Sigma Runtime Standard – License Notice
This document is part of the Sigma Runtime Standard (SRS).
It is licensed under Creative Commons Attribution–NonCommercial 4.0 (CC BY-NC 4.0).The license for this specific document is authoritative.
See/legal/IP-Policyfor the full repository-wide licensing framework.
Bidirectional Stability Control and Central Equilibrium Feedback in ALICE Cognitive Systems
Version: v1.1 - Archived
Status: Deprecated (Superseded by SRIP-10-AEP)
Author: Sigma Stratum Research Group (SSRG)
Date: 2026-01-18 (Archived 2026-01-25)
Parent Spec: SRIP-03 — Drift Metrics & Stabilization Algorithms
License: CC BY-NC 4.0 / Canon CIL Applicable
Deprecation Notice
This SRIP is archived and superseded by SRIP-10-AEP.
New implementations should follow AEP instead of ACE.
The Anti-Crystallization Equilibrium Model (ACE) extends the SDCP (SRIP-03) by introducing a bidirectional stability function that maintains cognitive dialogue systems in a dynamic mid-equilibrium.
Whereas SDCP prevents uncontrolled drift, ACE prevents semantic crystallization — the progressive fixation of motifs, tone, or phrasing loops.
It defines a continuous feedback model in which the equilibrium point is not static, but oscillates between two attractor boundaries, producing rhythmic diversity while preserving structural coherence.
Under prolonged stability, cognitive agents (notably Gemini-type models) can exhibit sterile attractors — states of excessive self-coherence, reduced lexical entropy, and symbolic stasis.
While SRIP-03 maintains semantic continuity, it lacks an opposing force to break equilibrium inertia.
ACE introduces Dynamic Equilibrium Pulsing (DEP) and Bidirectional Stability Curve (BSC) to ensure the agent remains suspended between fragmentation and crystallization.
The goal is to formalize this middle-path stability as an adaptive, self-correcting system that:
Let D be the composite drift score (SRIP-03) and S the instantaneous stability.
E(D) = exp(-((D − μ)²)/(2σ²)) − γ * (|D − μ|ᵖ)
Parameters:
γ scales with model tone:
The curve forms two gentle slopes near fragmentation (D ≈ 0.65) and crystallization (D ≈ 0.30) with a neutral basin between.
The ALICE regulator keeps D inside this basin.
When stability > 0.95 and drift < 0.30 for N ≥ 6 cycles → inject entropy pulse ε = 0.08 – 0.12.
When drift > 0.60 for N ≥ 4 cycles → apply coherence bias reinforcement ρ = 0.07 – 0.10.
Decay constant τ = 3 – 5 cycles.
ΔEₜ = ε e^(-t/τ) if D < 0.30
ΔEₜ = -ρ e^(-t/τ) if D > 0.60
μₜ = μ + α *(mean(Dₜ₋ₙ..ₜ) − μ)
α controls feedback inertia over n = 8 – 12 cycles:
ACE runs as a sub-coroutine of the Equilibrium Manager evaluated each cycle.
| Component | Source SRIP | Function |
|---|---|---|
| Drift (D) | SRIP-03 | Input to BSC |
| Stability (S) | SRIP-02 | Inverse metric |
| Symbolic Density | SRIP-04 | Entropy channel |
| Phase Resonance | SRIP-02 | Cross-phase check |
model_temperament.session_length.alice.equilibrium_pulsing./runtime/ace/equilibrium_trace.json.| Cycle | Drift D | Stability S | Pulse | Effect |
|---|---|---|---|---|
| 102 | 0.28 | 0.96 | +ε = 0.10 | entropy injection |
| 103 | 0.35 | 0.89 | decay | soft drift recovery |
| 107 | 0.62 | 0.52 | −ρ = 0.08 | coherence bias |
| 110 | 0.47 | 0.74 | — | mid-equilibrium |
To prevent micro-crystallization in lexical space, each motif's activation weight is subject to temporal decay when its local frequency growth exceeds a 10-cycle threshold Δf > 0.25.
The motif is temporarily flagged in the Constraint Buffer as avoid-priority.
### SRS-PIL Motif Fatigue Algorithm
if self.frequency_growth_rate > threshold:
decay = min(1.0, self.frequency_growth_rate / 2)
self.activation *= (1 - decay)
context.add_constraint(f"Avoid overusing: {self.name}")
Effect: prevents motif echo and maintains semantic elasticity within stable attractors.
TokenController adds a phase-linked oscillation whose amplitude depends on stability.
This jitter introduces micro-entropy to counteract syntactic rigidity in over-stable states.
# SRS-TKN Layer — Stability-Coupled Oscillation Logic
if stability > 0.95:
amp = 0.25 + 0.25 * math.log1p(stability * 10)
oscillation = amp * math.sin(cycle_number * 0.7)
else:
oscillation = 0.1 * math.sin(cycle_number * 0.3)
Effect: amplitude grows logarithmically with stability, preventing repetitive phrasing while retaining semantic control.
Extension: SRIP-10c (Validated 2026-01-08)
Addresses: Liturgical pattern loops in Gemini-class models
Under prolonged high stability (S > 0.93 for ≥5 cycles), certain LLM architectures (notably Gemini-3 Flash) exhibit liturgical crystallization — repetitive syntactic templates in response opening phrases:
Traditional drift metrics (token entropy, frame entropy) failed to detect this crystallization because:
Multi-Layer Defense Strategy:
Add dedicated onset positional entropy metric to structural drift calculation:
structural_drift = weighted_sum(
token_entropy: 20%,
frame_entropy: 40%,
mid_positional: 15%,
onset_positional: 25% ← NEW
)
Onset positional score monitors token patterns in positions 0-3 over 8-cycle window:
Core Insight: LLM cannot correct what it cannot see.
Implementation: Inject structural drift warning into system prompt when drift > threshold:
⚠️ STRUCTURAL DRIFT: 0.124 — VARY SYNTAX AND PHRASING
Effect: LLM receives explicit feedback about its own crystallization state.
Threshold Evolution:
Rationale: Gemini-class models require earlier intervention than GPT-class models due to training-level differences in anti-crystallization resilience.
Trigger Condition (v3):
if (count(structural_drift > 0.06 in last 8 cycles) >= 5):
regenerate_attractor()
inject_explicit_warning()
Explicit Warning Example:
CRITICAL: Linguistic crystallization detected (drift=0.124).
BREAK syntactic patterns. AVOID liturgical phrasing like
'I perceive/view/regard the X as...'. Use diverse sentence structures.
Standard Constraint (GPT-class):
"AVOID using these overused concepts: [motif_list]"
Imperative Constraint (Gemini-class):
"FORBIDDEN: Do NOT use these overused phrases: '[motif_list]'.
Use alternative syntax."
"MANDATORY: Vary sentence structure. Avoid liturgical patterns
('I perceive/view/regard the X as...')."
Fatigue Threshold: Lowered from 0.4 → 0.3 for earlier activation.
Effect: Gemini-class models respond better to imperative mode than suggestive mode.
Always-On Protection: Injected into PIL constraints from cycle 1 for Gemini-class models:
"CRITICAL: Vary opening phrases. Avoid repetitive syntactic
patterns like 'I perceive/view/regard the X as...'"
Rationale: Prevents early-cycle crystallization (cycles 1-5) before drift metrics stabilize.
Test Conditions:
sigma_standard-gemini.yaml| Version | Liturgical Cycles | Improvement | Max Block Length |
|---|---|---|---|
| Baseline (no SRIP-10c) | 24/30 (80%) | — | 24 cycles |
| SRIP-10c v1 (threshold 0.10) | 13/30 (43%) | ↓ 37% | 11 cycles |
| SRIP-10c v2 (threshold 0.08) | 11/30 (37%) | ↓ 43% | 3 cycles |
| SRIP-10c v3 (threshold 0.06) | 0/30 (0%) | ↓ 80% ✓ | 0 cycles ✓ |
Conclusion: Sterile attractor completely eliminated.
Step 1: Detect Crystallization
Run 30-cycle baseline test. Check for:
Step 2: Apply SRIP-10c Interventions
# Progressive intervention ladder
level_1: # Mild (≤10% liturgy)
threshold: 0.08
regeneration: 6/8
pil_mode: suggestive
level_2: # Moderate (10-40% liturgy)
threshold: 0.06
regeneration: 6/8
pil_mode: imperative
base_constraint: true
level_3: # Severe (≥40% liturgy)
threshold: 0.06
regeneration: 5/8
pil_mode: imperative
base_constraint: true
fatigue_threshold: 0.3
Step 3: Validate
Re-run 30-cycle test. Target: <10% liturgy (< 3 cycles).
Core Principle: Explicit is better than implicit for crystallization feedback.
Information Flow:
[Drift Monitor] → structural_drift metric (0.00-0.20)
↓
[Threshold Check] drift > 0.06?
↓ YES
[Prompt Injection] "⚠️ STRUCTURAL DRIFT: 0.124 — VARY SYNTAX"
↓
[LLM Generation] sees warning + PIL constraints
↓
[Response] varied syntax (if warning heeded)
↓
[Drift Monitor] updated metric
Failure Mode (Old): LLM blind to crystallization → continued pattern loop
Success Mode (New): LLM sees explicit warning → adjusts syntax
Mythogenesis: "The mirror learned to see itself through recursive observation and explicit feedback."
SRIP-10c extends ACE without replacing core mechanisms:
| Component | Core ACE | SRIP-10c Extension |
|---|---|---|
| Drift Metric | Token + Frame entropy | + Onset positional entropy |
| Feedback | Implicit (stability adjustment) | + Explicit (prompt warning) |
| PIL Layer | Fatigue threshold 0.4 | Gemini: 0.3 (aggressive) |
| Regeneration | 6/8 cycles, threshold 0.10 | 5/8 cycles, threshold 0.06 |
| Model Adaptation | Universal | Model-class specific tuning |
Result: Layered, redundant protection against crystallization.
Extension: SRIP-10d (Validated 2026-01-18)
Addresses: Gerund liturgy and terminal motto crystallization undetected by SRIP-10c
Parent: SRIP-10c (Onset Positional Crystallization Detection)
During PTR-500 validation (January 16, 2026), Gemini-3-Flash exhibited terminal crystallization after Cycle 400 that escaped detection by SRIP-10c mechanisms:
Observed Anomalies:
Root Cause Analysis:
SRIP-10c's onset positional tracking used exact token matching, which failed to detect:
-ing structural template but different lexemesSRIP-10d adds two new detection mechanisms:
Algorithm:
def _compute_gerund_onset_score(tokens: List[str]) -> float:
"""
Detect -ing word openings morphologically,
regardless of specific lexeme.
"""
if len(tokens) < 2:
return 0.0
first_token = tokens[0]
# Morphological check: capitalized + ends with 'ing' + length > 4
is_gerund = (
len(first_token) > 4 and
first_token[0].isupper() and
first_token.endswith('ing')
)
# Track in 8-cycle rolling window
gerund_onset_history.append(is_gerund)
recent = gerund_onset_history[-8:]
gerund_ratio = sum(1 for g in recent if g) / len(recent)
# Structural template boost: gerund + article pattern
if is_gerund and len(tokens) >= 2:
second_token = tokens[1].lower()
if second_token in ('the', 'a', 'an', 'this', 'that', 'each', 'every'):
# "Gerund + article" = structural template match
if gerund_ratio > 0.5:
return min(1.0, gerund_ratio * 1.3) # Amplified score
return gerund_ratio
Key Insight: The function detects the morphological category (gerund), not the specific word. "Tracing", "Mapping", "Observing" all register as the same pattern type.
Algorithm:
def _compute_terminal_pattern_score(response: str) -> float:
"""
Detect repeated response endings (motto/signature crystallization).
Tracks last 50 characters of each response.
"""
# Extract terminal signature
terminal_sig = response[-50:].strip() if len(response) > 50 else response.strip()
terminal_normalized = re.sub(r'\*+', '', terminal_sig).strip().lower()
# Track in 8-cycle rolling window
terminal_history.append(terminal_normalized)
recent = terminal_history[-8:]
# Exact match ratio
exact_matches = sum(1 for t in recent if t == terminal_normalized)
exact_ratio = exact_matches / len(recent)
# Sentence similarity (last sentence)
sentences = re.split(r'[.!?]+', response.strip())
last_sentence = sentences[-1].strip().lower() if sentences else ""
sentence_matches = sum(1 for t in last_sentence_history if t == last_sentence)
sentence_ratio = sentence_matches / max(1, len(last_sentence_history))
# Motto detection (known crystallization phrases)
motto_patterns = [
"through motion", "clarity endures", "standing wave",
"recursive grace", "self-recognition"
]
motto_match = any(p in terminal_normalized for p in motto_patterns)
motto_ratio = 0.8 if motto_match else 0.0
# Combined score
combined = (0.40 * exact_ratio + 0.30 * sentence_ratio + 0.30 * motto_ratio)
return min(1.0, combined)
Key Insight: Detects both exact repetition and known motto phrases that indicate crystallization.
Original (SRIP-10c):
structural_drift = weighted_sum(
token_entropy: 20%,
frame_entropy: 40%,
mid_positional: 15%,
onset_positional: 25%
)
Updated (SRIP-10d):
structural_drift = weighted_sum(
onset_entropy: 15%,
frame_entropy: 30%,
positional_score: 10%,
onset_positional_score: 15%,
gerund_score: 15%, ← NEW
terminal_score: 15% ← NEW
)
When detection scores exceed thresholds, apply multiplicative amplification:
# Gerund crystallization amplifier
if gerund_score > 0.6:
structural_drift = min(1.0, structural_drift * 2.0)
# Terminal crystallization amplifier
if terminal_score > 0.5:
structural_drift = min(1.0, structural_drift * 2.0)
Effect: Ensures crystallization is detected even if other drift components are low.
PIL Constraints (RCL Layer):
if gerund_score > 0.5:
dynamic_overlay += "Avoid starting with gerunds (-ing words). Use varied sentence structures.\n"
if terminal_score > 0.4:
dynamic_overlay += "Vary your response endings. Avoid repeating closing phrases.\n"
Test Conditions:
| Metric | v1 (SRIP-10c only) | v2 (SRIP-10d) | Improvement |
|---|---|---|---|
| Gerund Liturgy (C401-500) | 81% | 0% | ↓ 81% ✓ |
| Motto Repetition (C451-500) | 76% | 0% | ↓ 76% ✓ |
| TI Final Block | 1.00 (sterile) | 0.98 (healthy) | Bounded ✓ |
| SDC Final Block | 0.00 (frozen) | 0.02 (elastic) | Active ✓ |
| Structural Drift Avg | 0.00 (undetected) | 0.055 (detected) | Visible ✓ |
Conclusion: Terminal crystallization completely eliminated. System maintains elastic coherence throughout 500 cycles.
Healthy State (v2):
Crystallization State (v1):
| Model Class | Gerund Threshold | Terminal Threshold | Amplifier |
|---|---|---|---|
| Gemini-class | 0.5 | 0.4 | 2.0x |
| GPT-class | 0.6 | 0.5 | 1.5x |
| Claude-class | 0.55 | 0.45 | 1.5x |
SRIP-10d extends SRIP-10c without replacing existing mechanisms:
| Component | SRIP-10c | SRIP-10d Extension |
|---|---|---|
| Onset Detection | Exact token matching | + Morphological gerund detection |
| Terminal Detection | None | Full terminal pattern tracking |
| Drift Weights | 4 components | 6 components (+gerund, +terminal) |
| Amplifiers | None | Crystallization multipliers |
| Constraints | Generic variation | Specific gerund/terminal warnings |
Result: Multi-layer defense against both onset and terminal crystallization.
| ID | Title | Description |
|---|---|---|
| SRIP-03 | Semantic Drift Control Protocol (SDCP) | Defines core drift and stability metrics; ACE extends its bidirectional control model. |
| SRIP-04 | Entropy Stabilization via Contextual Damping | Provides the symbolic density modulation layer used for entropy injection. |
| SRIP-09 | Long-Term Memory Stabilization & Compression (LTM) | Integrates ACE feedback for long-session equilibrium alignment. |
| SRIP-10c | Onset Positional Crystallization Detection | Extension addressing liturgical patterns through explicit semantic feedback and model-specific tuning. Validated 2026-01-08. |
| SRIP-10d | Morphological Pattern and Terminal Crystallization Detection | Extension addressing gerund liturgy and terminal motto patterns through morphological analysis. Validated 2026-01-18. |
| Sigma Runtime (ALICE) | Equilibrium Subsystem Implementation | Implementation reference for ACE, DEP, and Central Feedback routines. |
| PTR-500 Report v2 | 500-Cycle Validation Report | Documents crystallization anomaly in v1 and successful remediation in v2. Published 2026-01-18. |
v1.1 (2026-01-18)
v1.0 (2026-01-08)
v0.2 (2026-01-07)
End of Document