invarlock 0.3.0__tar.gz → 0.3.1__tar.gz

{invarlock-0.3.0/src/invarlock.egg-info → invarlock-0.3.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: invarlock
-Version: 0.3.0
+Version: 0.3.1
 Summary: Edit‑agnostic robustness certificates for weight edits (InvarLock framework)
 Author-email: InvarLock Team <oss@invarlock.dev>
 Maintainer-email: InvarLock Maintainers <support@invarlock.dev>
@@ -112,7 +112,7 @@ they don’t, roll back safely.
 Technical: edit‑agnostic guard pipeline (invariants → spectral → RMT →
 variance) producing a machine‑readable Safety Certificate.
 
-> **Status:** 0.3.0 (pre‑1.0). Until 1.0, **minor** releases may be
+> **Status:** 0.3.1 (pre‑1.0). Until 1.0, **minor** releases may be
 > breaking. See CLI help and the CHANGELOG for updates.
 
 [![CI](https://img.shields.io/github/actions/workflow/status/invarlock/invarlock/ci.yml?branch=main&logo=github&label=CI)](https://github.com/invarlock/invarlock/actions/workflows/ci.yml)

{invarlock-0.3.0 → invarlock-0.3.1}/README.md

@@ -6,7 +6,7 @@ they don’t, roll back safely.
 Technical: edit‑agnostic guard pipeline (invariants → spectral → RMT →
 variance) producing a machine‑readable Safety Certificate.
 
-> **Status:** 0.3.0 (pre‑1.0). Until 1.0, **minor** releases may be
+> **Status:** 0.3.1 (pre‑1.0). Until 1.0, **minor** releases may be
 > breaking. See CLI help and the CHANGELOG for updates.
 
 [![CI](https://img.shields.io/github/actions/workflow/status/invarlock/invarlock/ci.yml?branch=main&logo=github&label=CI)](https://github.com/invarlock/invarlock/actions/workflows/ci.yml)

{invarlock-0.3.0 → invarlock-0.3.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "invarlock"
-version = "0.3.0"
+version = "0.3.1"
 description = "Edit‑agnostic robustness certificates for weight edits (InvarLock framework)"
 authors = [{ name = "InvarLock Team", email = "oss@invarlock.dev" }]
 maintainers = [{ name = "InvarLock Maintainers", email = "support@invarlock.dev" }]

{invarlock-0.3.0 → invarlock-0.3.1}/src/invarlock/__init__.py

@@ -12,7 +12,7 @@ For torch-dependent functionality, see subpackages under `invarlock.*`:
 - `invarlock.eval`: Metrics, guard-overhead checks, and certification
 """
 
-__version__ = "0.3.0"
+__version__ = "0.3.1"
 
 # Core exports - torch-independent
 from .config import CFG, Defaults, get_default_config

{invarlock-0.3.0 → invarlock-0.3.1}/src/invarlock/_data/runtime/profiles/ci_cpu.yaml

@@ -11,5 +11,10 @@ dataset:
   final_n: 120
   stride: 512
 
+primary_metric:
+  acceptance_range:
+    min: 0.95
+    max: 1.15
+
 context:
   telemetry_profile: "ci_cpu"

{invarlock-0.3.0 → invarlock-0.3.1}/src/invarlock/cli/commands/doctor.py

@@ -326,8 +326,14 @@ def doctor_command(
     try:
         import torch
 
+        torch_version = getattr(torch, "__version__", None)
         if not json_out:
-            console.print(f"[green]✅ PyTorch {torch.__version__}[/green]")
+            if torch_version:
+                console.print(f"[green]✅ PyTorch {torch_version}[/green]")
+            else:
+                console.print(
+                    "[yellow]⚠️  PyTorch present but version unavailable[/yellow]"
+                )
 
         # Device information
         from ..device import get_device_info

{invarlock-0.3.0 → invarlock-0.3.1}/src/invarlock/cli/commands/run.py

@@ -81,6 +81,137 @@ GUARD_OVERHEAD_THRESHOLD = 0.01
 SPLIT_ALIASES: tuple[str, ...] = ("validation", "val", "dev", "eval", "test")
 
 
+def _coerce_mapping(obj: object) -> dict[str, Any]:
+    """Best-effort conversion of config-like objects to plain dicts."""
+
+    if isinstance(obj, dict):
+        return obj
+    try:
+        raw = getattr(obj, "_data", None)
+        if isinstance(raw, dict):
+            return raw
+    except Exception:
+        pass
+    try:
+        dumped = obj.model_dump()  # type: ignore[attr-defined]
+        if isinstance(dumped, dict):
+            return dumped
+    except Exception:
+        pass
+    try:
+        data = vars(obj)
+        if isinstance(data, dict):
+            return data
+    except Exception:
+        pass
+    return {}
+
+
+def _resolve_pm_acceptance_range(
+    cfg: InvarLockConfig | dict[str, Any] | None,
+) -> dict[str, float]:
+    """Resolve primary-metric acceptance bounds from config/env with safe defaults."""
+
+    base_min = 0.95
+    base_max = 1.10
+
+    cfg_min = None
+    cfg_max = None
+    try:
+        cfg_map = _coerce_mapping(cfg) if cfg is not None else {}
+        pm_section = cfg_map.get("primary_metric") if isinstance(cfg_map, dict) else {}
+        pm_map = _coerce_mapping(pm_section)
+        acceptance = (
+            pm_map.get("acceptance_range") if isinstance(pm_map, dict) else None
+        )
+        if isinstance(acceptance, dict):
+            if acceptance.get("min") is not None:
+                try:
+                    cfg_min = float(acceptance["min"])
+                except (TypeError, ValueError):
+                    cfg_min = None
+            if acceptance.get("max") is not None:
+                try:
+                    cfg_max = float(acceptance["max"])
+                except (TypeError, ValueError):
+                    cfg_max = None
+    except Exception:
+        cfg_min = None
+        cfg_max = None
+
+    def _parse_env(name: str) -> float | None:
+        try:
+            raw = os.environ.get(name, "")
+            if raw is None or str(raw).strip() == "":
+                return None
+            return float(raw)
+        except Exception:
+            return None
+
+    env_min = _parse_env("INVARLOCK_PM_ACCEPTANCE_MIN")
+    env_max = _parse_env("INVARLOCK_PM_ACCEPTANCE_MAX")
+
+    has_explicit = any(v is not None for v in (cfg_min, cfg_max, env_min, env_max))
+    if not has_explicit:
+        return {}
+
+    min_val = (
+        env_min if env_min is not None else cfg_min if cfg_min is not None else base_min
+    )
+    max_val = (
+        env_max if env_max is not None else cfg_max if cfg_max is not None else base_max
+    )
+
+    try:
+        if min_val is not None and min_val <= 0:
+            min_val = base_min
+    except Exception:
+        min_val = base_min
+    try:
+        if max_val is not None and max_val <= 0:
+            max_val = base_max
+    except Exception:
+        max_val = base_max
+
+    try:
+        if max_val is not None and min_val is not None and max_val < min_val:
+            max_val = min_val
+    except Exception:
+        max_val = base_max
+
+    return {"min": float(min_val), "max": float(max_val)}
+
+
+def _free_model_memory(model: object | None) -> None:
+    """Best-effort cleanup to release GPU memory for a model object."""
+    if model is None:
+        return
+    try:
+        import gc
+
+        del model
+        gc.collect()
+        if torch is not None and torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+    except Exception:
+        # Cleanup should never raise; fallback is to proceed without cache purge
+        pass
+
+
+def _should_measure_overhead(profile_normalized: str) -> tuple[bool, bool]:
+    """Return (measure_guard_overhead, skip_overhead) derived from env/profile."""
+
+    skip_overhead_env = (
+        os.environ.get("INVARLOCK_SKIP_OVERHEAD_CHECK", "").strip().lower()
+    )
+    skip_overhead = skip_overhead_env in {"1", "true", "yes"}
+    measure_guard_overhead = (
+        profile_normalized in {"ci", "release"} and not skip_overhead
+    )
+    return measure_guard_overhead, skip_overhead
+
+
 def _choose_dataset_split(
     *, requested: str | None, available: list[str] | None
 ) -> tuple[str, bool]:
@@ -1671,6 +1802,7 @@ def run_command(
             "edit": edit_meta,
             "guards": guard_metadata,
         }
+        pm_acceptance_range = _resolve_pm_acceptance_range(cfg)
 
         console.print(f"🔌 Adapter: {adapter.name}")
 
@@ -1746,6 +1878,10 @@ def run_command(
             "plugins": plugin_provenance,
             "run_id": run_id,
         }
+        run_context.setdefault("primary_metric", {})["acceptance_range"] = (
+            pm_acceptance_range
+        )
+        run_context["pm_acceptance_range"] = pm_acceptance_range
         run_context["model_profile"] = {
             "family": model_profile.family,
             "default_loss": model_profile.default_loss,
@@ -2756,18 +2892,26 @@ def run_command(
 
                 restore_fn = _restore2
             else:
-                # reload path
+                # reload path - properly free GPU memory before setting to None
+                _free_model_memory(model)
                 model = None
                 restore_fn = None
         except Exception:
             # On any failure, fall back to reload-per-attempt path
+            _free_model_memory(model)
             model = None
             restore_fn = None
 
         # RETRY LOOP - All report processing inside loop
         attempt = 1
         profile_normalized = (profile or "").lower()
-        measure_guard_overhead = profile_normalized in {"ci", "release"}
+        measure_guard_overhead, skip_overhead = _should_measure_overhead(
+            profile_normalized
+        )
+        if skip_overhead and profile_normalized in {"ci", "release"}:
+            console.print(
+                "[yellow]⚠️  Overhead check skipped via INVARLOCK_SKIP_OVERHEAD_CHECK[/yellow]"
+            )
 
         while True:
             # Reset RNG streams each attempt to guarantee determinism across retries
@@ -2933,6 +3077,8 @@ def run_command(
             if env_flags:
                 meta_payload["env_flags"] = env_flags
             report["meta"].update(meta_payload)
+            if pm_acceptance_range:
+                report["meta"]["pm_acceptance_range"] = pm_acceptance_range
             report["meta"]["model_profile"] = {
                 "family": model_profile.family,
                 "default_loss": model_profile.default_loss,

{invarlock-0.3.0 → invarlock-0.3.1}/src/invarlock/reporting/certificate.py

@@ -13,6 +13,7 @@ from __future__ import annotations
 # mypy: ignore-errors
 import copy
 import hashlib
+import inspect
 import json
 import math
 import os
@@ -1322,24 +1323,36 @@ def make_certificate(
         capacity_tokens = None
         capacity_examples = None
 
-    validation_flags = _compute_validation_flags(
-        ppl_analysis,
-        spectral,
-        rmt,
-        invariants,
-        auto.get("tier", "balanced"),
-        ppl_metrics,
-        auto.get("target_pm_ratio"),
-        guard_overhead_section,
-        report.get("metrics", {}).get("primary_metric")
+    pm_acceptance_range = _resolve_pm_acceptance_range_from_report(report)
+
+    validation_kwargs = {
+        "ppl": ppl_analysis,
+        "spectral": spectral,
+        "rmt": rmt,
+        "invariants": invariants,
+        "tier": auto.get("tier", "balanced"),
+        "_ppl_metrics": ppl_metrics,
+        "target_ratio": auto.get("target_pm_ratio"),
+        "guard_overhead": guard_overhead_section,
+        "primary_metric": report.get("metrics", {}).get("primary_metric")
         if isinstance(report.get("metrics"), dict)
         else None,
-        moe_section,
-        {
+        "moe": moe_section,
+        "dataset_capacity": {
             "tokens_available": capacity_tokens,
             "examples_available": capacity_examples,
         },
-    )
+    }
+    try:
+        if (
+            "pm_acceptance_range"
+            in inspect.signature(_compute_validation_flags).parameters
+        ):
+            validation_kwargs["pm_acceptance_range"] = pm_acceptance_range
+    except Exception:  # pragma: no cover - defensive against patched functions
+        validation_kwargs["pm_acceptance_range"] = pm_acceptance_range
+
+    validation_flags = _compute_validation_flags(**validation_kwargs)
     # Enforce validation key allow-list to prevent surface drift
     _allowed_validation = _load_validation_allowlist()
     validation_filtered = {
@@ -2537,6 +2550,103 @@ def _build_provenance_block(
     return provenance
 
 
+def _resolve_pm_acceptance_range_from_report(
+    report: dict[str, Any] | None,
+) -> dict[str, float]:
+    """Resolve primary-metric acceptance bounds from report context/meta/env."""
+
+    base_min = 0.95
+    base_max = 1.10
+
+    def _safe_float(val: Any) -> float | None:
+        try:
+            if val is None:
+                return None
+            return float(val)
+        except Exception:
+            return None
+
+    cfg_min = None
+    cfg_max = None
+    ctx = report.get("context") if isinstance(report, dict) else None
+    if isinstance(ctx, dict):
+        pm_ctx = (
+            ctx.get("primary_metric")
+            if isinstance(ctx.get("primary_metric"), dict)
+            else {}
+        )
+        if isinstance(pm_ctx, dict):
+            cfg_min = _safe_float(pm_ctx.get("acceptance_range", {}).get("min"))
+            cfg_max = _safe_float(pm_ctx.get("acceptance_range", {}).get("max"))
+        if cfg_min is None or cfg_max is None:
+            alt = ctx.get("pm_acceptance_range")
+            if isinstance(alt, dict):
+                cfg_min = (
+                    cfg_min if cfg_min is not None else _safe_float(alt.get("min"))
+                )
+                cfg_max = (
+                    cfg_max if cfg_max is not None else _safe_float(alt.get("max"))
+                )
+
+    if (cfg_min is None or cfg_max is None) and isinstance(report, dict):
+        meta = report.get("meta")
+        if isinstance(meta, dict):
+            meta_range = meta.get("pm_acceptance_range")
+            if isinstance(meta_range, dict):
+                cfg_min = (
+                    cfg_min
+                    if cfg_min is not None
+                    else _safe_float(meta_range.get("min"))
+                )
+                cfg_max = (
+                    cfg_max
+                    if cfg_max is not None
+                    else _safe_float(meta_range.get("max"))
+                )
+
+    def _parse_env(name: str) -> float | None:
+        try:
+            raw = os.environ.get(name, "")
+            if raw is None or str(raw).strip() == "":
+                return None
+            return float(raw)
+        except Exception:
+            return None
+
+    env_min = _parse_env("INVARLOCK_PM_ACCEPTANCE_MIN")
+    env_max = _parse_env("INVARLOCK_PM_ACCEPTANCE_MAX")
+
+    has_explicit = any(v is not None for v in (cfg_min, cfg_max, env_min, env_max))
+    if not has_explicit:
+        return {}
+
+    min_val = (
+        env_min if env_min is not None else cfg_min if cfg_min is not None else base_min
+    )
+    max_val = (
+        env_max if env_max is not None else cfg_max if cfg_max is not None else base_max
+    )
+
+    try:
+        if min_val is not None and min_val <= 0:
+            min_val = base_min
+    except Exception:
+        min_val = base_min
+    try:
+        if max_val is not None and max_val <= 0:
+            max_val = base_max
+    except Exception:
+        max_val = base_max
+
+    try:
+        if max_val is not None and min_val is not None and max_val < min_val:
+            max_val = min_val
+    except Exception:
+        max_val = base_max
+
+    return {"min": float(min_val), "max": float(max_val)}
+
+
 def _compute_validation_flags(
     ppl: dict[str, Any],
     spectral: dict[str, Any],
@@ -2549,6 +2659,7 @@ def _compute_validation_flags(
     primary_metric: dict[str, Any] | None = None,
     moe: dict[str, Any] | None = None,
     dataset_capacity: dict[str, Any] | None = None,
+    pm_acceptance_range: dict[str, float] | None = None,
 ) -> dict[str, bool]:
     """Compute validation flags for the certificate including canonical gates."""
     tier = (tier or "balanced").lower()
@@ -2569,7 +2680,25 @@ def _compute_validation_flags(
         "aggressive": 1.20,
         "none": 1.10,
     }
-    ratio_limit = tier_thresholds.get(tier, 1.10)
+    acceptance = pm_acceptance_range if isinstance(pm_acceptance_range, dict) else {}
+    ratio_min_bound = None
+    ratio_max_bound = None
+    try:
+        if acceptance.get("min") is not None:
+            ratio_min_bound = float(acceptance.get("min"))
+    except Exception:
+        ratio_min_bound = None
+    try:
+        if acceptance.get("max") is not None:
+            ratio_max_bound = float(acceptance.get("max"))
+    except Exception:
+        ratio_max_bound = None
+
+    ratio_limit = (
+        ratio_max_bound
+        if isinstance(ratio_max_bound, (int | float)) and math.isfinite(ratio_max_bound)
+        else tier_thresholds.get(tier, 1.10)
+    )
     if isinstance(target_ratio, int | float) and target_ratio > 0:
         ratio_limit = min(ratio_limit, float(target_ratio))
 
@@ -2636,9 +2765,18 @@ def _compute_validation_flags(
     tokens_ok_eff = tokens_ok or _tiny_relax
     # Apply hysteresis to ratio limit if needed
     ratio_limit_with_hyst = ratio_limit + max(0.0, hysteresis_ratio)
+    lower_bound_ok = True
+    if ratio_min_bound is not None and isinstance(ratio_vs_baseline, (int | float)):
+        try:
+            lower_bound_ok = math.isfinite(float(ratio_vs_baseline)) and (
+                float(ratio_vs_baseline) >= float(ratio_min_bound)
+            )
+        except Exception:
+            lower_bound_ok = True
     compression_acceptable = (
         isinstance(ratio_vs_baseline, int | float)
         and math.isfinite(ratio_vs_baseline)
+        and lower_bound_ok
         and ratio_vs_baseline <= ratio_limit_with_hyst
         and tokens_ok_eff
     )
@@ -2655,7 +2793,9 @@ def _compute_validation_flags(
         and all(isinstance(x, int | float) and math.isfinite(x) for x in ratio_ci)
     ):
         compression_acceptable = (
-            compression_acceptable and ratio_ci[1] <= ratio_limit_with_hyst
+            compression_acceptable
+            and ratio_ci[1] <= ratio_limit_with_hyst
+            and (ratio_min_bound is None or ratio_ci[0] >= ratio_min_bound)
         )
 
     # 3. RMT ε-rule compliance

{invarlock-0.3.0 → invarlock-0.3.1/src/invarlock.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: invarlock
-Version: 0.3.0
+Version: 0.3.1
 Summary: Edit‑agnostic robustness certificates for weight edits (InvarLock framework)
 Author-email: InvarLock Team <oss@invarlock.dev>
 Maintainer-email: InvarLock Maintainers <support@invarlock.dev>
@@ -112,7 +112,7 @@ they don’t, roll back safely.
 Technical: edit‑agnostic guard pipeline (invariants → spectral → RMT →
 variance) producing a machine‑readable Safety Certificate.
 
-> **Status:** 0.3.0 (pre‑1.0). Until 1.0, **minor** releases may be
+> **Status:** 0.3.1 (pre‑1.0). Until 1.0, **minor** releases may be
 > breaking. See CLI help and the CHANGELOG for updates.
 
 [![CI](https://img.shields.io/github/actions/workflow/status/invarlock/invarlock/ci.yml?branch=main&logo=github&label=CI)](https://github.com/invarlock/invarlock/actions/workflows/ci.yml)