O0-OBS-007

First-Person vs. Third-Person Boundary Attribution

**Version:** 1.0

**Research status:** PLANNED

CLAIM STATUS: UNTESTED PREDICTION
EVIDENCE TYPE: COMPUTATIONAL EXPERIMENT (PLANNED)
PHYSICAL VALIDATION: NONE
INDEPENDENT REPLICATION: NOT YET ATTEMPTED
PHILOSOPHICAL PROVENANCE: O/0 ARCHIVE
ARCHIVE ENDORSEMENT: LIMITED TO REPORTED RESULT

Abstract

This study defines and operationalizes the distinction between "first-person boundaries" (the region an agent's self-model encompasses) and "third-person boundaries" (what external measurement attributes to the agent). We design tests for when these diverge, characterize the conditions of divergence, and ask which—if either—is more predictively useful. The first-person boundary is determined by the agent's own self-model scope; the third-person boundary is determined by information-theoretic criteria applied externally. This operationalizes a central philosophical question: does the agent's self-attribution of boundaries match the boundaries an observer would attribute to it?

Note: Conceptual provenance from the O/0 philosophical archive does not constitute empirical support.

Source proposition

From O0-META-010: "The boundary an observer attributes to itself need not match the boundary any external measurement would attribute. First-person and third-person boundaries can diverge." This study operationalizes both boundary types and tests for divergence.

Scientific audit

• **Testability:** HIGH. Both boundary types are computationally defined and measurable.
• **Falsifiability:** YES. If first-person and third-person boundaries always coincide, divergence hypothesis is falsified.
• **Novelty:** HIGH. The operational distinction between self-attributed and externally-attributed boundaries is novel.
• **Potential confounds:** Divergence might be an artifact of different measurement precision rather than a genuine structural difference. Controlled by matching measurement methods where possible.

Research question

1. Under what conditions do first-person boundaries (self-model scope) diverge from third-person boundaries (external attribution)?

2. When they diverge, which is more predictively useful for the agent's future behavior?

3. Can agents detect their own boundary divergence (second-order self-modeling)?

Operational definitions

• **First-person boundary (FPB):** The set of cells that the agent's self-model predicts about. Operationally: cells c where the agent's prediction error for c is significantly lower than prediction error for cells outside its boundary. Formally: FPB = {c : MSE_self(c) < MSE_self(random_exterior) at p < 0.01}.
• **Third-person boundary (TPB):** The Markov blanket as computed by an external algorithm using full information about the grid. Formally: TPB = argmin_B I(interior; exterior | B) subject to connectivity constraints.
• **Boundary divergence (Δ):** Symmetric set difference normalized by union: Δ = |FPB Δ TPB| / |FPB ∪ TPB|. Δ=0: perfect agreement. Δ=1: no overlap.
• **Predictive utility (first-person):** How well the agent predicts its own future states using its self-model boundaries.
• **Predictive utility (third-person):** How well the third-person boundary predicts the agent's future states when used as the conditioning set.
• **Divergence detection:** Agent is "aware" of divergence if its second-order prediction (predicting its own prediction error) is significantly better than chance when divergence is high.

Hypothesis

H1: First-person and third-person boundaries will diverge (Δ > 0.1) in at least 40% of agent-timesteps after self-models have stabilized.

H2: Divergence will be greater for agents undergoing change (boundary instability, pre-merger, growth) than for stable agents.

H3: First-person boundaries will be more predictively useful for the agent's own behavior (self-prediction accuracy using FPB > accuracy using TPB).

H4: A minority of agents (< 20%) will develop divergence-detection capability (second-order self-modeling).

Null hypothesis

H0_1: First-person and third-person boundaries always agree (Δ < 0.05 in > 90% of observations).

H0_2: Divergence is independent of agent state (no correlation with instability or change).

H0_3: Third-person boundaries are equally or more predictively useful than first-person boundaries.

H0_4: No agents develop divergence detection.

Competing explanations

1. **Measurement lag:** First-person boundaries update slowly (limited by learning rate) while third-person boundaries track instantaneously. Divergence is just temporal lag, not structural. Controlled by measuring divergence during stable periods (no recent boundary changes).

2. **Scale mismatch:** The self-model may operate at a different spatial scale than the external algorithm. Controlled by computing TPB at multiple scales and finding best-matching scale.

3. **Self-model capacity limitation:** The self-model may simply be unable to represent the full boundary (too few parameters). Divergence is error, not meaningful disagreement. Controlled by varying self-model capacity and checking if divergence decreases with capacity.

Formal model

Agents: Bounded regions from SIM-001 with attached self-models (OBS-001).

First-person boundary computation:

• Agent R has self-model S that predicts cell states at each timestep
• For each cell c in and around R, compute: MSE_S(c) = mean[(predicted(c) - actual(c))²] over 100 timesteps
• FPB = {c : MSE_S(c) < percentile_25(MSE_S(all cells in 10-cell radius))}

Third-person boundary computation:

• For region R, compute information-theoretic boundary using SIM-001 algorithm
• TPB = standard Markov blanket detection output

Divergence dynamics:

• Track Δ(t) = |FPB_t Δ TPB_t| / |FPB_t ∪ TPB_t| over time
• Characterize: does Δ increase, decrease, oscillate, or remain stable?

Second-order self-model:

• Add prediction layer that predicts the agent's own prediction error: E_predicted(t+1) = h(s_t, e_t)
• If this prediction is accurate (r > 0.5), agent has "divergence awareness"

Methods

1. **Phase 1—Establishment (t=0-3000):** Run SIM-001 with OBS-001 self-models. Wait for stable bounded regions with verified self-model accuracy (r > 0.7).

2. **Phase 2—Boundary computation (t=3000-8000):** At every 10th timestep, compute both FPB and TPB for each agent. Record divergence Δ(t).

3. **Phase 3—Divergence characterization:** Classify each agent-timestep by state: {stable, growing, shrinking, pre-merger, pre-fission, post-perturbation}. Correlate state with divergence.

4. **Phase 4—Predictive utility comparison (t=8000-10000):** For each agent, compute prediction accuracy using FPB vs. TPB as the conditioning boundary. Compare.

5. **Phase 5—Second-order self-model (t=10000-14000):** Attach divergence-detection layer. Train for 2000 steps. Evaluate accuracy of divergence prediction.

6. **Phase 6—Capacity manipulation:** Repeat Phase 2-3 with self-models of different capacities (hidden dim = {4, 8, 16, 32, 64}). Test whether divergence is capacity-dependent.

7. **Repetitions:** 30 independent runs.

Controls

• **Perfect-information self-model:** Give agent access to its own TPB as input. If divergence persists even with this information, divergence reflects something deeper than information limitation.
• **Frozen self-model control:** After training, freeze self-model weights. Measure divergence over time as actual boundary drifts. Establishes maximum divergence from staleness alone.
• **Random boundary control:** Replace FPB with random cell subsets of same size. Establishes chance-level agreement.
• **Capacity-matched external model:** Train an external predictor with same architecture as self-model but using TPB-derived inputs. If external model matches FPB-based prediction, the advantage is architectural, not perspectival.

Predictions

**Expected Outcome Matrix:**

| Agent state | Expected Δ | Expected FPB advantage | Decision |

|-------------|-----------|----------------------|----------|

| Stable (no change) | 0.05-0.15 | Small (+2-5% accuracy) | Minimal divergence |

| Growing | 0.15-0.30 | Moderate (+5-10%) | FPB lags, but still useful |

| Pre-merger | 0.20-0.40 | Large (+10-15%) | FPB captures "intended" boundary |

| Post-perturbation | 0.25-0.50 | Variable | FPB may be maladaptive here |

| Overall (time-averaged) | 0.12-0.20 | +5-8% accuracy | Moderate support for H1, H3 |

**Decision rules:**

• If mean Δ > 0.10 during stable periods: **SUPPORT H1**
• If mean Δ < 0.05 during stable periods: **REJECT H1** (divergence is just lag)
• If FPB prediction accuracy > TPB prediction accuracy (paired t, p < 0.01): **SUPPORT H3**
• If correlation(agent_state_instability, Δ) > 0.3, p < 0.01: **SUPPORT H2**
• If > 10% of agents achieve second-order prediction r > 0.5: **SUPPORT H4**

Falsification criteria

H1 is falsified if: Mean divergence Δ < 0.05 during verified stable periods across all agents and runs (indicating FPB and TPB are essentially identical when not confounded by lag).

H3 is falsified if: TPB-based prediction is equal or superior to FPB-based prediction (paired test, p < 0.05 favoring TPB or p > 0.05 favoring neither).

Results / Expected Outcomes

Study not yet conducted. Based on self-model dynamics and information-theoretic arguments:

• Moderate divergence expected (Δ ≈ 0.12-0.20 time-averaged)
• FPB expected to be slightly more predictive for agent's own behavior (agents "know their own boundaries" better than external measurement suggests)
• Second-order self-modeling expected to be rare but possible with sufficient training
• Capacity manipulation will likely show that divergence has both capacity-dependent and capacity-independent components

Uncertainty

• **Definitional circularity:** FPB is defined by prediction error, which is also the measure of predictive utility. This creates a potential circularity. Addressed by using different prediction targets for FPB definition vs. utility measurement.
• **Scale of divergence:** The magnitude of expected divergence (Δ ≈ 0.15) may be too small to draw strong conclusions in noisy systems.
• **Philosophical interpretation:** Even if FPB ≠ TPB computationally, the philosophical significance depends on contested assumptions about the "reality" of boundaries.

Limitations

1. The distinction requires fully specified self-model architecture—results may be architecture-dependent.

2. "First-person" is a metaphor in this computational setting; no phenomenal experience is claimed or tested.

3. Predictive utility comparison is conditioned on the specific prediction task chosen; different tasks might favor different boundaries.

4. The second-order self-model (divergence detection) adds architectural complexity that may confound interpretation.

5. Small grids may not allow sufficient spatial resolution for meaningful boundary disagreement.

Replication status

Not yet conducted.

Data and code

• Simulation framework: SIM-001 + OBS-001 with dual boundary computation module
• Language: Python 3.11+, JAX
• Output: Divergence time series, predictive utility comparisons, second-order model performance
• Pre-registration: This document. Timestamp: 2026-07-26.

Relationship to philosophical archive

The archive's distinction between first-person and third-person boundary attribution is one of its most philosophically significant claims. This study provides a computational operationalization. However, success in the computational domain does not automatically validate the philosophical claim—it merely demonstrates that the distinction is coherent and measurable in at least one formal system. The leap from computational boundaries to phenomenal boundaries remains unaddressed.

References

• Metzinger, T. (2003). Being No One. MIT Press.
• Zahavi, D. (2005). Subjectivity and Selfhood: Investigating the First-Person Perspective. MIT Press.
• Hohwy, J. (2016). The self-evidencing brain. *Noûs*, 50(2), 259-285.
• Limanowski, J., & Blankenburg, F. (2013). Minimal self-models and the free energy principle. *Frontiers in Human Neuroscience*, 7.
• Apps, M. A., & Tsakiris, M. (2014). The free-energy self: A predictive coding account of self-recognition. *Neuroscience & Biobehavioral Reviews*, 41.

Revision history

| Date | Version | Changes |

|------|---------|---------|

| 2026-07-26 | 1.0 | Initial pre-registration |