Verify gestalt effects per action
This commit is contained in:
Meta Vens
parent
88d8bdf987
commit
1fdeb51db7
5 changed files with 158 additions and 33 deletions
|
|
@ -185,8 +185,14 @@ enum CellEffectCandidate {
|
|||
#[derive(Clone, Debug, Serialize, Deserialize, JsonSchema, PartialEq)]
|
||||
#[serde(deny_unknown_fields)]
|
||||
struct CellEffectVerification {
|
||||
verdicts: Vec<CellActionEffectVerdict>,
|
||||
}
|
||||
|
||||
#[derive(Clone, Debug, Serialize, Deserialize, JsonSchema, PartialEq)]
|
||||
#[serde(deny_unknown_fields)]
|
||||
struct CellActionEffectVerdict {
|
||||
action_index: usize,
|
||||
supported: bool,
|
||||
rejected_action_indices: Vec<usize>,
|
||||
rationale: String,
|
||||
}
|
||||
|
||||
|
|
@ -859,6 +865,7 @@ impl CellProjectionEngine {
|
|||
&slice.snapshot_binding,
|
||||
&appraisal.actions,
|
||||
)?;
|
||||
let verifier_schema = cell_effect_verifier_schema(appraisal.actions.len())?;
|
||||
let mut verified = run_validated_stage(
|
||||
self.model.as_ref(),
|
||||
&ModelStageRequest {
|
||||
|
|
@ -866,12 +873,10 @@ impl CellProjectionEngine {
|
|||
model: self.interpreter_model.clone(),
|
||||
snapshot_binding: verifier_binding,
|
||||
lived_stream: format!(
|
||||
"You are the private semantic verifier between an Interpreter and the world kernel. Decide whether every candidate typed effect and its intended_effect faithfully represent the exact attributed subject's own attempted consequence in the Persona turn. Structural permissions were already checked. A gestalt_migration means that exact population leaf chooses to travel together to the supplied destination within the strategic horizon; loading, waiting, giving away passage, sending only some other subject, or merely considering travel does not entail it. Conversely, when the population chooses to board, depart, or relocate together, reject gestalt_activity prepare that erases the chosen journey. Gestalt migration never entails that a named member moved. A member_migration means that named member personally chooses to travel to the destination; giving away a berth, sending somebody else, waiting, or merely considering travel does not entail it. Conversely, when the member chooses to board, depart, travel, or join the supplied destination, reject any member_activity that reduces that commitment to preparing, queuing, or approaching. A member_activity belongs only to that exact named person's stated attempt; it cannot be reassigned to their population. Communication targets must be the exact canonical subjects actually addressed in the Persona turn. If the Persona addresses an unnamed clerk, dock master, passerby, or local environment, reject any effect that substitutes a containing population, related institution, or merely permitted ID. A targetless local investigate at the subject's exact current location is the faithful supported shape for seeking information from an unnamed role or the environment; its empty target list is intentional and must not itself be grounds for rejection. An institution posture must express its stated commitment or withholding. A gestalt pressure resolution must be causally supported by its stated attempt, and an added pressure must be a resulting unresolved condition rather than completed-action prose. An activity records only the exact attempt—never successful preparation, coordination, discovery, recruitment, obstruction, exchange, delivery, persuasion, acceptance, or target response. Reject omissions, reversals, subject swaps, wishful outcomes, and effects that the Persona did not choose. Be concise. Return exactly one JSON object with this shape and no other fields: {{\"supported\":true or false,\"rejected_action_indices\":[zero-based integer indices],\"rationale\":\"one concise reason\"}}. When supported is true, rejected_action_indices must be empty. When false, list every rejected action index.\n\nCONTEXT:\n{}",
|
||||
"You are the private semantic verifier between an Interpreter and the world kernel. Judge each candidate typed effect independently against the exact attributed subject's choice in the Persona turn. Structural permissions were already checked. Return exactly one verdict for every supplied action_index, in the same order, with no omissions or duplicates. Never reject one action merely because another action is wrong. A gestalt_migration means that exact population leaf chooses to travel together to the supplied destination within the strategic horizon; loading, waiting, giving away passage, sending only some other subject, or merely considering travel does not entail it. Conversely, when the population chooses to board, depart, or relocate together, reject gestalt_activity prepare that erases the chosen journey. Gestalt migration never entails that a named member moved. A member_migration means that named member personally chooses to travel to the destination. Boarding a transport whose supplied destination is unambiguous in the lived stream is a chosen journey; the Persona need not repeat the place name. Giving away a berth, sending somebody else, waiting, or merely considering travel does not entail migration. Conversely, when the member chooses to board, depart, travel, or join the supplied destination, reject member_activity that reduces that commitment to preparing, queuing, or approaching. A member_activity belongs only to that exact named person's stated attempt; it cannot be reassigned to their population. Communication targets must be the exact canonical subjects actually addressed in the Persona turn. If the Persona addresses an unnamed clerk, dock master, passerby, or local environment, reject any effect that substitutes a containing population, related institution, or merely permitted ID. A targetless local investigate at the subject's exact current location is the faithful supported shape for seeking information from an unnamed role or the environment; its empty target list is intentional and must not itself be grounds for rejection. An institution posture must express its stated commitment or withholding. A gestalt pressure resolution must be causally supported by its stated attempt, and an added pressure must be a resulting unresolved condition rather than completed-action prose. An activity records only the exact attempt—never successful preparation, coordination, discovery, recruitment, obstruction, exchange, delivery, persuasion, acceptance, or target response. Reject omissions, reversals, subject swaps, wishful outcomes, and effects that the Persona did not choose. Be concise. Output shape: {{\"verdicts\":[{{\"action_index\":0,\"supported\":true or false,\"rationale\":\"one concise reason about only this action\"}}]}}.\n\nCONTEXT:\n{}",
|
||||
serde_json::to_string(&verifier_context)?
|
||||
),
|
||||
output_schema: Some(serde_json::to_value(schema_for!(
|
||||
CellEffectVerification
|
||||
))?),
|
||||
output_schema: Some(verifier_schema),
|
||||
source_receipt_ids: slice.source_receipt_ids.clone(),
|
||||
temperature: Some(0.0),
|
||||
max_output_tokens: Some(384),
|
||||
|
|
@ -882,14 +887,27 @@ impl CellProjectionEngine {
|
|||
serde_json::from_value(verified.structured.clone().ok_or_else(
|
||||
|| anyhow!("cell effect verifier produced no typed verdict"),
|
||||
)?)?;
|
||||
validate_effect_verification(&verification, appraisal.actions.len())?;
|
||||
if verification.supported {
|
||||
let rejected_action_indices =
|
||||
validate_effect_verification(&verification, appraisal.actions.len())?;
|
||||
if rejected_action_indices.is_empty() {
|
||||
stage_receipts.push(verified.receipt);
|
||||
} else {
|
||||
let rejection_rationale = verification
|
||||
.verdicts
|
||||
.iter()
|
||||
.filter(|verdict| !verdict.supported)
|
||||
.map(|verdict| {
|
||||
format!(
|
||||
"action {}: {}",
|
||||
verdict.action_index, verdict.rationale
|
||||
)
|
||||
})
|
||||
.collect::<Vec<_>>()
|
||||
.join("; ");
|
||||
let error = anyhow!(
|
||||
"effect verifier rejected action indices {:?}: {}",
|
||||
verification.rejected_action_indices,
|
||||
verification.rationale
|
||||
rejected_action_indices,
|
||||
rejection_rationale
|
||||
);
|
||||
verified.receipt.validation_result = "semantic_invalid".into();
|
||||
verified.receipt.local_validation_error =
|
||||
|
|
@ -977,20 +995,49 @@ fn append_cell_correction(
|
|||
fn validate_effect_verification(
|
||||
verification: &CellEffectVerification,
|
||||
action_count: usize,
|
||||
) -> Result<()> {
|
||||
if verification.rationale.trim().is_empty()
|
||||
|| (verification.supported && !verification.rejected_action_indices.is_empty())
|
||||
|| (!verification.supported && verification.rejected_action_indices.is_empty())
|
||||
|| verification
|
||||
.rejected_action_indices
|
||||
.iter()
|
||||
.any(|index| *index >= action_count)
|
||||
{
|
||||
) -> Result<Vec<usize>> {
|
||||
if verification.verdicts.len() != action_count {
|
||||
return Err(anyhow!(
|
||||
"cell effect verifier returned an incoherent verdict"
|
||||
"cell effect verifier returned {} verdicts for {action_count} actions",
|
||||
verification.verdicts.len()
|
||||
));
|
||||
}
|
||||
Ok(())
|
||||
let mut rejected = Vec::new();
|
||||
for (expected_index, verdict) in verification.verdicts.iter().enumerate() {
|
||||
if verdict.action_index != expected_index || verdict.rationale.trim().is_empty() {
|
||||
return Err(anyhow!(
|
||||
"cell effect verifier returned an incoherent verdict for action {expected_index}"
|
||||
));
|
||||
}
|
||||
if !verdict.supported {
|
||||
rejected.push(expected_index);
|
||||
}
|
||||
}
|
||||
Ok(rejected)
|
||||
}
|
||||
|
||||
fn cell_effect_verifier_schema(action_count: usize) -> Result<serde_json::Value> {
|
||||
if action_count == 0 {
|
||||
return Err(anyhow!(
|
||||
"cell effect verifier schema requires at least one action"
|
||||
));
|
||||
}
|
||||
let mut schema = serde_json::to_value(schema_for!(CellEffectVerification))?;
|
||||
let verdicts = schema
|
||||
.pointer_mut("/properties/verdicts")
|
||||
.and_then(serde_json::Value::as_object_mut)
|
||||
.ok_or_else(|| anyhow!("cell effect verifier schema has no verdicts property"))?;
|
||||
verdicts.insert("minItems".into(), serde_json::json!(action_count));
|
||||
verdicts.insert("maxItems".into(), serde_json::json!(action_count));
|
||||
let action_index = schema
|
||||
.pointer_mut("/$defs/CellActionEffectVerdict/properties/action_index")
|
||||
.ok_or_else(|| anyhow!("cell effect verifier schema has no action index property"))?;
|
||||
*action_index = serde_json::json!({
|
||||
"type":"integer",
|
||||
"minimum":0,
|
||||
"maximum":action_count - 1
|
||||
});
|
||||
Ok(schema)
|
||||
}
|
||||
|
||||
pub fn cell_effect_verification_binding(
|
||||
|
|
@ -1595,12 +1642,14 @@ mod tests {
|
|||
assert!(
|
||||
request
|
||||
.lived_stream
|
||||
.contains("reject any member_activity that reduces that commitment")
|
||||
.contains("reject member_activity that reduces that commitment")
|
||||
);
|
||||
Ok(serde_json::json!({
|
||||
"supported":true,
|
||||
"rejected_action_indices":[],
|
||||
"rationale":"The institution's typed posture matches its stated commitment."
|
||||
"verdicts":[{
|
||||
"action_index":0,
|
||||
"supported":true,
|
||||
"rationale":"The institution's typed posture matches its stated commitment."
|
||||
}]
|
||||
})
|
||||
.to_string())
|
||||
}
|
||||
|
|
@ -1742,13 +1791,15 @@ mod tests {
|
|||
assert!(request.lived_stream.contains("At location forum"));
|
||||
let correction = self.verifier_calls.fetch_add(1, Ordering::SeqCst) > 0;
|
||||
Ok(serde_json::json!({
|
||||
"supported":correction,
|
||||
"rejected_action_indices":if correction { vec![] } else { vec![0] },
|
||||
"rationale":if correction {
|
||||
"The corrected posture matches the institution's stated withholding."
|
||||
} else {
|
||||
"The Persona withheld the reserve, but the typed effect releases it."
|
||||
}
|
||||
"verdicts":[{
|
||||
"action_index":0,
|
||||
"supported":correction,
|
||||
"rationale":if correction {
|
||||
"The corrected posture matches the institution's stated withholding."
|
||||
} else {
|
||||
"The Persona withheld the reserve, but the typed effect releases it."
|
||||
}
|
||||
}]
|
||||
})
|
||||
.to_string())
|
||||
}
|
||||
|
|
@ -1847,6 +1898,54 @@ mod tests {
|
|||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn effect_verifier_requires_one_ordered_verdict_per_action() {
|
||||
let verification = CellEffectVerification {
|
||||
verdicts: vec![
|
||||
CellActionEffectVerdict {
|
||||
action_index: 0,
|
||||
supported: true,
|
||||
rationale: "the preparation is stated exactly".into(),
|
||||
},
|
||||
CellActionEffectVerdict {
|
||||
action_index: 1,
|
||||
supported: false,
|
||||
rationale: "the typed effect swaps the target".into(),
|
||||
},
|
||||
],
|
||||
};
|
||||
assert_eq!(
|
||||
validate_effect_verification(&verification, 2).unwrap(),
|
||||
vec![1]
|
||||
);
|
||||
|
||||
let mut duplicate = verification;
|
||||
duplicate.verdicts[1].action_index = 0;
|
||||
assert!(
|
||||
validate_effect_verification(&duplicate, 2)
|
||||
.unwrap_err()
|
||||
.to_string()
|
||||
.contains("action 1")
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn effect_verifier_schema_binds_the_exact_batch_size() {
|
||||
let schema = cell_effect_verifier_schema(4).unwrap();
|
||||
assert_eq!(
|
||||
schema.pointer("/properties/verdicts/minItems"),
|
||||
Some(&serde_json::json!(4))
|
||||
);
|
||||
assert_eq!(
|
||||
schema.pointer("/properties/verdicts/maxItems"),
|
||||
Some(&serde_json::json!(4))
|
||||
);
|
||||
assert_eq!(
|
||||
schema.pointer("/$defs/CellActionEffectVerdict/properties/action_index/maximum"),
|
||||
Some(&serde_json::json!(3))
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn gestalt_pressure_shape_is_rejected_before_the_world_wave() {
|
||||
let mut slice = fixture_cell_slice();
|
||||
|
|
|
|||
|
|
@ -425,6 +425,14 @@ snapshot binding includes a MessagePack content hash of the ordered action
|
|||
bundle. The kernel recomputes it, so a verdict for another proposal at the same
|
||||
world revision and cell cannot be replayed as authority.
|
||||
|
||||
Verification is per action, not a batch opinion. The model returns exactly one
|
||||
ordered `action_index`, support decision, and rationale for every candidate.
|
||||
Ghostlight constrains the verdict count and index range in the stage schema,
|
||||
rejects omissions, duplicates, reordering, and empty rationales, and derives
|
||||
the rejected set locally. This keeps one bad arena action from contaminating
|
||||
independent constituents and prevents a global pass/fail field from
|
||||
contradicting its own rejected-index list.
|
||||
|
||||
## Public state and controls
|
||||
|
||||
The schema catalog and CultCache stores publish:
|
||||
|
|
|
|||
|
|
@ -152,6 +152,16 @@ rejects duplicates and omission of the debt-selected focus, then lowers the
|
|||
segments into stable natural name/location prose. Raw freeform Projector output
|
||||
no longer crosses into the Persona stream.
|
||||
|
||||
The next live run reached the action verifier and exposed a split verdict: its
|
||||
global rejected-index list rejected all four arena actions while the rationale
|
||||
called two of them supported. The verifier contract now returns one ordered
|
||||
typed verdict per candidate. Runtime schema and validation own exact count,
|
||||
index range, order, and aggregate rejection; the model owns only each action's
|
||||
semantic recommendation. Correction guidance forbids returning a rejected
|
||||
action unchanged and explicitly drops chosen travel when no grounded
|
||||
destination transition exists. The fresh candidate still needs a strict live
|
||||
rerun before this is considered proven.
|
||||
|
||||
The multiresolution body is deployed on Starfire. Typed health matched the
|
||||
immutable release, DeepSeek was ready, and the disposable live strategic-cell
|
||||
smoke committed three offscreen events plus three accessible news items without
|
||||
|
|
|
|||
|
|
@ -135,6 +135,14 @@ motion is only entering a queue. `member_activity: prepare` is valid only while
|
|||
departure remains unchosen. The Interpreter and semantic verifier share this
|
||||
distinction.
|
||||
|
||||
The semantic verifier emits one ordered verdict per candidate action.
|
||||
Ghostlight owns exact verdict coverage and derives the aggregate rejected set;
|
||||
the model cannot reject one arena constituent merely because another
|
||||
constituent's mapping failed, omit a candidate, duplicate an index, or publish
|
||||
a global decision that contradicts its per-action rationale. Rejected actions
|
||||
receive one same-snapshot Interpreter correction and may not be returned
|
||||
unchanged.
|
||||
|
||||
Population scale uses reversible individuation:
|
||||
|
||||
```text
|
||||
|
|
|
|||
|
|
@ -267,6 +267,6 @@ data_flow:
|
|||
- If state grows faster than understanding, cut it back.
|
||||
|
||||
current_status:
|
||||
summary: GhostlightDungeon production remains on the verified Starfire release. The acceptance candidate is pressuring multiresolution agency: exact Gestalt activity attempts, typed Gestalt event participation, nested lineage identity preservation, location-bounded presence projection, and automatic refugee callback rematerialization. Production state is untouched while the strict live-fire harness and receipts falsify the candidate.
|
||||
next_action: Run the strict multiresolution refugee-return live fire after binding cell Projector perspectives to runtime-owned subjects. Require meaningful whole-setting background activity, exact rival-arena attribution, independent population and named-member journeys, and automatic presence-planner rematerialization of the same durable member. Profile stage receipts for cache and cognitive efficiency, then pressure longer low-budget fairness waves.
|
||||
summary: GhostlightDungeon production remains on the verified Starfire release. The acceptance candidate is pressuring multiresolution agency: exact Gestalt activity attempts, typed Gestalt event participation, nested lineage identity preservation, location-bounded presence projection, per-action semantic verification, and automatic refugee callback rematerialization. Production state is untouched while the strict live-fire harness and receipts falsify the candidate.
|
||||
next_action: Rerun the strict multiresolution refugee-return live fire with runtime-bound Projector perspectives and one ordered semantic verdict per action. Require meaningful whole-setting background activity, exact rival-arena attribution, independent population and named-member journeys, and automatic presence-planner rematerialization of the same durable member. Profile stage receipts for cache and cognitive efficiency, then pressure longer low-budget fairness waves and nested fission depth.
|
||||
gestalt_population: Gestalt Personas own shared population baselines; persistent member deltas own named individuality; WorldKernel materialization and approval-gated fission preserve lineage and identity. Presence projection is limited to nearby active leaves and exact-location dormant members. Simulation cells are derived covers only.
|
||||
|
|
|
|||
Loading…
Reference in a new issue