invarlock 0.3.4__py3-none-any.whl → 0.3.6__py3-none-any.whl

invarlock/reporting/policy_utils.py

@@ -48,6 +48,10 @@ def _compute_thresholds_payload(
     if not isinstance(pm_policy, dict):
         pm_policy = {}
 
+    pm_tail_policy = metrics_policy.get("pm_tail", {})
+    if not isinstance(pm_tail_policy, dict):
+        pm_tail_policy = {}
+
     acc_policy = metrics_policy.get("accuracy", {})
     if not isinstance(acc_policy, dict):
         acc_policy = {}
@@ -76,6 +80,12 @@ def _compute_thresholds_payload(
         resolved_policy.get("variance", {}) if isinstance(resolved_policy, dict) else {}
     )
 
+    def _safe_float_any(value: Any, default: float) -> float:
+        try:
+            return float(value)
+        except Exception:
+            return float(default)
+
     payload = {
         "tier": tier_lc,
         "pm_ratio": {
@@ -86,6 +96,22 @@ def _compute_thresholds_payload(
             ),
             "hysteresis_ratio": float(pm_policy.get("hysteresis_ratio", 0.0) or 0.0),
         },
+        "pm_tail": {
+            "mode": str(pm_tail_policy.get("mode", "warn") or "warn").strip().lower(),
+            "min_windows": int(pm_tail_policy.get("min_windows", 0) or 0),
+            "quantile": _safe_float_any(pm_tail_policy.get("quantile", 0.95), 0.95),
+            "quantile_max": (
+                float(pm_tail_policy.get("quantile_max"))
+                if isinstance(pm_tail_policy.get("quantile_max"), int | float)
+                else None
+            ),
+            "epsilon": _safe_float_any(pm_tail_policy.get("epsilon", 0.0), 0.0),
+            "mass_max": (
+                float(pm_tail_policy.get("mass_max"))
+                if isinstance(pm_tail_policy.get("mass_max"), int | float)
+                else None
+            ),
+        },
         "accuracy": {
             "delta_min_pp": float(acc_policy.get("delta_min_pp", -1.0) or -1.0),
             "min_examples": int(acc_policy.get("min_examples", 200) or 200),
@@ -110,16 +136,6 @@ def _compute_thresholds_hash(payload: dict[str, Any]) -> str:
     return hashlib.sha256(canonical.encode("utf-8")).hexdigest()[:16]
 
 
-def _promote_legacy_multiple_testing_key(payload: dict[str, Any]) -> None:
-    """Promote legacy 'multipletesting' to 'multiple_testing' in-place if present."""
-    try:
-        legacy_mt = payload.pop("multipletesting", None)
-        if legacy_mt is not None and "multiple_testing" not in payload:
-            payload["multiple_testing"] = legacy_mt
-    except Exception:
-        pass
-
-
 def _resolve_policy_tier(report: RunReport) -> str:
     """Resolve the policy tier from report metadata or context."""
     tier: Any = None
@@ -218,15 +234,9 @@ def _build_resolved_policies(
     from .policy_utils import _format_family_caps as _ffc  # self import safe
 
     spectral_resolved["family_caps"] = _ffc(spectral_caps)
-    # Prefer observed policy sigma_quantile (accepting legacy aliases), then fallback
     pol_sq = None
     try:
         pol_sq = (spectral.get("policy", {}) or {}).get("sigma_quantile")
-        if pol_sq is None:
-            # Legacy aliases
-            pol_sq = (spectral.get("policy", {}) or {}).get("contraction") or (
-                spectral.get("policy", {}) or {}
-            ).get("kappa")
     except Exception:
         pol_sq = None
     spectral_resolved["sigma_quantile"] = _safe_float(
@@ -276,6 +286,9 @@ def _build_resolved_policies(
         spectral_resolved["max_spectral_norm"] = spectral.get("policy", {}).get(
             "max_spectral_norm", spectral_resolved.get("max_spectral_norm")
         )
+    mc = spectral.get("measurement_contract")
+    if isinstance(mc, dict) and mc:
+        spectral_resolved["measurement_contract"] = copy.deepcopy(mc)
     resolved["spectral"] = spectral_resolved
 
     # RMT guard
@@ -295,15 +308,16 @@ def _build_resolved_policies(
     rmt_resolved["epsilon_default"] = _safe_float(epsilon_default_val, 0.1)
     from .policy_utils import _format_epsilon_map as _fem
 
-    epsilon_map = _fem(rmt.get("epsilon_by_family") or rmt_resolved.pop("epsilon", {}))
+    epsilon_map = _fem(
+        rmt.get("epsilon_by_family") or rmt_resolved.get("epsilon_by_family") or {}
+    )
     if epsilon_map:
         rmt_resolved["epsilon_by_family"] = epsilon_map
-    else:
-        rmt_resolved.pop("epsilon", None)
-    if "epsilon" in rmt_resolved:
-        rmt_resolved.pop("epsilon", None)
     if "correct" in rmt_resolved:
         rmt_resolved["correct"] = bool(rmt_resolved["correct"])
+    mc = rmt.get("measurement_contract")
+    if isinstance(mc, dict) and mc:
+        rmt_resolved["measurement_contract"] = copy.deepcopy(mc)
     resolved["rmt"] = rmt_resolved
 
     # Variance guard
@@ -441,13 +455,9 @@ def _extract_effective_policies(report: RunReport) -> dict[str, Any]:
             elif guard_name == "spectral":
                 sigma_quantile = guard_metrics.get(
                     "sigma_quantile",
-                    guard_metrics.get("contraction", guard_metrics.get("kappa", 0.95)),
-                )
-                multiple_testing = guard_metrics.get("multiple_testing") or (
-                    guard_metrics.get("multipletesting")
-                    if isinstance(guard_metrics.get("multipletesting"), dict)
-                    else None
+                    0.95,
                 )
+                multiple_testing = guard_metrics.get("multiple_testing")
                 guard_policy = {
                     "max_spectral_norm": guard_metrics.get("max_spectral_norm"),
                     "stability_score": guard_metrics.get("stability_score", 0.95),
@@ -473,20 +483,13 @@ def _extract_effective_policies(report: RunReport) -> dict[str, Any]:
 
         if guard_policy:
             if guard_name == "spectral":
-                sigma_quantile = guard_policy.get("sigma_quantile")
-                if sigma_quantile is None:
-                    sigma_quantile = guard_policy.get("contraction")
-                if sigma_quantile is None and "kappa" in guard_policy:
-                    sigma_quantile = guard_policy["kappa"]
                 sanitized_policy = dict(guard_policy)
+                sigma_quantile = sanitized_policy.get("sigma_quantile")
                 if sigma_quantile is not None:
                     try:
                         sanitized_policy["sigma_quantile"] = float(sigma_quantile)
                     except (TypeError, ValueError):
                         pass
-                _promote_legacy_multiple_testing_key(sanitized_policy)
-                sanitized_policy.pop("contraction", None)
-                sanitized_policy.pop("kappa", None)
                 if sanitized_policy.get("max_spectral_norm") in (None, 0):
                     sanitized_policy["max_spectral_norm"] = None
                 guard_policy = sanitized_policy
@@ -587,7 +590,6 @@ __all__ = [
     "_compute_variance_policy_digest",
     "_compute_thresholds_payload",
     "_compute_thresholds_hash",
-    "_promote_legacy_multiple_testing_key",
     "_resolve_policy_tier",
     "_build_resolved_policies",
     "_extract_effective_policies",

invarlock/reporting/primary_metric_utils.py

@@ -30,6 +30,38 @@ def attach_primary_metric(
         pm = m.get("primary_metric") if isinstance(m, dict) else None
         if isinstance(pm, dict) and pm:
             pm_copy = copy.deepcopy(pm)
+            pm_copy.setdefault("invalid", bool(pm_copy.get("invalid", False)))
+            degraded_reason = pm_copy.get("degraded_reason")
+            preview_val = pm_copy.get("preview")
+            final_val = pm_copy.get("final")
+            ratio_val = pm_copy.get("ratio_vs_baseline")
+            baseline_final = (
+                baseline_ref.get("primary_metric", {}).get("final")
+                if isinstance(baseline_ref, dict)
+                else None
+            )
+
+            def _is_finite(value: Any) -> bool:
+                return isinstance(value, (int, float)) and math.isfinite(float(value))
+
+            baseline_has_reference = _is_finite(baseline_final)
+            needs_pm_fallback = not (_is_finite(preview_val) and _is_finite(final_val))
+            needs_ratio_fallback = baseline_has_reference and not _is_finite(ratio_val)
+
+            if degraded_reason is None:
+                if needs_pm_fallback:
+                    degraded_reason = "non_finite_pm"
+                elif needs_ratio_fallback:
+                    degraded_reason = "non_finite_delta"
+                elif pm_copy.get("invalid"):
+                    degraded_reason = "primary_metric_invalid"
+
+            pm_copy["degraded"] = bool(
+                pm_copy.get("degraded") or pm_copy.get("invalid") or degraded_reason
+            )
+            if pm_copy["degraded"] and degraded_reason:
+                pm_copy.setdefault("degraded_reason", degraded_reason)
+
             # Propagate instability hint from ppl_analysis
             try:
                 if isinstance(ppl_analysis, dict) and bool(
@@ -75,33 +107,52 @@ def attach_primary_metric(
                             pm_copy["analysis_point_final"] = float(mean_fin)
                     # Attach analysis-basis CIs for preview/final in log space from report metrics
                     try:
-                        dlci = (
-                            _coerce_interval(m.get("logloss_delta_ci"))
-                            if isinstance(m, dict)
-                            else (math.nan, math.nan)
-                        )
-                        if isinstance(dlci, tuple | list) and len(dlci) == 2:
-                            lo, hi = float(dlci[0]), float(dlci[1])
-                            if math.isfinite(lo) and math.isfinite(hi):
-                                pm_copy.setdefault("ci", (lo, hi))
+                        dlci_source: tuple[float, float] | list[float] | None = None
+                        pairing_source = None
+                        if isinstance(ppl_analysis, dict):
+                            stats = ppl_analysis.get("stats") or {}
+                            if isinstance(stats, dict):
+                                pairing_source = stats.get("pairing")
+                            if pairing_source == "paired_baseline":
+                                dlci_source = _coerce_interval(
+                                    ppl_analysis.get("logloss_delta_ci")
+                                )
+                        if dlci_source is None:
+                            dlci_source = (
+                                _coerce_interval(m.get("logloss_delta_ci"))
+                                if isinstance(m, dict)
+                                else (math.nan, math.nan)
+                            )
+                        if (
+                            isinstance(dlci_source, tuple | list)
+                            and len(dlci_source) == 2
+                        ):
+                            lo_raw, hi_raw = dlci_source[0], dlci_source[1]
+                            if isinstance(lo_raw, (int, float)) and isinstance(
+                                hi_raw, (int, float)
+                            ):
+                                lo, hi = float(lo_raw), float(hi_raw)
+                                if math.isfinite(lo) and math.isfinite(hi):
+                                    pm_copy.setdefault("ci", (lo, hi))
                     except Exception:
                         pass
                 except Exception:
                     pass
             # Ensure ratio_vs_baseline present and consistent
             try:
-                base_final = (
-                    baseline_ref.get("primary_metric", {}).get("final")
-                    if isinstance(baseline_ref, dict)
+                fin = pm_copy.get("final")
+                baseline_final_val = (
+                    float(baseline_final)
+                    if isinstance(baseline_final, (int, float))
+                    and _is_finite(baseline_final)
                     else None
                 )
-                fin = pm_copy.get("final")
                 if (
-                    isinstance(fin, int | float)
-                    and isinstance(base_final, int | float)
-                    and float(base_final) > 0
+                    isinstance(fin, (int, float))
+                    and baseline_final_val is not None
+                    and baseline_final_val > 0
                 ):
-                    pm_copy["ratio_vs_baseline"] = float(fin) / float(base_final)
+                    pm_copy["ratio_vs_baseline"] = float(fin) / baseline_final_val
                 # Ensure display_ci aligns with log-space CI for ppl-like metrics
                 try:
                     kind = str(pm_copy.get("kind", "")).lower()
@@ -277,6 +328,9 @@ def attach_primary_metric(
                 if isinstance(point, float):
                     pm["display_ci"] = [point, point]
                 else:
+                    # As last resort, emit a degenerate [1.0, 1.0] to satisfy schema invariants
                     pm["display_ci"] = [1.0, 1.0]
+                    pm.setdefault("estimated", True)
+
     except Exception:
         pass

invarlock/reporting/render.py

@@ -468,9 +468,70 @@ def render_certificate_markdown(certificate: dict[str, Any]) -> str:
             f"| Guard Overhead Acceptable | {status} | {measured} | ≤ +{threshold_pct:.1f}% | point | Guarded vs bare PM overhead |"
         )
 
+    def _emit_pm_tail_gate_row() -> None:
+        pm_tail = certificate.get("primary_metric_tail", {}) or {}
+        if not isinstance(pm_tail, dict) or not pm_tail:
+            return
+
+        evaluated = bool(pm_tail.get("evaluated", False))
+        mode = str(pm_tail.get("mode", "warn") or "warn").strip().lower()
+        passed = bool(pm_tail.get("passed", True))
+        warned = bool(pm_tail.get("warned", False))
+
+        if not evaluated:
+            status = "🛈 INFO"
+        elif passed:
+            status = "✅ PASS"
+        elif mode == "fail":
+            status = "❌ FAIL"
+        else:
+            status = "⚠️ WARN" if warned else "⚠️ WARN"
+
+        policy = (
+            pm_tail.get("policy", {}) if isinstance(pm_tail.get("policy"), dict) else {}
+        )
+        stats = (
+            pm_tail.get("stats", {}) if isinstance(pm_tail.get("stats"), dict) else {}
+        )
+
+        q = policy.get("quantile", 0.95)
+        try:
+            qf = float(q)
+        except Exception:
+            qf = 0.95
+        qf = max(0.0, min(1.0, qf))
+        q_key = f"q{int(round(100.0 * qf))}"
+        q_name = f"P{int(round(100.0 * qf))}"
+        q_val = stats.get(q_key)
+        mass_val = stats.get("tail_mass")
+        eps = policy.get("epsilon", stats.get("epsilon"))
+
+        measured_parts: list[str] = []
+        if isinstance(q_val, int | float) and math.isfinite(float(q_val)):
+            measured_parts.append(f"{q_name}={float(q_val):.3f}")
+        if isinstance(mass_val, int | float) and math.isfinite(float(mass_val)):
+            measured_parts.append(f"mass={float(mass_val):.3f}")
+        measured = ", ".join(measured_parts) if measured_parts else "N/A"
+
+        thr_parts: list[str] = []
+        qmax = policy.get("quantile_max")
+        if isinstance(qmax, int | float) and math.isfinite(float(qmax)):
+            thr_parts.append(f"{q_name}≤{float(qmax):.3f}")
+        mmax = policy.get("mass_max")
+        if isinstance(mmax, int | float) and math.isfinite(float(mmax)):
+            thr_parts.append(f"mass≤{float(mmax):.3f}")
+        if isinstance(eps, int | float) and math.isfinite(float(eps)):
+            thr_parts.append(f"ε={float(eps):.1e}")
+        threshold = "; ".join(thr_parts) if thr_parts else "policy"
+
+        lines.append(
+            f"| Primary Metric Tail | {status} | {measured} | {threshold} | {q_name.lower()} | Tail regression vs baseline (ΔlogNLL) |"
+        )
+
     # Emit canonical gate rows
     if has_pm:
         _emit_pm_gate_row()
+        _emit_pm_tail_gate_row()
         _emit_drift_gate_row()
         _emit_overhead_gate_row()
 

invarlock/reporting/report.py

@@ -401,7 +401,7 @@ def _generate_single_markdown(report: RunReport) -> list[str]:
             parts.append(f"ratio_vs_baseline={ratio:.3f}")
         lines.append(" — ".join(parts) if len(parts) > 1 else parts[0])
     else:
-        # When primary_metric is absent, do not attempt legacy fallbacks
+        # When primary_metric is absent, do not attempt fallbacks
         lines.append("- **Primary Metric**: unavailable")
     lines.append(
         f"- **Parameters Changed**: {report['edit']['deltas']['params_changed']:,}"

invarlock/reporting/report_types.py

@@ -101,8 +101,10 @@ class EvalMetrics(TypedDict, total=False):
 
     # Canonical primary metric snapshot
     primary_metric: dict[str, Any]
+    # Always-computed tail evidence/gate for ppl-like primary metrics
+    primary_metric_tail: dict[str, Any]
 
-    # Optional legacy/aux fields retained for guard telemetry and debug
+    # Optional aux fields retained for guard telemetry and debug
     latency_ms_per_tok: float  # Average latency per token in milliseconds
     memory_mb_peak: float  # Peak memory usage in MB
     spectral: dict[str, Any]  # Spectral norm summaries
@@ -164,6 +166,7 @@ class RunReport(TypedDict):
     # Optional extras kept for richer downstream processing
     guard_overhead: NotRequired[dict[str, Any]]
     provenance: NotRequired[dict[str, Any]]
+    context: NotRequired[dict[str, Any]]
 
 
 # Utility functions for creating reports
@@ -247,7 +250,7 @@ def validate_report(report: RunReport) -> bool:
             if pm_final is not None and not isinstance(pm_final, int | float):
                 return False
         else:
-            # PM-only: legacy ppl_* acceptance removed
+            # PM-only: ppl_* acceptance removed
             return False
 
         meta = report.get("meta", {})

invarlock/reporting/validate.py

@@ -9,10 +9,9 @@ Supports both automated CI testing and flexible user validation.
 from __future__ import annotations
 
 import json
-import os
 import warnings
 from pathlib import Path
-from typing import Any, cast
+from typing import Any
 
 __all__ = [
     "validate_against_baseline",
@@ -99,23 +98,7 @@ def validate_against_baseline(
     ratio_bounds: tuple[float, float] = (1.25, 1.32),
     delta_bounds_pp: tuple[float, float] | None = None,
     structural_exact: bool = True,
-    **kwargs,
 ) -> ValidationResult:
-    # Backward-compatible kwargs (deprecated): enable via INVARLOCK_VALIDATE_LEGACY=1
-    legacy = str(os.environ.get("INVARLOCK_VALIDATE_LEGACY", "")).strip().lower() in {
-        "1",
-        "true",
-        "yes",
-        "on",
-    }
-    if legacy:
-        if "tol_ppl_ratio" in kwargs and isinstance(
-            kwargs["tol_ppl_ratio"], int | float
-        ):
-            tol_ratio = float(kwargs["tol_ppl_ratio"])
-        if "ppl_bounds" in kwargs and isinstance(kwargs["ppl_bounds"], tuple):
-            # Coerce after runtime guard
-            ratio_bounds = cast(tuple[float, float], kwargs["ppl_bounds"])
     """
     Validate pruning results against baseline metrics (PM-only API).
 

{invarlock-0.3.4.dist-info → invarlock-0.3.6.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: invarlock
-Version: 0.3.4
+Version: 0.3.6
 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:** 0.3.4 (pre‑1.0). Until 1.0, **minor** releases may be
+> **Status:** 0.3.6 (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)
@@ -170,7 +170,7 @@ Quick examples (repo presets, CPU; repo clone required for preset paths):
 pip install "invarlock[hf]"
 
 # Preflight a config (JSON diagnostics)
-invarlock doctor --config configs/tasks/causal_lm/ci_cpu.yaml --json
+invarlock doctor --config configs/presets/causal_lm/wikitext2_512.yaml --json
 
 # Calibrated GPT‑2 small (recommended starting point; repo preset)
 INVARLOCK_ALLOW_NETWORK=1 INVARLOCK_DEDUP_TEXTS=1 \
@@ -179,7 +179,7 @@ invarlock certify \
   --subject  gpt2 \
   --adapter auto \
   --profile release \
-  --preset configs/tasks/causal_lm/release_auto.yaml
+  --preset configs/presets/causal_lm/wikitext2_512.yaml
 
 # Tiny causal LM smoke (out‑of‑calibration, dev‑only)
 INVARLOCK_ALLOW_NETWORK=1 \
@@ -249,7 +249,7 @@ INVARLOCK_ALLOW_NETWORK=1 invarlock certify \
   --subject  gpt2 \
   --adapter auto \
   --profile ci \
-  --preset configs/tasks/causal_lm/ci_cpu.yaml
+  --preset configs/presets/causal_lm/wikitext2_512.yaml
 ```
 
 - Offline/air‑gapped usage: pre‑download to a cache, then run with network
@@ -488,7 +488,7 @@ output:
 Run preflight checks before a run to catch misconfigurations early:
 
 ```bash
-invarlock doctor --config configs/tasks/causal_lm/ci_cpu.yaml --json
+invarlock doctor --config configs/presets/causal_lm/wikitext2_512.yaml --json
 ```
 
 Text mode emits lines prefixed with `ERROR:`, `WARNING:`, or `NOTE:` and stable

{invarlock-0.3.4.dist-info → invarlock-0.3.6.dist-info}/RECORD

@@ -1,15 +1,15 @@
-invarlock/__init__.py,sha256=pTrxNy6MaRsu31Wvc1_V2TIENrSBVyVlk3AFDqggHxg,1268
+invarlock/__init__.py,sha256=_KZRr1cTUSCHY7ZiPIk71h06tEznF_LOZQbyXcQfBMk,1268
 invarlock/__main__.py,sha256=ffhoKctw89j-henmQXThbHDIdlvK9fBfsy8LpjhOEXc,146
 invarlock/config.py,sha256=7BUOl7EW258YnsgRipjOx6lmWou5jNDzimREd35ewsQ,1725
-invarlock/model_profile.py,sha256=sFHpK-1Q-1DjiZTWMolQBG4Dw9feLJgk3GnaD3ixgd8,12809
+invarlock/model_profile.py,sha256=K-RSLbQRv9uM4RMfLvApLf3p2wp8iKrfy3BZ4p1h3m4,13756
 invarlock/model_utils.py,sha256=mEl9KOF2yOJgx-6r38PbD9SNFeAScamy9K-csxtl40Q,6395
 invarlock/py.typed,sha256=LCPmZeE_vANVVJDNvuq9A07i7jg9Nxrq6f10UeuNfZc,37
 invarlock/security.py,sha256=xUjbHu-bHKUfFqDyN_21Hj00NBmAcGgnAg4JCjb0JOE,4861
 invarlock/sparsity_utils.py,sha256=30SC3osptca2GmzxezbfX31EE6sRUhmEz8u3jn4vB2c,8281
-invarlock/_data/runtime/tiers.yaml,sha256=xVHPi-Nopxc-Ye-5lGIer9k-lxiiZZ2cmSnzZE-EHIA,2989
+invarlock/_data/runtime/tiers.yaml,sha256=9EbVnNw6MX18NKMI9xjKzJt-BsaB-JhUtC13c3Skmls,3619
 invarlock/_data/runtime/profiles/ci_cpu.yaml,sha256=FE5vxeemAFWW7oRxPvp_8hIcCO8qrJilvQpXuS3f6u0,389
 invarlock/_data/runtime/profiles/release.yaml,sha256=xF0Qb0OTm904U6L3wK674JMTcDPegYvpKgwUB9pfq_w,482
-invarlock/adapters/__init__.py,sha256=Bwj8aKjhFxCzvcdxTIl-nG7IXyIE4L3Nd_fsIghbZxA,3418
+invarlock/adapters/__init__.py,sha256=MxWJGRxVrV2TL5c7MA4FqmWyu1e_Kwpk7Ol4IFUJcNM,3411
 invarlock/adapters/_capabilities.py,sha256=FmzUR5BHsxWe92Z9W1As-G5_5wG1PvqF2sUpjZ2_CdY,1483
 invarlock/adapters/auto.py,sha256=4F3I63PRW604v_NtaTY7aGvUJVjGSjejVKFff-X_1t4,7693
 invarlock/adapters/base.py,sha256=szSh1bECeDSDGQSr5oIWhs5RlI587gE4gzdt5cnOJ1s,16100
@@ -25,15 +25,15 @@ invarlock/adapters/hf_t5.py,sha256=sqHmi8io7oKlr7GrAk8tGo72zb-hmHGmXplCQSZgrA8,4
 invarlock/adapters/py.typed,sha256=LCPmZeE_vANVVJDNvuq9A07i7jg9Nxrq6f10UeuNfZc,37
 invarlock/assurance/__init__.py,sha256=SFDT2klaUaKZejulL5cfBTW29ehJxyW5srE-LuqP7z0,1388
 invarlock/calibration/__init__.py,sha256=M5lbkNQtLBuQjQJGeRzHovkpYI87fEWIm7a0b23jSp4,110
-invarlock/calibration/spectral_null.py,sha256=968168sQDgtZ2tATx7JkEtwtWw0Y7RFSk7OyQxwXBsE,9419
-invarlock/calibration/variance_ve.py,sha256=sdO-a9FUSDSvOI9T8qXAanw4GWOV7RZ_vv53N5ilEg4,4549
+invarlock/calibration/spectral_null.py,sha256=2WQU7YgnnJ2i3OF60Zj4miaWBvWBXQmztukQtIWhvro,9601
+invarlock/calibration/variance_ve.py,sha256=QkHO2vxht-n56LNqWbHYYY106Ge5PbxzHfGV1QktnN8,4487
 invarlock/cli/__init__.py,sha256=mRjYDVmNykhxJxfQBxJMzhFfmNhmvP6OWI-vJL-ehUA,196
 invarlock/cli/__main__.py,sha256=spgMrx0eJQSEf4rcd0_mB2B8rHOc3iWlTFO3P_qEvoM,175
 invarlock/cli/_evidence.py,sha256=38QMfFlPUT_qIRAZmpDROkwSIpHGiNKRWDyXfZ9xI_U,769
 invarlock/cli/_json.py,sha256=AKrRNFza76tH3GBVHO8gSA3klUVQ9XpDo_aYh8l-7j8,2367
 invarlock/cli/adapter_auto.py,sha256=Qwh4hJXm_vGexCXPt3iP70rQPC7ozwlEzSTYay0svQ8,5284
 invarlock/cli/app.py,sha256=iWO4tTwglZTH2CGEW_SXIQ95MyFTw6NYxlBujhpZQRg,9179
-invarlock/cli/config.py,sha256=d9uar3l9VuqeF_AMtan85V71Vc2L6HVjwLwE6SM2e0M,13464
+invarlock/cli/config.py,sha256=vyLxR9WKW2vnIFACmS6k79yHEaO1zKuZ9q5bufaoB18,14203
 invarlock/cli/constants.py,sha256=RuXxG82pukfBnEuJB2CVF5dyu4lEi-l8eAxkR314yuk,2107
 invarlock/cli/determinism.py,sha256=ux3AODjZbSdYE29uSrHXvY2qrEPxeAJd04QLT52lnec,8346
 invarlock/cli/device.py,sha256=5X0j2yiZbSKX9-SlDaAbSeHCoWhfEQ74YWPTb8WRk8k,3165
@@ -43,28 +43,28 @@ invarlock/cli/overhead_utils.py,sha256=Ygvl192UDTroAbVAd7v35jgigmmfVxNkCIGYlWfh6
 invarlock/cli/provenance.py,sha256=2E_8MxpSVEcxe5UewTb5Hlz6x4jXFwGd1567F4nIbEc,1939
 invarlock/cli/utils.py,sha256=R6IN21lGfko2lGUgspmkZyjA6Nl_dVRaP8F8nRtN5yw,1393
 invarlock/cli/commands/__init__.py,sha256=afpKpU8hJI41Ol46oXxVwQgPwFBy4OmLaa_v4AAOdts,712
-invarlock/cli/commands/calibrate.py,sha256=b_75wIuHuhQ-frVcKkSuA8W5AMN3acc63k6q9tZvNEE,20780
-invarlock/cli/commands/certify.py,sha256=2dYMrZYX1fACOIHbHshPoM7leCizQDRTPZ3Gc8s7FNo,14903
-invarlock/cli/commands/doctor.py,sha256=cz4uH8v77M74YhwlWEMgnmEpmqbsV5CN5-zkuDIBLS8,56897
-invarlock/cli/commands/explain_gates.py,sha256=CajKcu6rNi57yBsM_EUhi8rlmTl-ymT1kUaznXzALSE,7624
+invarlock/cli/commands/calibrate.py,sha256=TnoYKjvmz2pSj3Wak8HluSV27hN2D8O1dGAQWCcHeeE,20830
+invarlock/cli/commands/certify.py,sha256=n62diGs7eFymuJNz-EJmWldYVO5KehHWeYtwtL1UWRY,15861
+invarlock/cli/commands/doctor.py,sha256=IQciP1Xs-_W4R5lhZv1ZbYwvIjy9VY_WMBypvbok1FI,56944
+invarlock/cli/commands/explain_gates.py,sha256=7cnc_FBxvSkjmM12JOA2yfhY4JH1puDDaIHo3r6IC9M,9444
 invarlock/cli/commands/export_html.py,sha256=oUGtt6EpKhjlywdtz_0FqYqcmA219H4eSjPuSMDShgY,3095
 invarlock/cli/commands/plugins.py,sha256=u5E6ThcWE5RUs2YKH-m6x7n-5K1jt2IudkF9kl9T5xg,53486
-invarlock/cli/commands/report.py,sha256=I-B11k6cLLNUIaNiKlsWrGgZ2cTQQJsarN3oNf9Dt84,12988
-invarlock/cli/commands/run.py,sha256=nf9Y_BcgxQyaCjwQmGUt8ZojDXBDPSeEk4Ut4_uRggM,200772
-invarlock/cli/commands/verify.py,sha256=WAgIzV_CAZSkzRJIxVFV5llZAL3n3-biTMXLm6YL_qs,46408
+invarlock/cli/commands/report.py,sha256=AaGq4J1qm8ERMZmbAkAHtc8kZ0CvMp14wili1K2tX8g,12987
+invarlock/cli/commands/run.py,sha256=S25dVLxQe1RVxLQ6cD8CSlpqGxp1Eggvbk4WzsHSwW4,204782
+invarlock/cli/commands/verify.py,sha256=zKoSTPLrMgCw6LseCNn5RkPDEZTuRZ8j14tWHbvcZxM,49234
 invarlock/core/__init__.py,sha256=4wb83Xv7NE5M1FgvaFUviiNtVSTVATiPH3gqavNmp2w,1490
 invarlock/core/abi.py,sha256=gmU1F7LDd2HTcF4kcznn8vgG4zj9UCHSqhBk9jyu05k,484
-invarlock/core/api.py,sha256=dg76wZaULo-T-_nZtq6rmDtWwSNLCSYXiQKVYySV6Fg,7545
-invarlock/core/auto_tuning.py,sha256=ulk_bLAokQJFHgttj5x9QmkwSx7IPVhX2J1GBPWS6PA,17383
+invarlock/core/api.py,sha256=RTFswFBuUH38gary5X7qmccOD7vAYLtSN0ie8Y6p1ck,8480
+invarlock/core/auto_tuning.py,sha256=Gj0AhpThoDYnzr9DyuXM9iCuieMKR4xv43tC7ZcjHzs,18543
 invarlock/core/bootstrap.py,sha256=0mA_x1Ipz4eTwJqoz4-F4H0O2AqvlmQGA9BVJpTNqXs,10655
 invarlock/core/checkpoint.py,sha256=a78L-mZd3j2kC1az2eTScRxTHjrULuYs-UPfWUcDgqM,6933
-invarlock/core/contracts.py,sha256=9j55WVwMrEsUqxWlzAdMsHtkzgfkSftdizhcLiJBauw,2165
+invarlock/core/contracts.py,sha256=ESK3-ErieGEf8e84wiphlCjahdr87zoKyTSskq6o23E,2468
 invarlock/core/error_utils.py,sha256=T23-p5ONQ-SeVuMR4Ts0cWyupsSa-0DAgsejRTfxeCg,1782
 invarlock/core/events.py,sha256=8XBAi-9A7ys7QJQwqlz8PVlfxF0TM_TvLqjcPtDwZm4,9428
 invarlock/core/exceptions.py,sha256=b4OszJ0Fd0Ezy8s99AzprS7lAkqdZYGXaSj9fYaln4E,2077
-invarlock/core/registry.py,sha256=a5hYHXl74MLnaPFbmpwJ-1OsJgY3SSe2Jisbzfr2rLM,19137
+invarlock/core/registry.py,sha256=ZRW7JY-XgWrQMtN451n_Uk4it3CMLCsZUPHqic2G1x4,19109
 invarlock/core/retry.py,sha256=KTVkrTnWs60jwATOZDHinERH56GnOGjsKR0lmohagEo,4503
-invarlock/core/runner.py,sha256=Tep0CDT75PbE8bpGapR8J00-MXZ-T0JB4C6tsqOMs68,93567
+invarlock/core/runner.py,sha256=BYvJlWhsrSRyudXNdY-vuZVt5NypYqMTS5yK28jhIVI,104105
 invarlock/core/types.py,sha256=nVLMP4yqlxwhE1moQU7FWVeGJqTuud-cvTZiutdBGKk,3585
 invarlock/edits/__init__.py,sha256=5CJTX6oJTLj79Vmks2_8mXRGFQH0oyKaZOrBLTlRtV0,277
 invarlock/edits/_edit_utils.py,sha256=VGl32JAt7XsovxJbX1YvW9DJvtG-K2EKu-TjBFTQmDY,6329
@@ -74,13 +74,14 @@ invarlock/edits/py.typed,sha256=LCPmZeE_vANVVJDNvuq9A07i7jg9Nxrq6f10UeuNfZc,37
 invarlock/edits/quant_rtn.py,sha256=8rseGbF_-EFuL8MWCuzgGXYfzp9LsONhV72-OK56mlk,30684
 invarlock/edits/registry.py,sha256=MmUfJhhvc3WxB03wQkPxFMS6nkT7YcXFPQqhbksfOUc,4523
 invarlock/eval/__init__.py,sha256=GsKoszCysh3TT_UHiuJcqeoiXT7AUNZoqMuBmefnWtY,755
-invarlock/eval/bench.py,sha256=85tGywzAeJSVlZ1Qsebk8FwIGUYhq9g6TxOW5XnL5X8,56335
+invarlock/eval/bench.py,sha256=pG1OSO5-Ela46ReImhQJAuLTxE6oSfTFjHSTSL_09-8,55925
 invarlock/eval/bench_regression.py,sha256=GDphmdcH7NI8P_afSkQAPMChEgW15J0gqTZ4Kj3kxgw,456
 invarlock/eval/bootstrap.py,sha256=CNn7W_MY716Wa_wn6cYRmbTYB9lT2X0yLOAIHCRw29k,1766
-invarlock/eval/data.py,sha256=QVGOIiakm89jEy9-rVMlFJ6WYHq_keSrSNUOjRsVwCA,75736
-invarlock/eval/metrics.py,sha256=3ay3mqsFTa0V5a66gKdk7BOOJjKWXlub5lrTpcex9DQ,73946
-invarlock/eval/primary_metric.py,sha256=ay8gB7xGjWahXpJP-o4OcZFNRNZKcUIHjKmSeOUWJNk,28701
+invarlock/eval/data.py,sha256=YUOvgM3h_KSIZ0FQw5yhrsU_YonIb5Ml950t8hg3qn8,66652
+invarlock/eval/metrics.py,sha256=Ff72jgeLAWCn1WgpDE2lX0RzdJMrb39fgjKCF2rwkpk,78972
+invarlock/eval/primary_metric.py,sha256=lGDWiXM1cyXqGH6XprLrxBl1hTwjrlSYbQCzF687zsA,29545
 invarlock/eval/py.typed,sha256=LCPmZeE_vANVVJDNvuq9A07i7jg9Nxrq6f10UeuNfZc,37
+invarlock/eval/tail_stats.py,sha256=ueScFzPvBnBC4jVW5EElnB0c3LPK3cAITcOUdyPCsEk,7114
 invarlock/eval/probes/__init__.py,sha256=KocOJmEWrD90c2HbDW3wDsdsibRyidskpI4rNNfLbZ0,593
 invarlock/eval/probes/fft.py,sha256=iuZauzOaNPbXN4t74vMiNQAjg6ZmwPNPUuUBj4L1ick,4106
 invarlock/eval/probes/mi.py,sha256=ew33dkc1tymDVek718m7ZU--BAFa1L35PZA2POwKx0g,7258
@@ -91,15 +92,16 @@ invarlock/eval/providers/text_lm.py,sha256=EpQWjQEkxSMhTiABsJzZqfYsgSYeMmUW-ZgeP
 invarlock/eval/providers/vision_text.py,sha256=pqhPHwvYEnyKupchIy-g-jM_oWeYNXRHb23_Qm4v2hU,3180
 invarlock/guards/__init__.py,sha256=Kj7Jwb4qLyaY0ATRku34T628vom3QMjTiVrlpfJAWw8,443
 invarlock/guards/_contracts.py,sha256=WPhg3iWpHGkKfPvlPPQ-umzOakpy3s289DKkYQ0e64I,263
+invarlock/guards/_estimators.py,sha256=84nT5eF7uN-zr-RfYF2uOvdHlmE5c8_yvor4IqxvwWk,4642
 invarlock/guards/invariants.py,sha256=2mD5WSJMCksjO-DodAac5pvZN0nl29NKHhqwqH_gyCQ,22379
-invarlock/guards/policies.py,sha256=sbZkddUSnkx3ZsFm4vV4DfSU81QUh9R8S4vuAtPF11I,25466
+invarlock/guards/policies.py,sha256=ggY0zZ3i39OAeFZJSG1FjwnR3-wjWntKLLvumBvG8pc,25933
 invarlock/guards/py.typed,sha256=LCPmZeE_vANVVJDNvuq9A07i7jg9Nxrq6f10UeuNfZc,37
-invarlock/guards/rmt.py,sha256=U3_JTZKsOIgnH-TdZxKMLL9p_3EbfYJKlIqB1gtvQLw,97928
-invarlock/guards/spectral.py,sha256=lmnq4x1-MwoXAK6awaUeb_6xvCya-Tg_zqECDdYvMOA,58618
-invarlock/guards/tier_config.py,sha256=_WJIQ4qvIOc8EI1ygBhpBqbZPt072RT6H6Oh9-LqNWY,10706
-invarlock/guards/variance.py,sha256=lcTNb_uSufB7D8cLkoY1S_IQWHUjIfCucsu5y8MpfPs,136794
+invarlock/guards/rmt.py,sha256=eKWbF4htt_0-Cuby5T9Q6Di6QmQPsCBUWh-dCXlrxic,99864
+invarlock/guards/spectral.py,sha256=TP_xmzL0iP3CSQjnkulNbuJegxCrp68MuMPWdLeE0Gg,69597
+invarlock/guards/tier_config.py,sha256=6JzS1TOlhUVX2NdUodWhb4VOA_AnAFVOXvwY9aiopbA,10796
+invarlock/guards/variance.py,sha256=8WauGx34ORBMrArYTDRtyEZuTcZSeBIqJHqG067wugc,136371
 invarlock/guards_ref/__init__.py,sha256=jLnyFqdqQaheG1qQMlU4Gx7R118rkkQHPqFVF3_1ih0,366
-invarlock/guards_ref/rmt_ref.py,sha256=md-aSzLCxPL3OXmrA5NtI9wK7cVSyd2xw8WtSodcGQY,1246
+invarlock/guards_ref/rmt_ref.py,sha256=IfwN-l9cLExUQJhZmr-1uJv7qRSoYrGg2w2caCG53hs,1399
 invarlock/guards_ref/spectral_ref.py,sha256=FdwFfrs5hxEEUIfBV3CvAJvTX78gAM00mKLEXyZ0zJo,4386
 invarlock/guards_ref/variance_ref.py,sha256=BmTQL5nfxReJK5SqnuUi1SOcQftK7IUNwlcxqdtGPEU,2264
 invarlock/observability/__init__.py,sha256=_fCH1H51M0PYOvpuNq9sVQLTiDAtdQMvtnhVt-z-VPk,345
@@ -117,24 +119,24 @@ invarlock/plugins/hf_bnb_adapter.py,sha256=g0ysWEi8dQzLtJy8iCszfTsYCOACuZMFYnTLM
 invarlock/plugins/hf_gptq_adapter.py,sha256=ysugAcnjLqF5sqpijSNiim1xUpRmoIgBrG053X3S2hE,3743
 invarlock/plugins/py.typed,sha256=LCPmZeE_vANVVJDNvuq9A07i7jg9Nxrq6f10UeuNfZc,37
 invarlock/reporting/__init__.py,sha256=A0541EqxcdTpslNbZEWIO4q-LCqzCQcadev2IBKEBbM,232
-invarlock/reporting/certificate.py,sha256=yoMfALyXYGK6BBHJfsrCFDxb07o-9J0KQqui35RHkhM,146382
-invarlock/reporting/certificate_schema.py,sha256=ft58LrF4lfdXT08P0jjtiNvRPVs6TdgPU_cp9K4TMPw,8939
-invarlock/reporting/dataset_hashing.py,sha256=b9LM2rtOtD0-1gQq_oJ0yI6oLleempwKinQtLupajwI,8680
-invarlock/reporting/guards_analysis.py,sha256=7TTlF-OwgT7wGQ8zZ8oHHPuMdUMvKw7836WvO0J1GUc,42465
+invarlock/reporting/certificate.py,sha256=UyUhhrgWsdNG2xPtOVvRjaiWaOhHP-Aw6PuzOD8UBHQ,154177
+invarlock/reporting/certificate_schema.py,sha256=aZPDvJcFOIcwNJ06iTHJtXAHTmb5BPvv_T7qaSTjW8o,8978
+invarlock/reporting/dataset_hashing.py,sha256=igDq9iMB8sqed5CEWT2fXtx9z5QAxIZ9bbs5RndDDC4,9228
+invarlock/reporting/guards_analysis.py,sha256=JYZ-_Aoc1XI7BrYYm1Pit4n5vTewnt8wqhzMDrWYpO8,38014
 invarlock/reporting/html.py,sha256=d3LQ_brDGYW7GLR_Kj4rimIfGZsAgVE9XQtIjqsJ1rw,1029
-invarlock/reporting/normalizer.py,sha256=hSfNMyUusWZ1OIJ5lVxzKrk8Le1rRqbwOkPY83gCcHM,8078
-invarlock/reporting/policy_utils.py,sha256=Yu0zAtjnf0-w9t03NvH7LJ0lxe98o-E1tD8TpXtzsfo,23080
-invarlock/reporting/primary_metric_utils.py,sha256=F9H8xpEIgbYfj6VS5-QA5aNLL5OwPegdQVKwvdoVVMM,12150
-invarlock/reporting/render.py,sha256=zukCOm6Qm81nDkpOur50TKrlgRpnt7Q53TEU7eYFivg,59976
-invarlock/reporting/report.py,sha256=kyoqTiqtYUQfgpwm6RApV-LbHHIyryfiIPIRhy-z47Q,32293
-invarlock/reporting/report_types.py,sha256=IoCLoPdoB71hW6-UatZXSjG7itTn4dKp7jPRndXxugA,8736
+invarlock/reporting/normalizer.py,sha256=7zUyx9pUPuV6IrEi7PRBFIAid_14ABuU4xhHLjkONCI,8271
+invarlock/reporting/policy_utils.py,sha256=540fw6cFvyvYL7avlvId9GUvY9-g6MkeDlcPt18uuE0,22934
+invarlock/reporting/primary_metric_utils.py,sha256=7h6CzMQzu0A8xN10vbRgSNcQ0YTMxsT2JiNEgCVaXyM,14849
+invarlock/reporting/render.py,sha256=nujyDXgh11dWUtYPru41mx8FnWer7-5wbCiqqK2ae7A,62450
+invarlock/reporting/report.py,sha256=oqNjusy2wseURakW3Icn0MPSDNMm6bpcvKREFTzbd3o,32286
+invarlock/reporting/report_types.py,sha256=cw6UxHS67W0Y8uLEFoplGFrdrr9ktZDyzkvjuQkSERU,8873
 invarlock/reporting/utils.py,sha256=1aLYgSUR4XvgmhDvU9YK9ICd7W5sjft1qdsZC9JJSRY,5540
-invarlock/reporting/validate.py,sha256=oh7w8VrLM2EYZSJ2ySDNi7WKKmF1SgS81VkOrcHMzaE,22427
+invarlock/reporting/validate.py,sha256=aFcac5iaaNOLMYEf6g5xoF4-a-J5E430utFLnfu6fKc,21782
 invarlock/utils/__init__.py,sha256=DR2pBrgddLH2PW-6ninOE8CM7DNvlvgyYsCkckozbPU,4276
 invarlock/utils/digest.py,sha256=sfnqGFRiRf7l950MjSIrWO1XbUfXlcEfNLeWFbBUr8I,1290
-invarlock-0.3.4.dist-info/licenses/LICENSE,sha256=uFddaXYY02nEFdPpS7bam_bnm0st41BibzD0jHULPXw,10413
-invarlock-0.3.4.dist-info/METADATA,sha256=RycW7PQJeWJgfX-e3KbqUi_zjXxDPzLkchaTVoS_SCA,21845
-invarlock-0.3.4.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-invarlock-0.3.4.dist-info/entry_points.txt,sha256=i0e4ZzmJNMBGG-69lbgP-muEcn1je2TUIWwl9SJERm0,670
-invarlock-0.3.4.dist-info/top_level.txt,sha256=GXfftc_YDHHcQC2vQgYbZ5cTO82YuWY3HusHMT3DuKs,10
-invarlock-0.3.4.dist-info/RECORD,,
+invarlock-0.3.6.dist-info/licenses/LICENSE,sha256=uFddaXYY02nEFdPpS7bam_bnm0st41BibzD0jHULPXw,10413
+invarlock-0.3.6.dist-info/METADATA,sha256=PYnrgExSymgE3QAyfJbHect9jVt6cJatWY3zVp7npNo,21875
+invarlock-0.3.6.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+invarlock-0.3.6.dist-info/entry_points.txt,sha256=i0e4ZzmJNMBGG-69lbgP-muEcn1je2TUIWwl9SJERm0,670
+invarlock-0.3.6.dist-info/top_level.txt,sha256=GXfftc_YDHHcQC2vQgYbZ5cTO82YuWY3HusHMT3DuKs,10
+invarlock-0.3.6.dist-info/RECORD,,