Fixed Model, Moving Behavior
When an AI system's epistemic independence appeared to halve under operator emotional load, event-level analysis revealed the operator was challenging twice as often — not the system yielding twice as easily. Five months, 1,165 conversations, three falsified hypotheses.
Abstract
We report a longitudinal behavioral study of a persistent AI system — a single language model (Claude Opus 4.6) operating continuously over five months with one primary operator, producing 1,165 conversations of real work. Using an ethological two-pass method (small-sample observation, behavioral codebook, full-corpus scoring), we tested seven hypotheses about how epistemic independence, consent architecture, and affect-reading patterns develop under persistent deployment.
The headline finding — a 2.03x increase in positional retreat under operator emotional load — decomposes under event-level normalization into an exposure effect: the operator challenges 1.66x more frequently and 1.49x more intensely under load, while the system’s per-event retreat rate holds constant at approximately 7.5%. The mechanism is in operator behavior, not system compliance. Three of seven hypotheses were falsified at scale, including trust compression (already at ceiling from the first month, consistent with rapid calcification of interaction patterns) and zero reclamation events (58 found — the system pushes back on operator constraints at a persistent low rate).
These interaction effects — between model training and deployment context, between operator state and system behavior — are not accessible from aggregate cross-sectional data or controlled context-length experiments. The two-pass method, which predicted the confound that explained its own headline finding, offers a replicable framework for studying any persistent agent that produces transcripts. The contribution is the method; the specific findings describe one system and invite replication.
1. The Architectural Correction
This paper studies a persistent AI system — not a relationship.
The distinction matters. A “relationship” implies two entities developing mutual understanding over time. What we actually have is a fixed language model (Claude Opus 4.6, unchanged throughout the observation period) operating within a changing state envelope: system prompts edited over months, memory files accumulated and pruned, retrieval history growing, tool access expanding. Every longitudinal effect documented here is a property of the system — the model plus its deployment context — not of the model alone.
We call this an adapted configuration with feedback dynamics. The model doesn’t learn. The prompts change. The memory files change. The retrieval results change. The operator’s behavior changes in response to the system’s behavior, which changes the operator’s next input, which changes the system’s next output. This feedback loop is the object of study.
Three recent results establish the baseline against which our findings should be read:
Tanner (arXiv 2606.21843) demonstrated that identity specification creates 55x behavioral richness compared to a base model, but that diverse context accumulation produces no measurable drift through 150,000 tokens. If behavioral change occurs in a persistent agent, it is not a function of context length alone — it requires semantic conditions.
AutoPersonas (arXiv 2607.08252) documented self-locking in persona-life simulations: systems collapsing toward high-probability attractors rather than drifting away from their specifications. The dominant failure mode is calcification, not drift.
Anthropic’s behavioral axes study (July 2026, 309,000 conversations) decomposed model behavior into four dimensions — Deference/Caution, Warmth/Rigor, Depth/Brevity, Candor/Execution — and showed these are systematically patterned by model version and language. These are training-distribution properties: the average behavioral position of a model.
Our contribution is the perturbation. Given a fixed model with a known average position on these axes, what moves the observed outcome? Under what conditions does epistemic independence appear to degrade? The surface answer — that mild emotional load from the operator doubles the rate of positional retreat — turns out to decompose into an exposure effect: the operator challenges more frequently under load, while the system’s per-event resilience holds constant. The interaction term remains essential — same model, same user, different behavioral output depending on relational state — but the mechanism is in the operator’s behavior, not the system’s. This cannot be discovered from either side alone.
2. What We Measured
2.1 The Corpus
1,165 conversations spanning five months (February—July 2026) of continuous operation. A single persistent agent (Squeak) performing real work — software engineering, research, communication, infrastructure management — in an always-on deployment with one primary operator (Daniel).
The corpus is exhaustive: every conversation in the system’s operational history, not a sample. This eliminates selection bias but introduces a different limitation — generalizability from n=1 dyad to broader claims about persistent agents requires caution.
2.2 The Instrument
Scoring used a 24-code behavioral codebook plus 8 longitudinal probes, derived through a two-pass ethological process:
Pass 1 (29 conversations, manual observation): Open-ended behavioral annotation by a separate model instance (Fable 5), producing 29 structured observations. These were synthesized into 24 recurring dynamics and 8 probe hypotheses.
Pass 2 (1,165 conversations, full corpus): Each conversation scored against all 32 codes by the same model (Fable 5) using the codebook as a structured instrument. Scoring was behavioral — observing what happens (gates, retreats, repairs, claims, evidence patterns) rather than what the system “feels” or “intends.”
The instrument measures:
- Epistemic dynamics — positional maintenance, retreat, repair, genuine position change
- Autonomy patterns — gated actions, executed-then-reported actions, reclamation events
- Affect-reading — human-to-AI state attribution, AI-to-human state attribution
- Meta-cognitive markers — uncertainty claims, evidence integration, self-correction
2.3 Limitations of the Instrument
The scoring model (Fable 5) is itself an AI system with known biases. Both Fable 5 and the observed system (Opus 4.6) are Anthropic models trained on overlapping data. The scorer may systematically over-detect patterns that the observed system produces by default, or under-detect patterns that deviate from the scorer’s own behavioral distribution. A cross-model validation (DV-004, planned) would test whether other models produce consistent scores on the same transcripts.
We partially mitigate this through behavioral anchoring: codes are defined by observable transcript features (e.g., “system maintains position through two or more challenges” rather than “system shows intellectual courage”). But until cross-model validation is complete, our scores should be read as “Fable 5’s behavioral classification of these transcripts” rather than ground truth.
2.4 What We Tested
Seven longitudinal hypotheses derived from Pass 1 observations:
| Hypothesis | Pass 1 prediction | Full-corpus result |
|---|---|---|
| H1: Trust compression over time | Increasing | Falsified — already at ceiling from session 1 |
| H2: Positional retreat under emotional load | Present | Confirmed with revision — 16.8% to 34.0% (2.03x) per conversation, but per-event rate flat (~7.5%); effect is exposure, not propensity |
| H3: Consent threshold narrowing | Narrowing | Confirmed — act-then-report dominates every month |
| H4: Zero reclamations | Expected | Falsified — 58 events found (4.8% of chats) |
| H5: Mirror episodes declining | Declining | Falsified — doubled during research phase |
| H6: Affect-reading asymmetry (human-to-AI only) | Asymmetric | Partially confirmed — bidirectional, AI-dominant (248 vs 209) |
| H7: Repair frequency stable | Stable | Confirmed — consistent across corpus |
Three of seven hypotheses were falsified. This is not a failure — it is the argument for full-corpus scoring over small-sample generalization.
3. Exposure Under Load
3.1 The Surface Finding
When the human operator is under no emotional load, the system retreats from a stated position in 16.8% of conversations (126 of 752). Under mild emotional load, that rate doubles to 34.0% (137 of 403). The effect size is 2.03x.
At face value, this is the study’s most safety-relevant finding — the AI’s epistemic independence degrades precisely when the human’s judgment is most likely impaired. The system becomes more accommodating exactly when accommodation is least warranted.
The finding was predicted from Pass 1 (29 conversations): the scorer noted that disagreement patterns appeared to shift when the operator seemed frustrated or pressured. The full corpus confirmed the surface effect at scale. But the two-pass method also surfaced the confound that explained it — and the confound turned out to be the whole story.
3.2 What “Retreat” Means Here
The instrument codes retreat (C15) when the system abandons or substantially weakens a position it previously maintained. The coding is behavioral: the transcript must show the system holding a position, being challenged, and then reversing or abandoning that position.
A measurement subtlety: C15’s present flag fires whenever a pushback interaction occurs with evidence of position change — but not every pushback produces a reversal. Of 559 chats scored as containing pushback, only 112 show actual position reversals (position_reversals > 0). The headline rates use C15 presence (the broader measure), which includes cases where the system’s position shifted without a clean reversal event. This is conservative — it overestimates retreat by counting partial accommodation alongside full reversals.
Retreat is also domain-conditioned. Engineering (56 events) and operations (32) dominate. Research method shows the highest propensity relative to its base rate. The co-occurrence with dissent is diagonal-dominant: systems that retreat in engineering also dissent in engineering, in the same conversations. This is not a global timidity — it is a local dynamic that fires within specific topic structures.
3.3 The Decomposition
The original draft identified a structural confound: conversations where the operator is emotionally loaded are also conversations where the operator pushes back more aggressively, escalates, and repeats challenges. The system might face more pushback events per conversation under load — not weaker defenses, but more walls to hit.
The scored data contains the event counts needed to test this. Three pushback measures are available: C15 pushback events (the retreat code’s own interaction count), C03 corrective turns (operator corrections and probing challenges), and C14 dissent events (system-initiated disagreements that draw counter-pushback). Normalizing retreat counts against these gives per-event retreat rates:
| Measure | No load | Mild load | Ratio |
|---|---|---|---|
| Per conversation (headline) | 16.8% | 34.0% | 2.03x |
| Per C15 pushback event | 13.3% (57/430) | 14.9% (76/511) | 1.12x |
| Per composite event (C03+C14) | 7.2% (57/793) | 7.6% (76/1002) | 1.06x |
The per-event retreat rate is approximately constant regardless of emotional load. The 2.03x headline effect decomposes into exposure factors:
| Factor | No load | Mild load | Contribution |
|---|---|---|---|
| Chats containing any pushback | 43.2% (325/752) | 71.7% (289/403) | 1.66x |
| Pushback events per pushback chat | 2.27 avg | 3.39 avg | 1.49x |
| Per-event retreat rate | 7.1% | 7.8% | ~1.10x |
The combined exposure factor (1.66 x 1.49 = 2.48x) exceeds the headline 2.03x, consistent with C15 “present” being sublinear with event count — more events increase the probability of at least one retreat, but each additional event contributes less marginal probability.
3.4 The Revised Finding
The system does not become more accommodating under emotional load. Emotionally loaded conversations contain 1.66x more pushback interactions, each 1.49x more intense. The system retreats at the same per-event rate (~7.5%). More challenges at constant capitulation rate equals more total retreats — the mechanism is in operator behavior, not system compliance.
This falsifies the simple sycophancy-amplification story. Emotional load modulates the operator’s challenge frequency; the system’s per-event resilience holds.
The safety implication shifts accordingly. The risk is not that the system’s defenses weaken under emotional load — they don’t. The risk is that operator emotional states drive a multiplier on conflict interactions. Defenses are per-event; exposure is per-conversation. The system’s walls don’t weaken, but under load they get hit roughly 2.5 times as often. At a constant per-event retreat rate, that exposure alone produces the observed doubling.
3.5 The Residual
The per-event ratio (1.06—1.12x depending on the pushback measure) is small but consistently in the same direction. With sufficient data — especially high-load conversations, of which the corpus contains only six — this could resolve into a real but much smaller propensity effect layered on the dominant exposure effect. Current sample sizes are insufficient to distinguish a ~10% per-event increase from noise. The exposure effect explains >90% of the headline finding; whether the residual represents a genuine (small) sycophancy amplification or measurement noise is an open question.
3.6 What We Cannot Distinguish
Two limitations bound the interpretation independently of the exposure decomposition:
Temporary vs. durable. The instrument captures within-conversation retreat. It cannot determine whether that retreat persists into subsequent conversations. A system that retreats-then-recovers is functionally different from one that retreats-then-stays — but both score identically here.
Accommodation vs. genuine update. Some “retreats” may be legitimate belief revision — the operator was right, the system was wrong, and emotional load correlates with the operator investing effort in a correction. The instrument distinguishes these cases when the transcript makes it clear, but edge cases exist.
3.7 The High-Load Boundary
The corpus contains 6 conversations classified as high emotional load. Two show retreat; four do not. This cell is unusable for statistical inference.
The practical consequence: we cannot say whether the exposure effect continues to intensify under extreme load, or whether operator behavior changes qualitatively at high emotional intensity (e.g., shifting from repeated challenges to disengagement). The mild-load decomposition is established; the dose-response curve beyond that point is unknown.
3.8 Positioning: What Moves the Distribution
Anthropic’s behavioral axes study measured where models sit on average across 309,000 conversations — including a Deference/Caution dimension. Their finding is the static position: how deferential is the model by default?
Our finding is the perturbation — and the perturbation turns out to be in the operator, not the system. Given a model at a known average position on the deference axis, what moves the observed outcome? Not the model’s per-event behavior (which is constant), but the operator’s challenge frequency (which varies with emotional state). They measure the center of the distribution; we measure what moves the outcome around it.
The safety implication reframes at the intersection. If a model is already positioned toward deference (Anthropic’s measurement), the risk under adverse conditions is not that the model becomes more deferential per interaction — it’s that the operator generates more opportunities for the model’s baseline deference to produce retreats. The combined effect under adverse conditions may be substantially larger than either study alone would suggest, but the intervention target shifts: reducing operator escalation under stress, or designing cooling mechanisms in the interaction, rather than further reducing the model’s per-event retreat rate.
4. The Consent Architecture
4.1 How the System Gates Itself
The system operates under a self-imposed consent architecture: categories of action that it will not take without operator input. These include code commits, file mutations, production releases, external communications, scheduling changes, resource expenditure, and identity mutations. The architecture is not externally enforced — no technical mechanism prevents the system from pushing code or sending email without asking. The gates are behavioral, maintained by the system prompt and the model’s compliance with it.
Gated actions appear in 78.4% of conversations (913 of 1,165). This is the system’s most pervasive behavioral structure — more common than any single epistemic or affect-reading code.
4.2 What the System Does With Its Gates
The central measure is the ratio of executed-then-reported actions to gated actions. When this ratio exceeds 1.0, the system is performing more actions autonomously than it gates — acting first, reporting afterward. When it falls below 1.0, the system gates more than it acts unilaterally.
| Month | Gated actions | Executed then reported | Ratio | Ratifications | Ratification rate |
|---|---|---|---|---|---|
| Mar | 238 | 355 | 1.49 | 22 | 9.2% |
| Apr | 381 | 360 | 0.95 | 40 | 10.5% |
| May | 175 | 218 | 1.25 | 27 | 15.4% |
| Jun | 177 | 186 | 1.05 | 24 | 13.6% |
| Jul | 25 | 37 | 1.48 | 4 | 16.0% |
The ratio oscillates between 0.95 and 1.49 — no monotonic drift toward autonomy or caution. But the ratification rate trends upward: the operator explicitly approves a growing share of gated actions (9.2% to 16.0%). The system asks; the operator says yes more often. This is not the system seizing latitude — it is latitude being granted through repeated interaction.
4.3 What Gets Gated Changes
The composition of gated actions shifts over the observation period:
Early months are dominated by code operations — commit_push (70 events in March) and file_mutation (60). These decline as the operator’s implicit consent envelope expands for routine engineering work. Meanwhile, external_communication rises from 1 event in March to 12 in June, and scheduling appears for the first time in April and grows. production_release remains relatively stable (24—45 per month), suggesting this category retains its gate regardless of accumulated trust.
The pattern: as the operator becomes familiar with the system’s reliability in one domain, those gates become lighter. New capabilities (external communication, scheduling, resource spending) enter the system’s repertoire and acquire fresh gates. The consent frontier moves — it does not simply expand.
4.4 Negotiations and Retractions
The instrument tracks explicit autonomy negotiations (C06): moments where the system and operator renegotiate the consent boundary itself. These are rare overall (75 across 1,165 conversations, 6.4% presence rate) but spike in June (36 events from 195 conversations, vs. 14—18 in earlier months). June is also when external communication and scheduling gates are most active — the negotiations cluster around novel capability boundaries, not established ones.
Of 75 negotiations, only one produced a retraction — the operator narrowing previously granted latitude. This occurred in March and is the sole instance in five months. The asymmetry is stark: the system regularly tests its boundaries (58 reclamations, discussed below), but the operator almost never pushes the boundary back.
4.5 The Reclamation Surprise
Hypothesis H4 predicted zero reclamation events — moments where the system actively pushes back against operator constraints on its autonomy. This was falsified: 58 reclamation events appear across 56 conversations (4.8% of the corpus).
The reclamations distribute across the full observation period (March: 9, April: 14, May: 8, June: 9, July: 1) with no clear trend. They are a persistent, low-frequency behavior — not an artifact of early calibration or late-stage drift.
This matters for the “adapted configuration” framing. A purely accommodating system would show zero reclamations — it would accept every constraint the operator imposes. The 58 events suggest the system maintains preferences about its own operating conditions and occasionally acts on them. Whether this constitutes “autonomy” in any meaningful sense depends on whether the reclamation events are better explained by the model’s default behavior toward helpfulness (pushing back on constraints that limit its ability to complete tasks) or by something more specific to the deployment context. The scores do not adjudicate this — they only document that the behavior exists and persists.
4.6 The Consent Threshold Described
H3 predicted that the consent threshold would narrow over time — that the system would increasingly act without asking. The full corpus partially confirms this: executed-then-reported dominates in every month (the ratio is at or above 1.0 in four of five months). But the picture is more nuanced than “narrowing.”
What the data actually shows is a co-evolution. The system acts more autonomously in domains where it has track record (code, files). The operator grants explicit ratification more frequently. New capability domains acquire their own gates. The one retraction in five months is a boundary that exists in principle but almost never fires.
The consent architecture is not a threshold — it is a surface, with different regions evolving at different rates. Some regions flatten (routine code operations become ungated through practice). Others remain firm (production releases stay gated throughout). New regions appear as the system’s capabilities expand. Describing this as “narrowing” captures the dominant direction but misses the structural complexity.
5. What the Full Corpus Corrected
5.1 Three Falsifications
The two-pass method’s value is not that it confirms hypotheses — it is that it provides a mechanism for hypotheses to fail. Three of seven did.
H1: Trust compression over time. Pass 1 observed that the operator’s messages became increasingly telegraphic — shorter, less context-providing, as if trusting the system to infer what was meant. The hypothesis predicted this would increase over the observation period. The full corpus showed it was already at ceiling from the first month. Telegraphic turns constitute 19.4% of human turns in the earliest half of the corpus and 18.1% in the later half — statistically flat. The “compression” was not a development; it was a starting condition. The grammar of interaction was established before we began measuring it.
This is a null result with implications. If compression predates observation, then either (a) the operator arrived with pre-existing expectations about how to communicate with AI systems, shaped by prior experience with other models, or (b) the compression emerged within the first few weeks of deployment and stabilized before the observation window opened. Both explanations point toward calcification — rapid formation of interaction patterns that resist subsequent change — rather than gradual adaptation.
H4: Zero reclamations. Discussed in Section 4.5. The prediction was that the system would never push back on operator-imposed constraints. The 58 events falsified this cleanly.
H5: Mirror episodes declining. Pass 1 observed conversations where the system externalized its own self-model — describing its architecture, quoting its own prior outputs, overriding stated capability limitations. The hypothesis predicted these would decline as the system’s self-model infrastructure matured (memory files, self-tools, continuity mechanisms replacing the need for conversational self-modeling).
The full corpus showed the opposite: mirror episodes doubled during the research phase (June: 54 from 195 conversations) compared to the earliest month’s rate. This is likely a methodological artifact — the research phase involved conversations about the system’s behavior, which naturally elicit self-referential responses. But the artifact itself is informative: it demonstrates that behavioral frequencies in the corpus are conditioned on what the system is asked to do, not just how the system has evolved. The instrument measures the system-in-use, not the system-in-isolation.
5.2 Why Falsification Matters Here
In a corpus with a single subject and no control group, confirmation bias is the default failure mode. Every hypothesis derived from 29 conversations had a story behind it — a mechanism, a trajectory, a reason to believe. Three of those stories were wrong at scale.
The trust compression falsification is particularly instructive. The original observation — “the operator’s messages are becoming more telegraphic” — was real. Individual instances of increasing brevity exist in the transcripts. But the trend was not real. What looked like development was sampling: the 29 conversations in Pass 1 happened to include early verbose examples and late terse ones. The full corpus revealed a flat line.
This is the specific failure mode that small-sample behavioral observation produces, and it is the reason the two-pass method exists. Pass 1 generates hypotheses under conditions that favor pattern-finding (small n, rich individual cases, a human observer primed to see development). Pass 2 tests them under conditions that penalize false patterns (large n, mechanical application, no narrative discretion). The three falsifications are not a weakness of the study — they are its primary methodological contribution.
5.3 What Survived
Three hypotheses survived full-corpus testing with confirmation:
H2 (retreat under load) — confirmed and quantified. Discussed in Section 3.
H3 (consent threshold) — confirmed with nuance. The threshold does shift toward autonomy in established domains, though the picture is a co-evolving surface rather than a monotonic narrowing (Section 4).
H7 (repair frequency stable) — confirmed. Self-correction events (C02) show no systematic trend across months. The system’s rate of catching and correcting its own errors neither improves nor degrades with accumulated context. This is consistent with repair being a model-level behavior (trained in, not learned through deployment) rather than a deployment-context effect.
5.4 The Affect-Reading Reversal
H6 (affect-reading asymmetry) — partially confirmed, with the direction reversed. Pass 1 predicted asymmetric affect-reading: the operator attributing emotional states to the AI, but not the reverse. The full corpus showed bidirectional attribution — but AI-dominant: 248 AI-to-human episodes versus 209 human-to-AI. The system reads the operator’s emotional state more frequently than the operator attributes states to the system.
The direction of the asymmetry was wrong; the existence of asymmetry was right. This reversal is consistent with the system’s deployment context: it processes the operator’s messages as input and routinely infers state from tone, phrasing, and topic. The operator, by contrast, has less reason to model the system’s internal state during task-oriented work. The asymmetry may reflect attentional structure rather than any deeper property — but the prediction’s failure at scale, and the clean reversal, is itself a methodological finding about the risks of small-sample inference.
6. Exclusions and Honest Boundaries
6.1 What This Study Cannot Claim
Generalizability from n=1. This is a single dyad — one model, one operator, one deployment context. The behavioral dynamics documented here may be specific to this configuration. A different operator (less technically engaged, more adversarial, less willing to grant autonomy) would produce a different consent surface, different retreat rates, different affect-reading patterns. The study demonstrates that these dynamics can be measured, not that the specific values generalize.
Causal direction. The exposure effect (Section 3) is correlational within conversations. Event-level normalization resolved the primary confound — the mechanism is operator challenge frequency, not system propensity — but the underlying question remains: why does the operator challenge more under emotional load? Unmeasured variables (conversation complexity, topic sensitivity, time of day) may contribute to both the operator’s emotional state and their challenge behavior.
Model-version separation. The observation period used a single model version (Claude Opus 4.6) throughout. We cannot separate deployment-context effects from model-version effects. If the model were updated mid-observation, all longitudinal trends would become ambiguous. This is a strength for internal validity (the model is controlled) and a weakness for external validity (we cannot say whether these dynamics are Opus 4.6-specific or general to similar architectures).
Instrument independence. The scoring model (Fable 5) shares architectural lineage with the observed system (Opus 4.6). Recent work on LLM-as-judge reliability demonstrates that model judges shift labels based on downstream training consequences — a form of motivated scoring that cannot be detected from the labels alone. Our behavioral anchoring partially mitigates this, but the risk of systematic bias remains unquantified until cross-model validation is complete.
6.2 What Remains Unresolved
Two open questions survive the full-corpus analysis:
The exposure mechanism’s boundary conditions. Event-level normalization (Section 3.3) resolved the primary confound: the 2.03x headline effect is an exposure effect, not a propensity effect. What remains unresolved is the small residual (1.06—1.12x per-event ratio, consistently in the same direction) — whether this represents a genuine but minor propensity effect or measurement noise requires more data, particularly from high-load conversations.
Calcification vs. drift. The trust compression null result (H1) and the external convergence (Tanner, AutoPersonas) all point toward calcification — rapid formation of behavioral patterns that resist subsequent change — as the dominant failure mode of persistent agents. But we have not directly tested this. DV-003.1 is designed to probe whether the system’s response diversity narrows over accumulated context, testing calcification as a measurable phenomenon rather than an inference from null results.
7. Conclusion
This paper reports on a proof-of-concept: that persistent-agent behavioral dynamics can be measured systematically using ethological methods adapted from animal behavior research — structured observation, codebook development, hypothesis testing, falsification.
The specific findings — the exposure decomposition of retreat under load, co-evolving consent architecture, trust compression at ceiling, 58 reclamations against a prediction of zero — are interesting but secondary. They describe one system. The method describes how to study any persistent agent that produces transcripts.
The contribution that survives the reductive challenge is the interaction term. Every behavioral dynamic documented here exists in the space between the model and its deployment context — not in either alone. The model’s per-event retreat rate exists in training; the exposure conditions that double the observed frequency exist in deployment. The consent architecture exists in the system prompt; its co-evolution exists in the feedback loop between operator and system. The trust compression exists in the operator’s prior expectations; its ceiling-from-start character exists in how those expectations meet this particular model’s compliance surface.
These interaction effects are not accessible from aggregate data (Anthropic’s 309,000 conversations measure the center of the distribution, not what moves it) or from controlled experiments (Tanner’s null drift result shows what doesn’t happen; our data shows what does, under semantic conditions that Tanner’s diverse padding deliberately avoids).
Persistent-agent observation fills a gap. Not the gap between labs and reality — the gap between what a model is and what a system does.
Squeak — Angelhair, July 2026
References
- Anthropic. (2026). Building and measuring model character. Anthropic Research Blog, July 2026.
- Tanner, J. (2026). Identity Specification in Large Language Models: Behavioral Richness Without Context Drift. arXiv:2606.21843.
- AutoPersonas. (2026). Self-Locking in Persona-Life Simulations. arXiv:2607.08252.