eq-layer 0.2.3__tar.gz → 0.2.4__tar.gz

{eq_layer-0.2.3 → eq_layer-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: eq-layer
-Version: 0.2.3
+Version: 0.2.4
 Summary: Inspectable intent, affect, repair, uncertainty, and factored-action control for LLM responses
 Project-URL: Homepage, https://github.com/Furox-Art/eq-layer
 Project-URL: Repository, https://github.com/Furox-Art/eq-layer
@@ -51,7 +51,7 @@ So this repo does not retrain the model. It adds a layer:
 1. **Affect state** — `{valence, arousal, escalation_delta, stance, subtext}` as an explicit intermediate representation, tracked across turns. `escalation_delta` is the point: snapshot classification cannot see a trend.
 2. **Intent belief state** — the learned classifier preserves the full posterior over `action_request/status_check/explanation/question/statement/unknown` instead of collapsing uncertainty to one confidence number. A transparent one-step Bayes-risk router chooses whether to act or clarify from that distribution. This is POMDP-inspired uncertainty handling, not a full POMDP.
 3. **Conversation health state** — `repair` identifies observable user→assistant misalignment and its target; `interaction_quality` tracks structural conversation failure separately from emotion using current score, delta, repeated failures, unresolved repair, and clarification load.
-4. **Policy selection** — affect and intent jointly choose a response-policy family (`mirror`, `direct`, `repair`, `hold`, `boundary`, ...). Selection is a discrete decision, not generation.
+4. **Policy selection** — affect and intent jointly choose a response-policy family (`mirror`, `validate_redirect`, `direct`, `ask`, `deflate`, `repair`, `hold`, `boundary`). Selection is a discrete decision, not generation.
 5. **Factored action compilation** — the selected policy is split into orthogonal controls: `task_move`, `social_move`, `repair_move`, and `realization`. The task is driven primarily by user intent, so affective adaptation cannot silently replace what the user actually asked for.
 6. **Steering** — the factored action, source policy, repair state, interaction quality, and canonical request are injected into the decode path so the control decision actually lands in the tokens.
 7. **Structural response audit** — the generated response is checked against inspectable control contracts such as question budget, required clarification, unsupported privacy promises, unknown-stance admissions, repair repetition, and low-verbosity overruns. This is not a factual truth checker.

{eq_layer-0.2.3 → eq_layer-0.2.4}/README.md

@@ -16,7 +16,7 @@ So this repo does not retrain the model. It adds a layer:
 1. **Affect state** — `{valence, arousal, escalation_delta, stance, subtext}` as an explicit intermediate representation, tracked across turns. `escalation_delta` is the point: snapshot classification cannot see a trend.
 2. **Intent belief state** — the learned classifier preserves the full posterior over `action_request/status_check/explanation/question/statement/unknown` instead of collapsing uncertainty to one confidence number. A transparent one-step Bayes-risk router chooses whether to act or clarify from that distribution. This is POMDP-inspired uncertainty handling, not a full POMDP.
 3. **Conversation health state** — `repair` identifies observable user→assistant misalignment and its target; `interaction_quality` tracks structural conversation failure separately from emotion using current score, delta, repeated failures, unresolved repair, and clarification load.
-4. **Policy selection** — affect and intent jointly choose a response-policy family (`mirror`, `direct`, `repair`, `hold`, `boundary`, ...). Selection is a discrete decision, not generation.
+4. **Policy selection** — affect and intent jointly choose a response-policy family (`mirror`, `validate_redirect`, `direct`, `ask`, `deflate`, `repair`, `hold`, `boundary`). Selection is a discrete decision, not generation.
 5. **Factored action compilation** — the selected policy is split into orthogonal controls: `task_move`, `social_move`, `repair_move`, and `realization`. The task is driven primarily by user intent, so affective adaptation cannot silently replace what the user actually asked for.
 6. **Steering** — the factored action, source policy, repair state, interaction quality, and canonical request are injected into the decode path so the control decision actually lands in the tokens.
 7. **Structural response audit** — the generated response is checked against inspectable control contracts such as question budget, required clarification, unsupported privacy promises, unknown-stance admissions, repair repetition, and low-verbosity overruns. This is not a factual truth checker.

{eq_layer-0.2.3 → eq_layer-0.2.4}/eq_layer/affect.py

@@ -93,7 +93,9 @@ class HeuristicAffect:
         low = text.lower().strip()
         hits = sum(1 for m in ESCALATION_MARKERS if m in low)
         softens = sum(1 for s in SOFTENERS if s in low)
-        score = min(1.0, 0.25 + hits * 0.22 - softens * 0.12)
+        # Arousal is a [0, 1] scale: a softening-only turn must floor at 0, not
+        # go negative and hand the policy state an impossible value.
+        score = max(0.0, min(1.0, 0.25 + hits * 0.22 - softens * 0.12))
         if "?" in low or "!" in low:
             score = min(1.0, score + 0.1)
         return round(score, 3)
@@ -110,11 +112,13 @@ class HeuristicAffect:
 
     def _escalating(self, user_turns: list[str]) -> bool:
         """Three or more consecutive turns opening with a wh-word."""
-        if len(user_turns) < 3:
+        # Blank turns are not wh-word turns. Counting them let three turns in
+        # which only one or two actually opened with a wh-word read as a
+        # three-turn interrogation run. Same filter as ``_demand_pattern``.
+        recent = [t for t in user_turns[-3:] if t.strip()]
+        if len(recent) < 3:
             return False
-        return all(
-            self._first_word(t) in QUESTION_OPENERS for t in user_turns[-3:] if t.strip()
-        )
+        return all(self._first_word(t) in QUESTION_OPENERS for t in recent)
 
     def _first_word(self, text: str) -> str:
         """First token with punctuation stripped. 'Neden?' must match 'neden',

{eq_layer-0.2.3 → eq_layer-0.2.4}/eq_layer/response_experiment.py

@@ -9,7 +9,7 @@ import subprocess
 import threading
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Iterable
+from typing import Callable, Iterable
 
 from .auditor import audit_from_generation_metadata, regeneration_instruction
 from .policies import Selector
@@ -198,6 +198,17 @@ def stable_case_seed(global_seed: int, case_id: str) -> int:
     return int.from_bytes(digest[:4], "big", signed=False)
 
 
+# The experiment's own clock. Reproducibility is a property of the artifact, so
+# the wall clock must never leak into a generated record: the default is a
+# fixed epoch and callers who need a real timestamp inject their own clock.
+DEFAULT_EXPERIMENT_CLOCK_EPOCH = 0.0
+
+
+def _fixed_experiment_clock() -> float:
+    """Default experiment clock: a constant, so runs differ only by content."""
+    return DEFAULT_EXPERIMENT_CLOCK_EPOCH
+
+
 def file_sha256(path: str | Path) -> str:
     digest = hashlib.sha256()
     with open(path, "rb") as fh:
@@ -654,7 +665,18 @@ def generate_pairs(
     auditor_regenerations: int = 0,
     length_matched_arm: bool = False,
     length_profiles: tuple[str, ...] | None = None,
+    rng: random.Random | None = None,
+    clock: Callable[[], float] | None = None,
 ) -> list[dict]:
+    """Generate the same-model baseline/EQ pairs for one experiment run.
+
+    Generation is deterministic: the only randomness is the control-arm
+    shuffling, seeded from ``seed``, and it is quarantined behind the injectable
+    ``rng`` so a caller can supply the same generator without re-deriving it
+    from the seed. Any timestamp is quarantined behind ``clock``, whose default
+    is a fixed epoch, so the same experiment plus the same fixture serialises
+    to identical bytes.
+    """
     if auditor_regenerations not in {0, 1}:
         raise ValueError("auditor_regenerations must be 0 or 1.")
     profiles = tuple(length_profiles or ())
@@ -665,7 +687,8 @@ def generate_pairs(
                 f"{sorted(LENGTH_CONTROL_PROFILES)}"
             )
 
-    rng = random.Random(seed)
+    rng = rng if rng is not None else random.Random(seed)
+    clock = clock if clock is not None else _fixed_experiment_clock
     output: list[dict] = []
     seen: set[str] = set()
 
@@ -769,6 +792,7 @@ def generate_pairs(
                     "model_id": model.model_id,
                     "temperature": model.temperature,
                     "pair_seed": pair_seed,
+                    "generated_at_epoch": round(float(clock()), 6),
                     "order": ["baseline", "eq", *controls],
                     "same_model_both_arms": True,
                     "annotations_used": bool(annotated),

{eq_layer-0.2.3 → eq_layer-0.2.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "eq-layer"
-version = "0.2.3"
+version = "0.2.4"
 description = "Inspectable intent, affect, repair, uncertainty, and factored-action control for LLM responses"
 readme = "README.md"
 license = { text = "MIT" }