Model perceived state as attribution

This commit is contained in:
Meta Vens 2026-05-01 19:26:57 +01:00
commit ddf7898ea2
8 changed files with 59 additions and 35 deletions

View file

@ -192,10 +192,12 @@ This gives us:
- scene-sensitive activation without rewriting the whole person every turn
Canonical authored state should not include epistemic uncertainty about itself.
If a value is only an observer's read, a classifier output, or a draft import,
store that uncertainty as `confidence` on the perceived, inferred, or workflow
record instead. The character is not uncertain about our spreadsheet. The
spreadsheet may be uncertain about the character.
If a value is only an observer's read, store it as perceived state rather than
canonical state. Perceived dimensions should not expose canonical `mean` or
`plasticity`; they should track observed activation, attributed disposition,
confidence, and correction resistance. This lets the model represent
fundamental attribution bias: someone can behave a certain way under pressure
and be perceived as simply being that way.
## Culture As A Prior Field

View file

@ -39,20 +39,30 @@ Canonical character state does not include `certainty`. If a generated or
imported profile is provisional, track that in workflow provenance, draft
status, imported-source notes, or review status.
Perceived or inferred state can use `confidence`, because that is explicitly
about an observer's or classifier's read:
Perceived state uses a different shape because observers do not know canonical
baseline or plasticity. They see behavior, then often treat that behavior as
evidence of what the target is like.
```json
{
"mean": 0.7,
"observed_activation": 0.9,
"attributed_disposition": 0.7,
"confidence": 0.4,
"plasticity": 0.5,
"current_activation": 0.8
"correction_resistance": 0.6
}
```
That means the observer currently reads the target as high in this dimension,
but the read is weakly supported or uncertain.
| Field | Meaning |
| --- | --- |
| `observed_activation` | How strongly the observer thinks the target expressed the dimension in the observed event or scene. |
| `attributed_disposition` | How much the observer generalizes from the observed behavior into "this is what the target is like." |
| `confidence` | How strongly the observer trusts this read. |
| `correction_resistance` | How much counter-evidence is likely to be needed before the observer revises the attribution. |
This models fundamental attribution bias directly. A character can see someone
behave coldly during a crisis and conclude "they are cold," not "they were
activated toward distance under exceptional pressure." Correcting that
impression takes work.
## Label Set Policy

View file

@ -67,11 +67,12 @@ A person can be dispositionally suspicious without being maximally suspicious
in every scene. A person can be highly activated right now without rewriting
their whole soul because someone coughed near a status wound.
Canonical variables do not include confidence or certainty. Perceived and
inferred variables use `confidence` when a read can be wrong.
Canonical variables do not include confidence or certainty. Perceived variables
do not expose canonical `mean` or `plasticity`; they store observed activation,
attributed disposition, confidence, and correction resistance.
For full meanings of `mean`, `plasticity`, `current_activation`, perceived
`confidence`, and every required v0 label, see
For full meanings of canonical `mean`, `plasticity`, `current_activation`,
perceived attribution fields, and every required v0 label, see
`docs/architecture/agent-state-variable-glossary.md`.
## Required First-Class Dimensions
@ -143,13 +144,15 @@ An overlay records:
- observer
- target
- perceived dimensions
- perceived dimensions, including observed activation and attributed disposition
- beliefs
- distortions
This is where misunderstanding lives. Suspicious agents can overread threat.
Attachment-hungry agents can overread warmth. Avoidant agents can underread
care. None of that requires corrupting the canonical state.
care. Fundamental attribution bias also lives here: an observer can mistake
exceptional behavior for stable character. None of that requires corrupting the
canonical state.
## Dialogue Context Packs

View file

@ -744,16 +744,16 @@
"target_agent_id": "cat_marrigan",
"perceived_dimensions": {
"warmth": {
"mean": 0.44,
"plasticity": 0.55,
"current_activation": 0.4,
"confidence": 0.34
"observed_activation": 0.4,
"attributed_disposition": 0.44,
"confidence": 0.34,
"correction_resistance": 0.42
},
"suspicion": {
"mean": 0.83,
"plasticity": 0.28,
"current_activation": 0.9,
"confidence": 0.71
"observed_activation": 0.9,
"attributed_disposition": 0.83,
"confidence": 0.71,
"correction_resistance": 0.64
}
},
"beliefs": [

View file

@ -652,14 +652,19 @@
"inferred_state_variable": {
"type": "object",
"required": [
"mean",
"observed_activation",
"attributed_disposition",
"confidence",
"plasticity",
"current_activation"
"correction_resistance"
],
"additionalProperties": false,
"properties": {
"mean": {
"observed_activation": {
"type": "number",
"minimum": 0,
"maximum": 1
},
"attributed_disposition": {
"type": "number",
"minimum": 0,
"maximum": 1
@ -669,12 +674,7 @@
"minimum": 0,
"maximum": 1
},
"plasticity": {
"type": "number",
"minimum": 0,
"maximum": 1
},
"current_activation": {
"correction_resistance": {
"type": "number",
"minimum": 0,
"maximum": 1

View file

@ -9,3 +9,4 @@
{"ts": "2026-05-01T18:10:28+00:00", "type": "design", "status": "accepted", "note": "Split performed self from voice style in the v0 schema: presentation_strategy now includes abrasive_boundary and ironic_distance, while voice_style carries sarcasm, dryness, sincerity evasion, emotional evasion, and banter-as-boundary handles."}
{"ts": "2026-05-01T18:14:18+00:00", "type": "documentation", "status": "ok", "note": "Added agent-state variable glossary defining numeric components, current v0 labels, label-set policy, and why canonical fixtures include full required vectors rather than only strongly applicable labels."}
{"ts": "2026-05-01T18:21:07+00:00", "type": "design", "status": "accepted", "note": "Removed certainty from canonical authored state variables. Canonical character state now stores mean, plasticity, and current_activation; perceived/inferred dimensions use confidence when a read can be wrong."}
{"ts": "2026-05-01T18:26:40+00:00", "type": "design", "status": "accepted", "note": "Reworked perceived dimensions so observer reads do not know canonical mean or plasticity. Perceived state now tracks observed_activation, attributed_disposition, confidence, and correction_resistance to model fundamental attribution bias."}

View file

@ -74,7 +74,8 @@ accepted_design:
- Ghostlight should build drama and narrative pressure, then project character-local dialogue contexts so characters can breathe and act from their own partial state.
- Call of the Void pressure should come from Sol's whole unresolved inheritance being trapped inside Elysium, not from a lone expedition discovering there is no way home.
- The first schema contract lives in schemas/agent-state.schema.json with required first-class dimensions in schemas/agent-state.required-fields.json and a validator in tools/validate_agent_state.py.
- Canonical authored character variables do not carry certainty; confidence belongs on perceived, inferred, classifier, memory, belief, or workflow-provenance records.
- Canonical authored character variables do not carry certainty; perceived dimensions do not expose canonical mean or plasticity, and instead model observed activation, attributed disposition, confidence, and correction resistance.
- Fundamental attribution bias is first-class in perceived state: observers can mistake exceptional behavior for stable character, and correction should require counter-evidence.
current_status:
summary: Ghostlight now has the Epiphany-style persistence spine, the first architecture payload, and the initial canonical agent-state schema seam for Call of the Void-style authoring examples. The schema is still v0 and should be hardened before building prompt projection on top of it.

View file

@ -51,9 +51,16 @@ def validate_state_variable(obj: Any, path: str) -> None:
def validate_inferred_state_variable(obj: Any, path: str) -> None:
require(isinstance(obj, dict), f"{path} must be an object")
require_keys(obj, ["mean", "confidence", "plasticity", "current_activation"], path)
require_keys(
obj,
["observed_activation", "attributed_disposition", "confidence", "correction_resistance"],
path,
)
require("mean" not in obj, f"{path}.mean belongs in canonical state, not perceived state")
require("plasticity" not in obj, f"{path}.plasticity belongs in canonical state, not perceived state")
require("current_activation" not in obj, f"{path}.current_activation should be observed_activation in perceived state")
require("certainty" not in obj, f"{path}.certainty should be named confidence in perceived/inferred state")
for key in ["mean", "confidence", "plasticity", "current_activation"]:
for key in ["observed_activation", "attributed_disposition", "confidence", "correction_resistance"]:
require_0_1(obj[key], f"{path}.{key}")