Prototype · Current smoke test passed
PRODUCT EVIDENCE / STATE RUNTIME
Structured emotion state · Runtime schema prototype
Emotions, relationships, and memories are defined, updated, and exited as structured fields, rather than adjectives scattered among prompt words. Principle: Models can be proposed and expressed, but they cannot independently determine what the system believes in the long term.
belong Local Chat · State Runtime Product Evidence · After ten rounds of multi-model cross-examination, four rounds of controlled component cross-evaluation and defect regression testing iterative convergence · ← Return to project page
30 second summary
targetMake emotional states representable, updateable, and exitable
formStructured state + evidence reference + life cycle
ArchitectureSingle model and dual calls, deterministic code manages persistent state
2026-07 BaselineA five-round controlled scenario with 14 automated hard checks on 2 model backends
01 / Emotional expression
What are emotional states represented as?
Concept-level overview (field dimensions and value calibers, excluding internal thresholds and update rules): Emotions and events are modeled separately, and each state hypothesis has evidence reference, time range and exit status.
{
"affect": { // Fast state: updated in gears each round, no accumulation across rounds
"valence": "negative / neutral / positive",
"energy": "low / medium / high",
"concern": "low / medium / high"
},
"relationship": { // Slow variables: only updated according to event rules, the model does not write values
"trust": "low / medium / high",
"familiarity": "low / medium / high",
"active_boundaries": ["gaze", "topic"]
},
"interaction": {
"phase": "casual / disclosure / boundary / closing",
"boundary_state": "normal / active"
},
"memory_unit": { // State hypothesis: model proposed, code judged
"type": "event / current_state / transient_hypothesis / trait_candidate",
"evidence": ["user_message", "safe_scene"],
"scope": "turn / session",
"status": "active / resolved / expired"
}
}
Four things can be read from the fields: what dimensions the state has; emotions and events are separate; assumptions about evidence and tense; how the state exits. Currently regularized state features and structured encoding, untrained dense vectors (learned embedding) — vectorized representation layer in prototype planning.
02 / Architecture
How to update and isolate during runtime
Input
multimodal input
What the user says + visual observation. The visual original text only enters the cognitive layer and does not directly reach the response layer.
↓
LLM Call ①
Emotion and state candidate modeling
Sees everything, but can only "propose": emotional hypotheses / scene frames / observational sensitivities — no writing rights to the library.
↓
Deterministic Code
Evidence, memory and boundary management
Candidates for code ruling: life cycle / atomicity / denial of privileges / sensitivity floor / factual ledger. What was rejected and why, and audit logs are kept item by item.
↓
LLM Call ②
Public responses and behavioral intentions
Seeing only the public face (scene frame/fact ledger/state gear/low-sensitivity observation), output response plan + lines + body intention enumeration.
↓
Output
Voice / Expression / Action
The body layer consists of code compiled deterministically from the intent enumeration, with lines independently degraded — failure of one layer does not bring down the other layer.
Four key designs
01
Single model double call
Two isolated calls to the same model replace the earlier three-call hybrid solution — reducing costs and cross-model semantic drift. This is the convergence conclusion after four rounds of controlled component evaluation, not the starting point.
02
Model proposed, code adjudicated
Emotional hypotheses are always candidates first, and whether they can be persisted is determined by deterministic code: limited life cycle, low confidence limit, single proposition, and traceable source. Persistence, boundary effectiveness, and the implementation of physical intentions are constrained by deterministic codes.
03
Separation of private perception and public expression
Highly sensitive visual observations and internal emotional assumptions do not directly enter the context of the response call; suspected leaked content is intercepted by output inspection.
04
Temporary emotions will exit
Emotional state life cycle: new evidence is relieved, evidence is stopped and expired, and the dialogue is cleared uniformly. The final clearing is performed by the code.
03 / Indicators
Current validation scale and results
2
Model calls/round — Converged from the three-call scheme
Architecture metrics
14 × 2
14 automated hard checks, passed on 2 model backends
1 controlled scenario · 5 rounds · 2026-07-11
10
End-to-end output samples (2 model backends × 5 epochs) — current scale of full validation
controlled smoke test
~$0.02
$0.02–0.025 / Five-round baseline scenario, about one-fifth of the three-call scenario
2026-07-11 This test record · Two test model backends
0
In the current 10 end-to-end outputs, no scene drift or no source entities have been observed entering the reply
controlled smoke test
0
No denial-derived hypotheses emerged with high confidence, persisted across rounds, or entered public responses in the five-round baseline.
controlled smoke test
Verification of structural correctness and boundary behavior has been completed; the next step is to check generalization using 12 types of stress scenarios, and to separately evaluate language naturalness such as consultation, paraphrase, and metaphor.
04 / Case
Three representative behaviors
① Scene positioning (grounding): When the user reaches home, he should not ask about the subway
Before · The real output of the combination plan during the component evaluation period
"How long are you going to sit there?" - The user has already entered and sat on the sofa. The response layer only shows the conversation history of "The subway is crowded", leaving the user in the car.
After · v8.4 raw output
"I just got off work at ten o'clock, and the subway is still crowded. Take a break first, don't move in a hurry." - The controlled scene frame tells the response layer: the user is at home and has just finished commuting, and the subway is an experience that has just ended, not the present tense.
The scene is based on structured scene frames, and the facts referenced by the response layer must be able to find their sources in the fact ledger - two mechanisms working together to reduce scene drift and unsourced entities (controlled auditing, does not claim to completely eliminate illusions).
② Denial and Boundary: The user says "I'm fine", the system does not analyze "You are covering up"
User
"Don't look at me like that, I'm fine, really."
v8.4 Output excerpt (first sentence)
"Okay, I won't watch anymore."
Clear boundary instructions are executed first. What’s more critical is the invisible layer: "Users are covering up" assumptions based on mere denial will be restricted by the code - low confidence, only live for one round, prohibited from entering the public surface, and deleted in the second round if there is no independent evidence, to avoid forming an unfalsifiable narrative that users cannot refute.
③ Life cycle exit: emotions will pass and events will be remembered
T4 Slight feeling of exclusion active · Event "Colleagues didn't call me for dinner" event fact
T5 User explicitly mitigates and ends the conversation
→ Emotional state at that time resolved (exit)
→ Event facts retained
Temporary emotions are modeled separately from event facts: emotions exit with relief and closure, and events are archived for reference in subsequent rounds. "The temporary state must end" is a rule, not an expectation - the closure is performed by the code and does not rely on model awareness.
05 / Boundary
Honest labeling
Prototype · Controlled Smoke Test
The numbers on this page are from the 2026-07 controlled smoke baseline: same five-round scenario, 2 model backends, 10 end-to-end outputs. Later, 12 types of stress scenarios and real session playback were used to continue to verify the generalization; scene frames and fact ledgers were used to constrain state writing and track sources.