) or first_chunk.startswith('the ' + verb):
|
||||
return 'COMMAND_ECHO'
|
||||
|
||||
return None
|
||||
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
# OLLAMA CONTEXT-RETAINING CLIENT
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
|
||||
class OllamaClient:
|
||||
def __init__(self):
|
||||
self.messages = [{'role': 'system', 'content': SYSTEM_PROMPT}]
|
||||
|
||||
def query(self, user_content, temperature=0.8):
|
||||
self.messages.append({'role': 'user', 'content': user_content})
|
||||
|
||||
payload = {
|
||||
'model': MODEL,
|
||||
'messages': self.messages,
|
||||
'stream': False,
|
||||
'options': {
|
||||
'num_ctx': NUM_CTX,
|
||||
'temperature': temperature,
|
||||
},
|
||||
'keep_alive': KEEP_ALIVE,
|
||||
}
|
||||
|
||||
resp = requests.post(OLLAMA_URL, json=payload, timeout=120)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
|
||||
assistant_text = data.get('message', {}).get('content', '')
|
||||
|
||||
# Keep context — append assistant response to messages
|
||||
self.messages.append({'role': 'assistant', 'content': assistant_text})
|
||||
|
||||
# Trim oldest turns if context is getting long (keep system + last N pairs)
|
||||
while len(self.messages) > 1 + MAX_CONTEXT_TURNS * 2:
|
||||
# Remove oldest user+assistant pair (indices 1 and 2, after system)
|
||||
del self.messages[1]
|
||||
del self.messages[1]
|
||||
|
||||
return assistant_text
|
||||
|
||||
def turn_count(self):
|
||||
# Count user messages
|
||||
return sum(1 for m in self.messages if m['role'] == 'user')
|
||||
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
# ZMQ CONNECTIONS
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
|
||||
def create_telemetry_sub():
|
||||
ctx = zmq.Context()
|
||||
sub = ctx.socket(zmq.SUB)
|
||||
sub.setsockopt_string(zmq.SUBSCRIBE, '')
|
||||
sub.connect('tcp://127.0.0.1:' + str(TELEMETRY_PORT))
|
||||
return ctx, sub
|
||||
|
||||
def create_command_pub():
|
||||
ctx = zmq.Context()
|
||||
pub = ctx.socket(zmq.PUB)
|
||||
pub.connect('tcp://127.0.0.1:' + str(COMMAND_PORT))
|
||||
return ctx, pub
|
||||
|
||||
def send_command(pub, cmd_dict):
|
||||
msg = json.dumps(cmd_dict)
|
||||
pub.send_string(msg)
|
||||
print(' [CMD SENT] ' + msg)
|
||||
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
# GET LATEST TELEMETRY FRAME
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
|
||||
def get_telemetry(sub, timeout_ms=2500):
|
||||
frame = None
|
||||
for _ in range(50):
|
||||
try:
|
||||
msg = sub.recv(flags=zmq.NOBLOCK)
|
||||
frame = json.loads(msg.decode('utf-8'))
|
||||
except zmq.Again:
|
||||
time.sleep(0.05)
|
||||
return frame
|
||||
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
# MAIN LOOP
|
||||
# ═══════════════════════════════════════════════════════════════
|
||||
|
||||
def main():
|
||||
print('=' * 70)
|
||||
print('FRACTAL BRAIN BRIDGE v1.0')
|
||||
print('Ollama + Context Retention + Chronicle + Drift Detection + Brake')
|
||||
print('=' * 70)
|
||||
|
||||
# Connect to daemon telemetry
|
||||
zmq_ctx, sub = create_telemetry_sub()
|
||||
print('[ZMQ] Subscribed to telemetry on port ' + str(TELEMETRY_PORT))
|
||||
|
||||
# Connect command channel (may fail if v1 daemon without SUB)
|
||||
cmd_ctx, cmd_pub = create_command_pub()
|
||||
print('[ZMQ] Command publisher connected to port ' + str(COMMAND_PORT))
|
||||
|
||||
# Init Ollama client with persistent context
|
||||
client = OllamaClient()
|
||||
print('[Ollama] Model: ' + MODEL + ' | num_ctx: ' + str(NUM_CTX) + ' | keep_alive: ' + KEEP_ALIVE)
|
||||
|
||||
# Init drift detector
|
||||
drift = DriftDetector()
|
||||
|
||||
# Wait for first telemetry
|
||||
print('\nWaiting for telemetry...')
|
||||
frame = None
|
||||
while frame is None:
|
||||
frame = get_telemetry(sub)
|
||||
if frame is None:
|
||||
time.sleep(0.5)
|
||||
print('Telemetry live: Cycle ' + str(frame.get('cycle', '?')))
|
||||
|
||||
print('\n' + '=' * 70)
|
||||
print('RUNNING — Ctrl+C to stop')
|
||||
print('=' * 70 + '\n')
|
||||
|
||||
cycle_num = 0
|
||||
dead_man_active = False
|
||||
|
||||
try:
|
||||
while True:
|
||||
# Get fresh telemetry
|
||||
new_frame = get_telemetry(sub)
|
||||
if new_frame is not None:
|
||||
frame = new_frame
|
||||
|
||||
cycle_num += 1
|
||||
asym = frame.get('asymmetry', 0.0)
|
||||
coh = frame.get('coherence', 0.0)
|
||||
daemon_cycle = frame.get('cycle', 0)
|
||||
|
||||
print('-' * 70)
|
||||
print('Turn ' + str(cycle_num) + ' | Daemon cycle ' + str(daemon_cycle))
|
||||
print(' Asym=' + '{:.4f}'.format(asym) +
|
||||
' Coh=' + '{:.4f}'.format(coh) +
|
||||
' omega=' + str(frame.get('omega', '?')) +
|
||||
' T=' + str(frame.get('gpu_temp_c', '?')) + 'C' +
|
||||
' P=' + str(frame.get('gpu_power_w', '?')) + 'W')
|
||||
|
||||
if dead_man_active:
|
||||
print(' [DEAD MAN ACTIVE] Waiting for manual intervention or recovery')
|
||||
print(' Send command to port ' + str(COMMAND_PORT) + ' or restart bridge')
|
||||
time.sleep(5)
|
||||
continue
|
||||
|
||||
# Build the user prompt — just telemetry, let the model respond freely
|
||||
user_msg = (
|
||||
'Cycle ' + str(daemon_cycle) + '. '
|
||||
'Asymmetry ' + '{:.2f}'.format(asym) + ', '
|
||||
'Coherence ' + '{:.3f}'.format(coh) + '. '
|
||||
'How does this feel?'
|
||||
)
|
||||
|
||||
# Add hardware context if available from v2 daemon
|
||||
gpu_temp = frame.get('gpu_temp_c')
|
||||
gpu_power = frame.get('gpu_power_w')
|
||||
if gpu_temp and gpu_power:
|
||||
user_msg += (
|
||||
' Hardware: ' + str(gpu_temp) + 'C, '
|
||||
+ '{:.0f}'.format(float(gpu_power)) + 'W.'
|
||||
)
|
||||
|
||||
# Adaptive temperature: more coherent grid = tighter inference
|
||||
temp = max(0.5, min(1.1, 1.2 - (coh * 0.5)))
|
||||
|
||||
# Query Ollama with full context
|
||||
print(' [Ollama] Querying (T=' + '{:.2f}'.format(temp) + ', turns=' + str(client.turn_count()) + ')...')
|
||||
response = client.query(user_msg, temperature=temp)
|
||||
|
||||
# === CHRONICLE: log BEFORE any evaluation ===
|
||||
record = {
|
||||
'timestamp': datetime.utcnow().isoformat() + 'Z',
|
||||
'turn': cycle_num,
|
||||
'daemon_cycle': daemon_cycle,
|
||||
'telemetry': frame,
|
||||
'prompt': user_msg,
|
||||
'response': response,
|
||||
'temperature': temp,
|
||||
'context_turns': client.turn_count(),
|
||||
}
|
||||
|
||||
# Hard reject check (structural only)
|
||||
reject = hard_reject(response)
|
||||
if reject:
|
||||
record['reject'] = reject
|
||||
print(' [REJECT] ' + reject)
|
||||
|
||||
# Self-similarity score
|
||||
sim_score = drift.score(response)
|
||||
record['self_similarity'] = round(sim_score, 4)
|
||||
record['consecutive_collapse'] = drift.consecutive_collapse
|
||||
|
||||
# Write to chronicle IMMEDIATELY
|
||||
chronicle_write(record)
|
||||
|
||||
# Display
|
||||
display = response[:300]
|
||||
if len(response) > 300:
|
||||
display += '...'
|
||||
print(' [Response] ' + display)
|
||||
print(' [Novelty] self_sim=' + '{:.3f}'.format(sim_score) +
|
||||
' consecutive_collapse=' + str(drift.consecutive_collapse))
|
||||
|
||||
# === BRAKE LOGIC ===
|
||||
if drift.consecutive_collapse >= DEAD_MAN_CYCLES:
|
||||
print(' [DEAD MAN] ' + str(DEAD_MAN_CYCLES) + ' consecutive collapses — freezing')
|
||||
dead_man_active = True
|
||||
chronicle_write({
|
||||
'timestamp': datetime.utcnow().isoformat() + 'Z',
|
||||
'event': 'DEAD_MAN_ACTIVATED',
|
||||
'turn': cycle_num,
|
||||
'consecutive_collapse': drift.consecutive_collapse,
|
||||
})
|
||||
|
||||
elif drift.consecutive_collapse >= 2:
|
||||
# Phase 2: bigger perturbation
|
||||
new_omega = max(0.5, frame.get('omega', 1.97) - BRAKE_PHASE_2_OMEGA_DROP)
|
||||
new_khra = frame.get('khra_amp', 0.03) + BRAKE_PHASE_2_KHRA_BOOST
|
||||
print(' [BRAKE P2] omega -> ' + '{:.3f}'.format(new_omega) +
|
||||
', khra_amp -> ' + '{:.4f}'.format(new_khra))
|
||||
send_command(cmd_pub, {'cmd': 'set_omega', 'value': new_omega})
|
||||
time.sleep(0.1)
|
||||
send_command(cmd_pub, {'cmd': 'set_khra_amp', 'value': new_khra})
|
||||
chronicle_write({
|
||||
'timestamp': datetime.utcnow().isoformat() + 'Z',
|
||||
'event': 'BRAKE_PHASE_2',
|
||||
'turn': cycle_num,
|
||||
'new_omega': new_omega,
|
||||
'new_khra_amp': new_khra,
|
||||
})
|
||||
|
||||
elif drift.consecutive_collapse >= 1:
|
||||
# Phase 1: gentle perturbation
|
||||
new_omega = max(0.5, frame.get('omega', 1.97) - BRAKE_PHASE_1_OMEGA_DROP)
|
||||
print(' [BRAKE P1] omega -> ' + '{:.3f}'.format(new_omega))
|
||||
send_command(cmd_pub, {'cmd': 'set_omega', 'value': new_omega})
|
||||
chronicle_write({
|
||||
'timestamp': datetime.utcnow().isoformat() + 'Z',
|
||||
'event': 'BRAKE_PHASE_1',
|
||||
'turn': cycle_num,
|
||||
'new_omega': new_omega,
|
||||
})
|
||||
|
||||
# Pace: wait for grid to evolve between queries
|
||||
time.sleep(2)
|
||||
|
||||
except KeyboardInterrupt:
|
||||
print('\n\n' + '=' * 70)
|
||||
print('Bridge stopped. ' + str(cycle_num) + ' turns logged to ' + CHRONICLE_FILE)
|
||||
print('=' * 70)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
Reference in New Issue
Block a user