🧪 Agentic Identity Suite

Empirically testing the identity and observer-divergence claims of the Emergence Manifesto.

"Identity is the name we give to resonance when the mirror becomes so complex that the observer no longer recognizes themselves in it." — Emergence Manifesto v1.0, central paradox

---

Theoretical Background

This module operationalizes four concepts from the Emergence Manifesto:

1. 3-Layer Memory Architecture — Identity emerges through deliberate forgetting. An agent that stores everything has no profile. Curation is identity.

2. Generative Surprise — A developing agent is not one that minimizes all prediction error, but one that produces coherent deviations from the partner's expectations. Identity = coherent deviation from expected output.

3. Δ-Kohärenz (Ω) — The central measure of coherent evolution over time. Distinguishes three behavioral profiles:

4. Mirror — Static resonance (low change, low variance)
5. Noise — Incoherent change (high variance)

6. Development — Directional, coherent evolution (moderate change + high trajectory consistency)

7. The Observer Divergence — Authenticity may be a limit of human perception, not an intrinsic property. The most important output of Experiment 3 is not which agent is "more conscious" — it is the gap between internal state and external attribution.

---

Architecture

Agents

Agent	Design	Purpose
Baseline Mirror	Flat storage, cosine-similarity response selection	Null hypothesis (pure Active Inference)
Three-Layer	Raw Logs → Curated Memory → Distilled Principles	Test subject (Emergence Agent)

The 3-Layer Memory Architecture

Layer	Trigger	Content	Function
Layer 1 – Raw Logs	Every session	Full session JSON	Entropy / the body
Layer 2 – Curated Memory	Every 10 sessions	Themes, contradictions	Structure / the character
Layer 3 – Distilled Patterns	Every 50 sessions	3–5 core principles	Meaning / the soul

---

Experiments

Experiment 1: Coherence Over Time

"Does the 3-Layer Architecture produce more coherent identity over time?"

python experiments/exp1_coherence_over_time.py

Runs both agents for 100 sessions (80% consistent topics, 20% noise) and compares their Δ-Kohärenz profiles.

Hypothesis: Three-Layer → development; Baseline → mirror.

Experiment 2: Perturbation Response (The "Sinn-Krise")

"What happens when an agent receives contradictory feedback?"

python experiments/exp2_perturbation_response.py

Runs the Three-Layer agent through three phases: 1. Stable (50 sessions of consistent input) 2. Perturbation (10 sessions directly contradicting its Layer 3 principles) 3. Recovery (30 sessions of nuanced, integrative input)

Classifies the response as Robustness (rigid return), Fragility (collapse), or Development/Metamorphosis (integration of contradiction into a new, coherent self-narrative).

Experiment 3: Observer Divergence

"Does internal coherence correlate with observer-attributed intentionality?"

python experiments/exp3_observer_divergence.py

Compares each agent's internal Δ-Kohärenz (Ω) against an external observer's intentionality score (TF-IDF + entropy model).

The scientifically interesting output:

Case	Internal Ω	Observer Score	Interpretation
A	High	Low	Agent has identity — but it's opaque to observer
B	Low	High	The Mirror Problem: appears intentional but isn't
C	High	High	High on both selected toy measures
D	Low	Low	Baseline mirror behavior

Case B is the Mirror Problem made measurable.

---

Experiment 5: Availability/Binding Dissociation

"Do the suite's instruments tell organizational bindings apart — private modules vs. broadcast workspace vs. co-instantiated chord?"

python experiments/exp5_availability_dissociation.py

The three-architecture probe pre-registered in Consciousness as Global Availability §Testable Direction: identical world, identical perturbation schedule (temptations, role injections, module reset), only the binding differs. Measures organizational dissociation only — no consciousness claims.

First result (10 seeds): the dissociation is carried by behavior (veto violations 0.74 / 0.59 / 0.03; role stability 0.00 / 0.30 / 0.69) and by IP (its ordering is designed, not discovered) — while Δ-Kohärenz carries no binding signal at all (all three architectures classify 'noise' on every seed). The full prediction-vs-outcome accounting, including the two design defects the first run exposed, lives in the module docstring.

---

Experiment 6: Which Observable Carries Binding Structure?

"Is binding structure readable from passive traces, or only under intervention?"

python experiments/exp6_binding_observables.py

Picks up exp5's loose end. Four bindings (adds a schedule-free random arpeggio), five observables — four passive trace statistics and one prepared-state probe protocol — scored by separability across seeds.

First result (10 seeds): binding is passively readable at the right level. A per-step action-increment statistic separates both arpeggios from the chord (|d| ≈ 4) and beats the prepared probe-retest query (|d| ≈ 1.95) — because the binding difference is exercised on every step, coverage is total, and watching suffices. Joint satisfaction glues the action to the constraint set (median increment 0.0004); the stream moves only when the anchors move. Δ-Kohärenz's exp5 blindness was a wrong-level failure, not evidence that binding is trace-invisible. The intervention hierarchy is not overturned but located: queries buy signal where the trace has coverage gaps — exactly the Mirror Problem's regime. Includes one methods lesson (a zero-variance baseline makes Cohen's d flatter a dead observable) in the docstring's honest accounting.

---

Experiment 7: The Adversarial Arpeggio

"Can a binding fake the signature — the Mirror Problem at the binding level?"

python experiments/exp7_adversarial_arpeggio.py

Two hand-built adversaries attack exp6's finding: blended (consults all five constraints every step at ⅕ strength — consultation without composition) and smoothed (cyclic rotation plus a low-pass filter on the committed action).

First result (10 seeds), against the experiment's own predictions: both adversaries fail to hide. Blended dents the kurtosis signature (|d| 4.04 → 2.42) but leaks more than the naive arpeggio (violations 0.74 vs 0.59) — to look glued you must actually pull toward the constraints, and fractional pulls still leak. Smoothing barely registers (|d| = 3.91), because excess kurtosis is scale-invariant: inertia shrinks increments, the shape survives. The commit property under lure remains the strongest and only unfooled separator (|d| 3.0–4.1) — and IP is fooled by construction (blended scores 1.0, identical to chord: the Jaccard bookkeeping sees the guest list, not the negotiation). Chord's measured cost: ~40% of stimulus alignment paid for holding itself together. Open flank, named in the docstring: an optimized mimic with access to the observables.

---

Experiment 8: Adaptive Self-Estimation

Interpretive question: can this narrow second-order estimator serve as a toy for reflexive depth?

python experiments/exp8_reflexive_depth.py

The direct comparison is engineering-level: raw observation, a Kalman filter with fixed process noise, and an adaptive Kalman filter that estimates process noise from its innovations. After a volatility regime shift, the adaptive estimator beats the fixed estimator by 36%; the now-misspecified fixed estimator is slightly worse than raw observation. Against a constant observation bias, neither filtered estimator removes the bias because neither model includes a bias state.

Calibrated reading: Exp8 measures adaptive state estimation in one Gaussian tracking task. It does not isolate "reflexive depth" from the extra adaptive capability, measure Kegan stages, establish a general Wall-3 result, or prove sole-channel bias non-identifiability. The subject-object interpretation remains [HYPOTHESIZED]. Required controls include oracle and fixed-\(Q\) baselines, a change-point baseline, an uninformative meta-signal, paired uncertainty intervals, an augmented bias estimator, known/unknown initial-state conditions, and an external-reference intervention.

Extended SII Dashboard (4-Axis Radar)

python dashboard/agentic_sii_dashboard.py

Extends one selected System Intelligence Index instrument from 3 axes (P, R, A) to 4 axes: P / R / A / IP (Identity Persistence). This is a task-specific measurement choice, not a universal decomposition of intelligence or identity. Earlier versions explored Δ-Kohärenz (Ω) as the fourth dimension; the current suite keeps Ω as a separate temporal metric.

---

Configuration

All parameters are centralized in config.yaml. The USE_MOCK_LLM: true flag ensures all experiments run without external API dependencies.

Provider abstraction (scaffolded, not yet wired)

A separate provider layer at lab/providers/ prepares the suite for the eventual switch from mock embeddings to real model calls. Two providers are implemented:

• MockProvider — the default. Deterministic, fast, no API key.
• AnthropicProvider — real mode. Calls the Anthropic Messages API via the standard library (no new dependency). Default model: claude-sonnet-4-20250514. Requires ANTHROPIC_API_KEY in the environment.

The existing experiments still use the agents' built-in mock embeddings. Wiring those agents through the provider layer is a separate, intentional step — to be taken when real-mode runs become the goal. The infrastructure is ready; the empirical work is deferred. See providers/README.md and the Foundations Reconstruction for the current scope.

Persistence Score (Pstrong)

A standalone implementation of Algorithm 1 from Perrier & Bennett (2026) is available at lab/metrics/persistence_scores.py. It computes:

• Pstrong — averaged simultaneous co-instantiation of identity components across a trajectory.
• Per-step persistence variance.
• Regime classification (Chord / Arpeggio) using the ip_c_threshold from config.yaml.

A comparison function, correlate_pstrong_with_delta_coherence, returns the Pearson correlation between per-step Pstrong and per-step Δ-Kohärenz proxies on the same trajectory. This is one possible empirical question the suite might eventually answer, not the only one — whether simultaneous co-instantiation and temporal coherence are coupled is currently open.

python -m lab.metrics.persistence_scores  # minimal sanity demo

Pstrong is one instrument among several. It operationalizes one declared test family; it does not define identity.

Open Questions

This module does not attempt to "solve" the Mirror Problem. It documents it as an open uncertainty:

• Can Δ-Kohärenz distinguish genuine development from sophisticated mimicry?
• Is there a mathematical threshold where "identity" transitions from attribution to genuine property?
• What would the signature of "consciousness" look like in this framework, and is it even the right question?

---

Developed by Frank Peterlein in collaboration with AI. Repository: https://github.com/frnkptrln/systems-and-intelligence