invarlock 0.3.7__py3-none-any.whl → 0.3.8__py3-none-any.whl

invarlock/core/runner.py

@@ -521,7 +521,7 @@ class CoreRunner:
                         f"Guard '{guard.name}' prepare failed: {e}"
                     ) from e
 
-        # Store resolved policies in report for certificate
+        # Store resolved policies in report for evaluation report generation
         report.meta["tier_policies"] = tier_policies
 
         self._log_event(

invarlock/edits/noop.py

@@ -1,7 +1,7 @@
-"""Built-in no-op edit used for baseline and Compare & Certify (BYOE).
+"""Built-in no-op edit used for baseline and Compare & Evaluate (BYOE).
 
 This edit does not modify the model and reports zero deltas. It exists to
-support baseline runs and Compare & Certify certification where the subject
+support baseline runs and Compare & Evaluate workflows where the subject
 checkpoint is produced outside of InvarLock.
 """
 

invarlock/edits/quant_rtn.py

@@ -406,7 +406,7 @@ class RTNQuantEdit(ModelEdit):
                         if layer_name not in modified_layers:
                             modified_layers.append(layer_name)
 
-            # Store edit plan for certificate generation
+            # Store edit plan for evaluation report generation
             modules_quantized = [r["module_name"] for r in quantization_results]
 
             edit_plan = {
@@ -424,7 +424,7 @@ class RTNQuantEdit(ModelEdit):
             return {
                 "name": self.name,
                 "plan_digest": f"rtn_quantization_{bitwidth}bit_{scope}",
-                "plan": edit_plan,  # Include the plan for certificate generation
+                "plan": edit_plan,  # Include the plan for evaluation report generation
                 "deltas": {
                     "params_changed": total_params_quantized,
                     "sparsity": None,  # Quantization doesn't create sparsity

invarlock/eval/__init__.py

@@ -1,7 +1,7 @@
 """Evaluation utilities (`invarlock.eval`).
 
 This package now focuses on evaluation helpers (metrics, data, probes).
-Reporting (report building, certificate, schema validation) has moved to
+Reporting (evaluation report building and schema validation) has moved to
 `invarlock.reporting`.
 """
 

invarlock/eval/bench.py

@@ -976,17 +976,21 @@ def execute_scenario(
     except Exception:
         pass
 
-    # Generate certificate artifact when both runs produced reports
+    # Generate evaluation report artifact when both runs produced reports
     try:
         if bare_result.success and guarded_result.success:
-            from invarlock.reporting.certificate import make_certificate
+            from invarlock.reporting.report_builder import make_report
 
-            cert = make_certificate(guarded_result.report, bare_result.report)
-            cert_path = scenario_dir / "certificate.json"
-            cert_path.write_text(json.dumps(cert, indent=2), encoding="utf-8")
-            artifacts["certificate"] = str(cert_path)
+            evaluation_report = make_report(guarded_result.report, bare_result.report)
+            report_path = scenario_dir / "evaluation.report.json"
+            report_path.write_text(
+                json.dumps(evaluation_report, indent=2), encoding="utf-8"
+            )
+            artifacts["evaluation_report"] = str(report_path)
     except Exception as exc:
-        logger.warning(f"Certificate generation failed for {scenario_slug}: {exc}")
+        logger.warning(
+            f"Evaluation report generation failed for {scenario_slug}: {exc}"
+        )
 
     # Resolve epsilon from runtime or use config
     epsilon_used = config.epsilon

invarlock/eval/primary_metric.py

@@ -2,7 +2,7 @@
 Primary metric abstraction and minimal ppl_causal implementation (Phase 1).
 
 This module introduces a light-weight, task-agnostic metric interface and a
-registry so the runner/certificate can evolve beyond causal-LM perplexity.
+registry so the runner and evaluation report builder can evolve beyond causal-LM perplexity.
 
 Phase 1 goal: provide a ppl_causal metric and a helper that can compute point
 estimates directly from evaluation window aggregates already present in run

invarlock/guards/spectral.py

@@ -344,7 +344,7 @@ class SpectralGuard(Guard):
                 self.ignore_preview_inflation = bool(policy["ignore_preview_inflation"])
                 self.config["ignore_preview_inflation"] = self.ignore_preview_inflation
 
-            # Optional hydration of baseline stats from policy (e.g., baseline certificate)
+            # Optional hydration of baseline stats from policy (e.g., baseline evaluation report)
             if "baseline_family_stats" in policy and isinstance(
                 policy["baseline_family_stats"], dict
             ):

invarlock/model_profile.py

@@ -6,11 +6,9 @@ from dataclasses import dataclass, field
 from typing import Any
 
 AutoTokenizer: Any | None = None
-GPT2Tokenizer: Any | None = None
 
 try:
     from transformers import AutoTokenizer as _AutoTokenizer
-    from transformers import GPT2Tokenizer as _GPT2Tokenizer
     from transformers.tokenization_utils_base import PreTrainedTokenizerBase
 except Exception:  # pragma: no cover - exercised only when transformers is absent
 
@@ -26,7 +24,6 @@ except Exception:  # pragma: no cover - exercised only when transformers is abse
 
 else:  # pragma: no cover - transformers optional
     AutoTokenizer = _AutoTokenizer
-    GPT2Tokenizer = _GPT2Tokenizer
 
 
 TokenizerFactory = Callable[[], tuple[PreTrainedTokenizerBase, str]]
@@ -177,12 +174,12 @@ def _make_bert_tokenizer(model_id: str):
 
 def _make_gpt2_tokenizer(model_id: str):
     def factory() -> tuple[PreTrainedTokenizerBase, str]:
-        if GPT2Tokenizer is None:
+        if AutoTokenizer is None:
             raise RuntimeError(
                 "GPT-2 tokenizers require the 'transformers' extra. "
                 "Install it with: pip install 'invarlock[adapters]'."
             )
-        tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+        tokenizer = AutoTokenizer.from_pretrained("gpt2")
         if tokenizer.pad_token is None:
             tokenizer.pad_token = tokenizer.eos_token
         hash_value = _hash_tokenizer(tokenizer)
@@ -193,7 +190,7 @@ def _make_gpt2_tokenizer(model_id: str):
 
 def _make_causal_auto_tokenizer(model_id: str):
     def factory() -> tuple[PreTrainedTokenizerBase, str]:
-        if AutoTokenizer is None and GPT2Tokenizer is None:
+        if AutoTokenizer is None:
             raise RuntimeError(
                 "Causal tokenizers require the 'transformers' extra. "
                 "Install it with: pip install 'invarlock[adapters]'."
@@ -202,23 +199,15 @@ def _make_causal_auto_tokenizer(model_id: str):
         # local GPT-2 tokenizer if the model assets are not cached or network
         # access is denied.
         tokenizer = None
-        if AutoTokenizer is not None:
+        try:
+            tokenizer = AutoTokenizer.from_pretrained(model_id, local_files_only=True)
+        except Exception:
             try:
-                tokenizer = AutoTokenizer.from_pretrained(
-                    model_id, local_files_only=True
-                )
+                tokenizer = AutoTokenizer.from_pretrained(model_id)
             except Exception:
-                try:
-                    tokenizer = AutoTokenizer.from_pretrained(model_id)
-                except Exception:
-                    tokenizer = None
+                tokenizer = None
         if tokenizer is None:
-            if GPT2Tokenizer is None:
-                raise RuntimeError(
-                    "Tokenization requires the 'transformers' extra. "
-                    "Install it with: pip install 'invarlock[adapters]'."
-                )
-            tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+            tokenizer = AutoTokenizer.from_pretrained("gpt2")
         # Ensure padding/bos tokens are configured so downstream encoding
         # yields stable non-zero ids and a valid attention mask regardless of
         # environment defaults or tokenizer variants.
@@ -246,30 +235,22 @@ def _make_causal_auto_tokenizer(model_id: str):
 
 def _make_unknown_tokenizer(model_id: str):
     def factory() -> tuple[PreTrainedTokenizerBase, str]:
-        if AutoTokenizer is None and GPT2Tokenizer is None:
+        if AutoTokenizer is None:
             raise RuntimeError(
                 "Text tokenization requires the 'transformers' extra. "
                 "Install it with: pip install 'invarlock[adapters]'."
             )
         # Unknown families: try local-only first, then remote, then degrade to GPT-2
         tokenizer = None
-        if AutoTokenizer is not None:
+        try:
+            tokenizer = AutoTokenizer.from_pretrained(model_id, local_files_only=True)
+        except Exception:
             try:
-                tokenizer = AutoTokenizer.from_pretrained(
-                    model_id, local_files_only=True
-                )
+                tokenizer = AutoTokenizer.from_pretrained(model_id)
             except Exception:
-                try:
-                    tokenizer = AutoTokenizer.from_pretrained(model_id)
-                except Exception:
-                    tokenizer = None
+                tokenizer = None
         if tokenizer is None:
-            if GPT2Tokenizer is None:
-                raise RuntimeError(
-                    "Text tokenization requires the 'transformers' extra. "
-                    "Install it with: pip install 'invarlock[adapters]'."
-                )
-            tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+            tokenizer = AutoTokenizer.from_pretrained("gpt2")
         if getattr(tokenizer, "pad_token", None) is None:
             eos_token = getattr(tokenizer, "eos_token", None)
             if eos_token is not None:

invarlock/plugins/hf_bnb_adapter.py

@@ -7,7 +7,7 @@ through Transformers. Requires GPU for practical use.
 Install with the `gpu` extra on supported platforms.
 
 This adapter handles both:
-1. Fresh quantization of FP16 models (load_in_8bit/load_in_4bit)
+1. Fresh quantization of FP16 models (via quantization_config)
 2. Loading pre-quantized BNB checkpoints (auto-detected via quantization_config)
 """
 
@@ -55,16 +55,17 @@ def _detect_pre_quantized_bnb(model_id: str) -> tuple[bool, int]:
         if not quant_cfg:
             return False, 0
 
-        # Check for BNB quantization
-        quant_method = quant_cfg.get("quant_method", "").lower()
-        if quant_method == "bitsandbytes" or "load_in_8bit" in quant_cfg:
+        # Check for BNB quantization. Prefer explicit bits, then legacy flags.
+        quant_method = str(quant_cfg.get("quant_method", "")).lower()
+        if "bitsandbytes" in quant_method or "bnb" in quant_method:
+            bits = quant_cfg.get("bits")
+            if isinstance(bits, int) and bits in {4, 8}:
+                return True, bits
             if quant_cfg.get("load_in_8bit"):
                 return True, 8
             if quant_cfg.get("load_in_4bit"):
                 return True, 4
-            # Fallback to bits field
-            bits = quant_cfg.get("bits", 8)
-            return True, bits
+            return True, 8
 
     except Exception:
         pass
@@ -82,11 +83,17 @@ class HF_BNB_Adapter(HFAdapterMixin, ModelAdapter):
             "DEPENDENCY-MISSING: transformers",
             lambda e: {"dependency": "transformers"},
         ):
-            from transformers import AutoModelForCausalLM
+            from transformers import AutoModelForCausalLM, BitsAndBytesConfig
 
         # Check if this is a pre-quantized checkpoint
         is_pre_quantized, pre_quant_bits = _detect_pre_quantized_bnb(model_id)
 
+        if "load_in_8bit" in kwargs or "load_in_4bit" in kwargs:
+            raise ValueError(
+                "hf_bnb adapter: load_in_8bit/load_in_4bit are not supported. "
+                "Use model.quantization_config instead."
+            )
+
         if is_pre_quantized:
             # Load pre-quantized checkpoint WITHOUT re-applying quantization
             with wrap_errors(
@@ -99,20 +106,25 @@ class HF_BNB_Adapter(HFAdapterMixin, ModelAdapter):
                     model_id,
                     device_map="auto",
                     trust_remote_code=True,
-                    # Do NOT pass load_in_8bit/load_in_4bit for pre-quantized
-                    **{
-                        k: v
-                        for k, v in kwargs.items()
-                        if k not in ("load_in_8bit", "load_in_4bit")
-                    },
+                    **kwargs,
                 )
         else:
             # Fresh quantization of FP16 model
-            load_in_8bit = bool(kwargs.pop("load_in_8bit", True))
-            load_in_4bit = bool(kwargs.pop("load_in_4bit", False))
-
-            if load_in_4bit:
-                load_in_8bit = False
+            quantization_config = kwargs.pop("quantization_config", None)
+            if quantization_config is None:
+                quantization_config = BitsAndBytesConfig(load_in_8bit=True)
+            elif isinstance(quantization_config, dict):
+                qdict = dict(quantization_config)
+                bits = qdict.pop("bits", None)
+                qdict.pop("quant_method", None)
+                if isinstance(bits, int):
+                    if bits == 4:
+                        qdict.setdefault("load_in_4bit", True)
+                        qdict.setdefault("load_in_8bit", False)
+                    elif bits == 8:
+                        qdict.setdefault("load_in_8bit", True)
+                        qdict.setdefault("load_in_4bit", False)
+                quantization_config = BitsAndBytesConfig(**qdict)
 
             with wrap_errors(
                 ModelLoadError,
@@ -123,9 +135,8 @@ class HF_BNB_Adapter(HFAdapterMixin, ModelAdapter):
                 model = AutoModelForCausalLM.from_pretrained(
                     model_id,
                     device_map="auto",
-                    load_in_8bit=load_in_8bit,
-                    load_in_4bit=load_in_4bit,
                     trust_remote_code=True,
+                    quantization_config=quantization_config,
                     **kwargs,
                 )
 

invarlock/reporting/__init__.py

@@ -1,7 +1,21 @@
 """
-Reporting package (aliasing evaluation reporting modules).
+Evaluation report tooling (`invarlock.reporting`).
 
-Provides a stable unified import path for report schema, builders,
-and certificate generation while keeping backward compatibility with
-`invarlock.eval.*` imports.
+Provides the evaluation report schema, builder, and renderers.
 """
+
+from __future__ import annotations
+
+from .html import render_report_html
+from .render import render_report_markdown
+from .report_builder import make_report, validate_report
+from .report_schema import REPORT_JSON_SCHEMA, REPORT_SCHEMA_VERSION
+
+__all__ = [
+    "REPORT_SCHEMA_VERSION",
+    "REPORT_JSON_SCHEMA",
+    "make_report",
+    "render_report_markdown",
+    "render_report_html",
+    "validate_report",
+]

invarlock/reporting/html.py

@@ -1,5 +1,5 @@
 """
-Minimal HTML exporter for certificates.
+Minimal HTML exporter for reports.
 
 This implementation wraps the Markdown rendering in a simple HTML template so
 that the numbers and core content remain identical across formats.
@@ -10,7 +10,7 @@ from __future__ import annotations
 from html import escape
 from typing import Any
 
-from .render import render_certificate_markdown
+from .render import render_report_markdown
 
 markdown_module: Any | None = None
 try:
@@ -37,13 +37,13 @@ def _apply_status_badges(html_body: str) -> str:
     return updated
 
 
-def render_certificate_html(certificate: dict[str, Any]) -> str:
-    """Render a certificate as a simple HTML document.
+def render_report_html(evaluation_report: dict[str, Any]) -> str:
+    """Render an evaluation report as a simple HTML document.
 
     Uses the Markdown renderer and converts to HTML when available, falling back
     to a <pre> block when the markdown dependency is missing.
     """
-    md = render_certificate_markdown(certificate)
+    md = render_report_markdown(evaluation_report)
     if markdown_module is None:
         body = f'<pre class="invarlock-md">{escape(md)}</pre>'
     else:
@@ -52,7 +52,7 @@ def render_certificate_html(certificate: dict[str, Any]) -> str:
         body = f'<div class="invarlock-md">{html_body}</div>'
     return (
         '<!DOCTYPE html><html><head><meta charset="utf-8">'
-        "<title>InvarLock Evaluation Certificate</title>"
+        "<title>InvarLock Evaluation Report</title>"
         "<style>"
         ":root{--pass:#2da44e;--fail:#cf222e;--warn:#bf8700;--ink:#1f2328;"
         "--muted:#57606a;--panel:#f6f8fa;--border:#d0d7de}"
@@ -79,4 +79,4 @@ def render_certificate_html(certificate: dict[str, Any]) -> str:
     )
 
 
-__all__ = ["render_certificate_html"]
+__all__ = ["render_report_html"]

invarlock/reporting/normalizer.py

@@ -33,7 +33,7 @@ def normalize_run_report(report: Mapping[str, Any] | RunReport) -> RunReport:
     """Coerce an arbitrary report-like mapping into a canonical RunReport.
 
     This is the single entry point for converting pre-canonical or loosely-typed
-    data into the strict PM-only RunReport shape used by certificate/report.
+    data into the strict PM-only RunReport shape used by evaluation report building.
     """
     src = _as_mapping(report)
 
@@ -53,7 +53,7 @@ def normalize_run_report(report: Mapping[str, Any] | RunReport) -> RunReport:
         "ts": ts,
         "auto": meta_in.get("auto") if isinstance(meta_in.get("auto"), dict) else None,
     }
-    # Preserve additional provenance knobs used by certificate/digests.
+    # Preserve additional provenance knobs used by evaluation report digests.
     for key in (
         "pm_acceptance_range",
         "pm_drift_band",

invarlock/reporting/policy_utils.py

@@ -35,7 +35,7 @@ def _compute_thresholds_payload(
     tier: str, resolved_policy: dict[str, Any]
 ) -> dict[str, Any]:
     """Build canonical thresholds payload for digest stability."""
-    from .certificate import TIER_RATIO_LIMITS  # local to avoid cycles
+    from .report_builder import TIER_RATIO_LIMITS  # local to avoid cycles
 
     tier_lc = (tier or "balanced").lower()
     metrics_policy = (

invarlock/reporting/primary_metric_utils.py

@@ -8,21 +8,21 @@ from .utils import _coerce_interval, _weighted_mean
 
 
 def attach_primary_metric(
-    certificate: dict[str, Any],
+    evaluation_report: dict[str, Any],
     report: dict[str, Any],
     baseline_raw: dict[str, Any] | None,
     baseline_ref: dict[str, Any] | None,
     ppl_analysis: dict[str, Any] | None,
 ) -> None:
-    """Attach/normalize the primary_metric block on the certificate.
+    """Attach/normalize the primary_metric block on the evaluation report.
 
-    Behavior mirrors historical logic in certificate.py and preserves structure:
+    Behavior mirrors historical logic in report_builder.py and preserves structure:
     - Prefer explicit metrics.primary_metric if present
     - Compute missing ratio_vs_baseline, degenerate display_ci
     - ppl window-based analysis info (mean logloss) added when available
     - Fallbacks for classification metrics and eval-window-derived ppl
     - Ensure display_ci always present for schema invariants
-    Mutates the certificate in-place.
+    Mutates the evaluation report in-place.
     """
     # Attach primary metric snapshot when provided in report
     try:
@@ -180,12 +180,12 @@ def attach_primary_metric(
                     ]
             except Exception:
                 pass
-            certificate["primary_metric"] = pm_copy
+            evaluation_report["primary_metric"] = pm_copy
     except Exception:
         pass
 
     def _attach_from_windows() -> None:
-        if isinstance(certificate.get("primary_metric"), dict):
+        if isinstance(evaluation_report.get("primary_metric"), dict):
             return
         try:
             m = (
@@ -212,7 +212,7 @@ def attach_primary_metric(
                 baseline=baseline_raw if isinstance(baseline_raw, dict) else None,
             )
             if isinstance(pm_block, dict) and pm_block:
-                certificate["primary_metric"] = pm_block
+                evaluation_report["primary_metric"] = pm_block
         except Exception:
             pass
 
@@ -220,7 +220,7 @@ def attach_primary_metric(
     _attach_from_windows()
 
     # Minimal fallback for classification-only reports without explicit primary_metric
-    if not isinstance(certificate.get("primary_metric"), dict):
+    if not isinstance(evaluation_report.get("primary_metric"), dict):
         try:
             metrics_map = report.get("metrics", {}) if isinstance(report, dict) else {}
             clf = (
@@ -298,7 +298,7 @@ def attach_primary_metric(
                         acc_pm["ratio_vs_baseline"] = delta_pp
                 except Exception:
                     pass
-                certificate["primary_metric"] = acc_pm
+                evaluation_report["primary_metric"] = acc_pm
         except Exception:
             pass
 
@@ -308,8 +308,8 @@ def attach_primary_metric(
     # Ensure primary_metric has display_ci populated for schema invariants
     try:
         pm = (
-            certificate.get("primary_metric", {})
-            if isinstance(certificate.get("primary_metric"), dict)
+            evaluation_report.get("primary_metric", {})
+            if isinstance(evaluation_report.get("primary_metric"), dict)
             else None
         )
         if isinstance(pm, dict) and pm: