invarlock 0.3.5__py3-none-any.whl → 0.3.7__py3-none-any.whl

invarlock/__init__.py

@@ -6,13 +6,13 @@ Core runtime package — torch-independent utilities, configuration, and interfa
 
 This package provides the foundation for the InvarLock GuardChain without heavy dependencies.
 For torch-dependent functionality, see subpackages under `invarlock.*`:
-- `invarlock.adapters`: Model adapters (HF GPT-2/BERT/LLaMA, auto)
+- `invarlock.adapters`: Model adapters (HF causal/MLM/seq2seq + auto)
 - `invarlock.guards`: Safety mechanisms (invariants, spectral, RMT, variance)
 - `invarlock.edits`: Built-in quantization and edit interfaces
 - `invarlock.eval`: Metrics, guard-overhead checks, and certification
 """
 
-__version__ = "0.3.5"
+__version__ = "0.3.7"
 
 # Core exports - torch-independent
 from .config import CFG, Defaults, get_default_config

invarlock/_data/runtime/tiers.yaml

@@ -1,8 +1,12 @@
-# Tier guard policy knobs for variance correction (balanced vs conservative)
+# Tier policy defaults (metrics gates + guard knobs) used at runtime.
 #
-# These values mirror the settings validated during the December 2025
-# calibration runs. They should be kept in sync with the policy digest
-# embedded in certificates and referenced by automation documentation.
+# Balanced and Conservative values are calibrated/validated against pilot/null
+# runs (Nov/Dec 2025) where applicable; Aggressive is research-oriented (not in
+# the safety case).
+#
+# Rationale by key: docs/reference/tier-policy-catalog.md
+# Calibration method: docs/assurance/09-tier-v1-calibration.md
+# Provenance/digest: docs/assurance/11-policy-provenance.md
 
 balanced:
   metrics:
@@ -11,6 +15,13 @@ balanced:
       min_tokens: 50000
       hysteresis_ratio: 0.002
       min_token_fraction: 0.01
+    pm_tail:
+      mode: warn
+      min_windows: 50
+      quantile: 0.95
+      quantile_max: 0.20
+      epsilon: 0.0001
+      mass_max: 1.0
     accuracy:
       delta_min_pp: -1.0
       min_examples: 200
@@ -20,7 +31,7 @@ balanced:
     deadband: 0.02
     min_abs_adjust: 0.012
     max_scale_step: 0.03
-    min_effect_lognll: 0.0009
+    min_effect_lognll: 0.0
     predictive_one_sided: true
     topk_backstop: 1
     max_adjusted_modules: 1
@@ -33,10 +44,10 @@ balanced:
     max_caps: 5
     max_spectral_norm: null
     family_caps:
-      ffn: 3.834
-      attn: 3.423
-      embed: 3.1
-      other: 3.1
+      ffn: 3.849
+      attn: 3.018
+      embed: 1.05
+      other: 0.0
     multiple_testing:
       method: bh
       alpha: 0.05
@@ -44,12 +55,12 @@ balanced:
   rmt_guard:
     deadband: 0.10
     margin: 1.5
-    epsilon_default: 0.10
+    epsilon_default: 0.01
     epsilon_by_family:
-      ffn: 0.10
-      attn: 0.08
-      embed: 0.12
-      other: 0.12
+      ffn: 0.01
+      attn: 0.01
+      embed: 0.01
+      other: 0.01
 
 conservative:
   metrics:
@@ -58,6 +69,13 @@ conservative:
       min_tokens: 20000
       hysteresis_ratio: 0.002
       min_token_fraction: 0.01
+    pm_tail:
+      mode: warn
+      min_windows: 50
+      quantile: 0.95
+      quantile_max: 0.12
+      epsilon: 0.0001
+      mass_max: 1.0
     accuracy:
       delta_min_pp: -0.5
       min_examples: 200
@@ -67,7 +85,7 @@ conservative:
     deadband: 0.03
     min_abs_adjust: 0.02
     max_scale_step: 0.015
-    min_effect_lognll: 0.0018
+    min_effect_lognll: 0.016
     predictive_one_sided: false
     topk_backstop: 0
     max_adjusted_modules: 0
@@ -78,24 +96,25 @@ conservative:
     deadband: 0.05
     scope: ffn
     max_caps: 3
+    max_spectral_norm: null
     family_caps:
-      ffn: 2.3
+      ffn: 3.849
       attn: 2.6
       embed: 2.8
       other: 2.8
     multiple_testing:
       method: bonferroni
-      alpha: 0.02
+      alpha: 0.000625
       m: 4
   rmt_guard:
     deadband: 0.05
     margin: 1.3
-    epsilon_default: 0.06
+    epsilon_default: 0.01
     epsilon_by_family:
-      ffn: 0.06
-      attn: 0.05
-      embed: 0.07
-      other: 0.07
+      ffn: 0.01
+      attn: 0.01
+      embed: 0.01
+      other: 0.01
 
 aggressive:
   metrics:
@@ -104,6 +123,13 @@ aggressive:
       min_tokens: 50000
       hysteresis_ratio: 0.002
       min_token_fraction: 0.01
+    pm_tail:
+      mode: warn
+      min_windows: 50
+      quantile: 0.95
+      quantile_max: 0.30
+      epsilon: 0.0001
+      mass_max: 1.0
     accuracy:
       delta_min_pp: -2.0
       min_examples: 200
@@ -111,27 +137,28 @@ aggressive:
       min_examples_fraction: 0.01
   variance_guard:
     deadband: 0.12
-    min_effect_lognll: 0.0005
+    min_effect_lognll: 0.033
   spectral_guard:
     sigma_quantile: 0.98
     deadband: 0.15
     scope: ffn
     max_caps: 8
+    max_spectral_norm: null
     family_caps:
-      ffn: 3.0
+      ffn: 3.849
       attn: 3.5
       embed: 2.5
       other: 3.5
     multiple_testing:
       method: bh
-      alpha: 0.1
+      alpha: 0.00078125
       m: 4
   rmt_guard:
     deadband: 0.15
     margin: 1.8
-    epsilon_default: 0.15
+    epsilon_default: 0.01
     epsilon_by_family:
-      ffn: 0.15
-      attn: 0.15
-      embed: 0.15
-      other: 0.15
+      ffn: 0.01
+      attn: 0.01
+      embed: 0.01
+      other: 0.01

invarlock/adapters/__init__.py

@@ -29,13 +29,11 @@ from .capabilities import (
 )
 
 _LAZY_MAP = {
-    "HF_BERT_Adapter": ".hf_bert",
-    "HF_GPT2_Adapter": ".hf_gpt2",
-    "HF_LLaMA_Adapter": ".hf_llama",
-    "HF_T5_Adapter": ".hf_t5",
-    "HF_ORT_CausalLM_Adapter": ".hf_onnx",
-    "HF_Causal_Auto_Adapter": ".auto",
-    "HF_MLM_Auto_Adapter": ".auto",
+    "HF_Causal_Adapter": ".hf_causal",
+    "HF_MLM_Adapter": ".hf_mlm",
+    "HF_Seq2Seq_Adapter": ".hf_seq2seq",
+    "HF_Causal_ONNX_Adapter": ".hf_causal_onnx",
+    "HF_Auto_Adapter": ".auto",
 }
 
 
@@ -76,7 +74,7 @@ class _RemovedComponent:
         return _RemovedComponent(self._name, self._replacement)
 
 
-# Placeholders for removed/legacy utilities referenced in tests
+# Placeholders for removed utilities referenced in tests
 HF_Pythia_Adapter = _RemovedComponent("HF_Pythia_Adapter")
 auto_tune_pruning_budget = _RemovedComponent("auto_tune_pruning_budget")
 run_auto_invarlock = _RemovedComponent("run_auto_invarlock")
@@ -91,13 +89,11 @@ run_invarlock = _RemovedComponent("run_invarlock", "invarlock.cli.run")
 quick_prune_gpt2 = _RemovedComponent("quick_prune_gpt2")
 
 __all__ = [
-    "HF_GPT2_Adapter",
-    "HF_BERT_Adapter",
-    "HF_LLaMA_Adapter",
-    "HF_T5_Adapter",
-    "HF_ORT_CausalLM_Adapter",
-    "HF_Causal_Auto_Adapter",
-    "HF_MLM_Auto_Adapter",
+    "HF_Causal_Adapter",
+    "HF_MLM_Adapter",
+    "HF_Seq2Seq_Adapter",
+    "HF_Causal_ONNX_Adapter",
+    "HF_Auto_Adapter",
     "BaseAdapter",
     "AdapterConfig",
     "AdapterInterface",

invarlock/adapters/auto.py

@@ -110,21 +110,26 @@ class _DelegatingAdapter(ModelAdapter):
 
     def _load_adapter(self, adapter_name: str) -> ModelAdapter:
         """Load an adapter by name."""
-        if adapter_name == "hf_llama":
-            HF_LLaMA_Adapter = _importlib.import_module(
-                ".hf_llama", __package__
-            ).HF_LLaMA_Adapter
-            return HF_LLaMA_Adapter()
-        elif adapter_name == "hf_bert":
-            HF_BERT_Adapter = _importlib.import_module(
-                ".hf_bert", __package__
-            ).HF_BERT_Adapter
-            return HF_BERT_Adapter()
-        elif adapter_name == "hf_gpt2":
-            HF_GPT2_Adapter = _importlib.import_module(
-                ".hf_gpt2", __package__
-            ).HF_GPT2_Adapter
-            return HF_GPT2_Adapter()
+        if adapter_name == "hf_causal":
+            HF_Causal_Adapter = _importlib.import_module(
+                ".hf_causal", __package__
+            ).HF_Causal_Adapter
+            return HF_Causal_Adapter()
+        if adapter_name == "hf_mlm":
+            HF_MLM_Adapter = _importlib.import_module(
+                ".hf_mlm", __package__
+            ).HF_MLM_Adapter
+            return HF_MLM_Adapter()
+        if adapter_name == "hf_seq2seq":
+            HF_Seq2Seq_Adapter = _importlib.import_module(
+                ".hf_seq2seq", __package__
+            ).HF_Seq2Seq_Adapter
+            return HF_Seq2Seq_Adapter()
+        if adapter_name == "hf_causal_onnx":
+            HF_Causal_ONNX_Adapter = _importlib.import_module(
+                ".hf_causal_onnx", __package__
+            ).HF_Causal_ONNX_Adapter
+            return HF_Causal_ONNX_Adapter()
         elif adapter_name == "hf_bnb":
             HF_BNB_Adapter = _importlib.import_module(
                 "invarlock.plugins.hf_bnb_adapter"
@@ -141,11 +146,11 @@ class _DelegatingAdapter(ModelAdapter):
             ).HF_GPTQ_Adapter
             return HF_GPTQ_Adapter()
         else:
-            # Default to GPT2 adapter
-            HF_GPT2_Adapter = _importlib.import_module(
-                ".hf_gpt2", __package__
-            ).HF_GPT2_Adapter
-            return HF_GPT2_Adapter()
+            # Default to causal adapter
+            HF_Causal_Adapter = _importlib.import_module(
+                ".hf_causal", __package__
+            ).HF_Causal_Adapter
+            return HF_Causal_Adapter()
 
     def _ensure_delegate_from_id(self, model_id: str) -> ModelAdapter:
         if self._delegate is not None:
@@ -172,14 +177,16 @@ class _DelegatingAdapter(ModelAdapter):
             self._delegate = self._load_adapter(quant_adapter)
             return self._delegate
 
-        # Fall back to class name inspection
+        # Fall back to lightweight class-name inspection (no transformers import).
         cls_name = getattr(model, "__class__", type(model)).__name__.lower()
-        if any(k in cls_name for k in ["llama", "mistral", "qwen", "yi"]):
-            self._delegate = self._load_adapter("hf_llama")
-        elif any(k in cls_name for k in ["bert", "roberta", "albert", "deberta"]):
-            self._delegate = self._load_adapter("hf_bert")
+        if any(k in cls_name for k in ["bert", "roberta", "albert", "deberta"]):
+            self._delegate = self._load_adapter("hf_mlm")
         else:
-            self._delegate = self._load_adapter("hf_gpt2")
+            cfg = getattr(model, "config", None)
+            if getattr(cfg, "is_encoder_decoder", False):
+                self._delegate = self._load_adapter("hf_seq2seq")
+            else:
+                self._delegate = self._load_adapter("hf_causal")
         return self._delegate
 
     def can_handle(self, model: Any) -> bool:  # pragma: no cover - trivial
@@ -206,21 +213,9 @@ class _DelegatingAdapter(ModelAdapter):
         raise AttributeError(item)
 
 
-class HF_Causal_Auto_Adapter(_DelegatingAdapter):
-    name = "hf_causal_auto"
+class HF_Auto_Adapter(_DelegatingAdapter):
+    name = "hf_auto"
 
     def load_model(self, model_id: str, device: str = "auto", **kwargs: Any) -> Any:
         delegate = self._ensure_delegate_from_id(model_id)
         return delegate.load_model(model_id, device=device, **kwargs)
-
-
-class HF_MLM_Auto_Adapter(_DelegatingAdapter):
-    name = "hf_mlm_auto"
-
-    def load_model(self, model_id: str, device: str = "auto", **kwargs: Any) -> Any:
-        # Force BERT-like adapter for MLM families
-        HF_BERT_Adapter = _importlib.import_module(
-            ".hf_bert", __package__
-        ).HF_BERT_Adapter
-        self._delegate = HF_BERT_Adapter()
-        return self._delegate.load_model(model_id, device=device, **kwargs)

invarlock/adapters/capabilities.py

@@ -359,7 +359,7 @@ def _detect_weight_tying(model: Any) -> dict[str, str]:
     tying: dict[str, str] = {}
 
     # Common weight tying patterns
-    # LLaMA/Mistral: lm_head.weight ↔ model.embed_tokens.weight
+    # Decoder embed_tokens style: lm_head.weight ↔ model.embed_tokens.weight
     if hasattr(model, "lm_head") and hasattr(model, "model"):
         inner = model.model
         if hasattr(inner, "embed_tokens"):
@@ -408,7 +408,7 @@ def _detect_primary_metric(model: Any) -> str:
             return "rouge"
         return "ppl_seq2seq"
 
-    # Decoder-only models (GPT-like, LLaMA-like)
+    # Decoder-only models (GPT-like, RoPE-style)
     return "ppl_causal"