invarlock 0.3.5__tar.gz → 0.3.6__tar.gz

{invarlock-0.3.5/src/invarlock.egg-info → invarlock-0.3.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: invarlock
-Version: 0.3.5
+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:** 0.3.5 (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-0.3.5 → invarlock-0.3.6}/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:** 0.3.5 (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)
@@ -64,7 +64,7 @@ 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 \
@@ -73,7 +73,7 @@ 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 \
@@ -143,7 +143,7 @@ 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
@@ -382,7 +382,7 @@ 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-0.3.5 → invarlock-0.3.6}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "invarlock"
-version = "0.3.5"
+version = "0.3.6"
 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-0.3.5 → invarlock-0.3.6}/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__ = "0.3.5"
+__version__ = "0.3.6"
 
 # Core exports - torch-independent
 from .config import CFG, Defaults, get_default_config

{invarlock-0.3.5 → invarlock-0.3.6}/src/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-0.3.5 → invarlock-0.3.6}/src/invarlock/adapters/__init__.py

@@ -76,7 +76,7 @@ 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")

{invarlock-0.3.5 → invarlock-0.3.6}/src/invarlock/calibration/spectral_null.py

@@ -148,7 +148,7 @@ def _selected_families_for_alpha(
 
 
 def summarize_null_sweep_reports(
-    reports: list[dict[str, Any]],
+    reports: list[object],
     *,
     tier: str,
     safety_margin: float = 0.05,
@@ -186,20 +186,25 @@ def summarize_null_sweep_reports(
         mt = _extract_multiple_testing(metrics)
         if mt:
             mt_method = str(mt.get("method", mt_method))
-            if mt.get("alpha") is not None:
-                mt_alpha = float(mt.get("alpha"))
-            if mt.get("m") is not None:
-                mt_m = int(mt.get("m"))
+            alpha_value = mt.get("alpha")
+            if alpha_value is not None:
+                try:
+                    mt_alpha = float(alpha_value)
+                except Exception:
+                    pass
+            m_value = mt.get("m")
+            if m_value is not None:
+                try:
+                    mt_m = int(m_value)
+                except Exception:
+                    pass
 
         fam_z = _extract_family_max_z(metrics)
         for fam, z in fam_z.items():
             family_max_z[fam] = max(family_max_z[fam], float(z))
 
-        selection = (
-            metrics.get("multiple_testing_selection")
-            if isinstance(metrics.get("multiple_testing_selection"), dict)
-            else {}
-        )
+        raw_selection = metrics.get("multiple_testing_selection")
+        selection = raw_selection if isinstance(raw_selection, dict) else {}
         pvals = selection.get("family_pvalues")
         if not isinstance(pvals, dict):
             pvals = {}

{invarlock-0.3.5 → invarlock-0.3.6}/src/invarlock/calibration/variance_ve.py

@@ -107,8 +107,6 @@ def summarize_ve_sweep_reports(
     evaluated = 0
 
     for report in reports:
-        if not isinstance(report, dict):
-            continue
         g = _extract_guard(report, "variance") or {}
         metrics = g.get("metrics", {}) if isinstance(g.get("metrics"), dict) else {}
         pg = metrics.get("predictive_gate")

{invarlock-0.3.5 → invarlock-0.3.6}/src/invarlock/cli/commands/calibrate.py

@@ -144,7 +144,9 @@ def null_sweep(
     ),
     n_seeds: int = typer.Option(10, "--n-seeds", min=1, help="Number of seeds to run."),
     seed_start: int = typer.Option(42, "--seed-start", help="Starting seed."),
-    profile: str = typer.Option("ci", "--profile", help="Run profile (ci|release)."),
+    profile: str = typer.Option(
+        "ci", "--profile", help="Run profile (ci|release|ci_cpu|dev)."
+    ),
     device: str | None = typer.Option(None, "--device", help="Device override."),
     safety_margin: float = typer.Option(
         0.05, "--safety-margin", help="Safety margin applied to κ recommendations."
@@ -363,7 +365,9 @@ def ve_sweep(
         "--target-enable-rate",
         help="Target expected VE enable rate (predictive-gate lower bound).",
     ),
-    profile: str = typer.Option("ci", "--profile", help="Run profile (ci|release)."),
+    profile: str = typer.Option(
+        "ci", "--profile", help="Run profile (ci|release|ci_cpu|dev)."
+    ),
     device: str | None = typer.Option(None, "--device", help="Device override."),
     safety_margin: float = typer.Option(
         0.0,

{invarlock-0.3.5 → invarlock-0.3.6}/src/invarlock/cli/commands/certify.py

@@ -22,9 +22,9 @@ from typing import Any
 import typer
 from rich.console import Console
 
+from ...core.exceptions import MetricsError
 from ..adapter_auto import resolve_auto_adapter
 from ..config import _deep_merge as _merge  # reuse helper
-from ..errors import InvarlockError
 
 # Use the report group's programmatic entry for report generation
 from .report import report_command as _report
@@ -98,7 +98,9 @@ def certify_command(
         "--device",
         help="Device override for runs (auto|cuda|mps|cpu)",
     ),
-    profile: str = typer.Option("ci", "--profile", help="Profile (ci|release)"),
+    profile: str = typer.Option(
+        "ci", "--profile", help="Profile (ci|release|ci_cpu|dev)"
+    ),
     tier: str = typer.Option("balanced", "--tier", help="Tier label for context"),
     preset: str | None = typer.Option(
         None,
@@ -152,9 +154,9 @@ def certify_command(
     # scenario), fall back to a minimal built-in universal preset so the
     # flag-only quick start works without cloning the repo.
     default_universal = (
-        Path("configs/tasks/masked_lm/ci_cpu.yaml")
+        Path("configs/presets/masked_lm/wikitext2_128.yaml")
         if eff_adapter == "hf_bert"
-        else Path("configs/tasks/causal_lm/ci_cpu.yaml")
+        else Path("configs/presets/causal_lm/wikitext2_512.yaml")
     )
     preset_path = Path(preset) if preset is not None else default_universal
 
@@ -185,6 +187,20 @@ def certify_command(
             model_block.pop("device", None)
             preset_data["model"] = model_block
 
+    default_guards_order = ["invariants", "spectral", "rmt", "variance", "invariants"]
+    guards_order = None
+    preset_guards = preset_data.get("guards")
+    if isinstance(preset_guards, dict):
+        preset_order = preset_guards.get("order")
+        if (
+            isinstance(preset_order, list)
+            and preset_order
+            and all(isinstance(item, str) for item in preset_order)
+        ):
+            guards_order = list(preset_order)
+    if guards_order is None:
+        guards_order = list(default_guards_order)
+
     # Create temp baseline config (no-op edit)
     # Normalize possible "hf:" prefixes for HF adapters
     norm_src_id = _normalize_model_id(src_id, eff_adapter)
@@ -199,9 +215,7 @@ def certify_command(
             },
             "edit": {"name": "noop", "plan": {}},
             "eval": {},
-            "guards": {
-                "order": ["invariants", "spectral", "rmt", "variance", "invariants"]
-            },
+            "guards": {"order": guards_order},
             "output": {"dir": str(Path(out) / "source")},
             "context": {"profile": profile, "tier": tier},
         },
@@ -292,15 +306,7 @@ def certify_command(
                 "model": {"id": norm_edt_id, "adapter": eff_adapter},
                 "edit": {"name": "noop", "plan": {}},
                 "eval": {},
-                "guards": {
-                    "order": [
-                        "invariants",
-                        "spectral",
-                        "rmt",
-                        "variance",
-                        "invariants",
-                    ]
-                },
+                "guards": {"order": guards_order},
                 "output": {"dir": str(Path(out) / "edited")},
                 "context": {"profile": profile, "tier": tier},
             },
@@ -325,12 +331,11 @@ def certify_command(
         raise typer.Exit(1)
 
     # CI/Release hard‑abort: fail fast when primary metric is not computable.
-    # Fall back to legacy ppl_* keys when primary_metric block is absent.
     try:
         prof = str(profile or "").strip().lower()
     except Exception:
         prof = ""
-    if prof in {"ci", "release"}:
+    if prof in {"ci", "ci_cpu", "release"}:
         try:
             with Path(edited_report).open("r", encoding="utf-8") as fh:
                 edited_payload = json.load(fh)
@@ -364,35 +369,49 @@ def certify_command(
             else None
         ) or "unknown"
 
-        # Enforce only when a metric block is present; skip for minimal stub reports
-        # Enforce only when a primary_metric block is present
+        # Enforce only when a primary_metric block is present; allow degraded-but-flagged metrics to emit certificates, but fail the task.
         has_metric_block = isinstance(pm, dict) and bool(pm)
         if has_metric_block:
-            # Treat non‑finite PM as hard error in CI/Release (after legacy fallback).
-            # Require a finite final value; preview is optional for legacy reports.
-            if not _finite(pm_final):
-                err = InvarlockError(
+            degraded = bool(pm.get("invalid") or pm.get("degraded"))
+            if degraded or not _finite(pm_final):
+                fallback = pm_prev if _finite(pm_prev) else pm_final
+                if not _finite(fallback) or fallback <= 0:
+                    fallback = 1.0
+                degraded_reason = pm.get("degraded_reason") or (
+                    "non_finite_pm"
+                    if (not _finite(pm_prev) or not _finite(pm_final))
+                    else "primary_metric_degraded"
+                )
+                console.print(
+                    "[yellow]⚠️  Primary metric degraded or non-finite; emitting certificate and marking task degraded. Primary metric computation failed.[/yellow]"
+                )
+                pm["degraded"] = True
+                pm["invalid"] = pm.get("invalid") or True
+                pm["preview"] = pm_prev if _finite(pm_prev) else fallback
+                pm["final"] = pm_final if _finite(pm_final) else fallback
+                pm["ratio_vs_baseline"] = pm_ratio if _finite(pm_ratio) else 1.0
+                pm["degraded_reason"] = degraded_reason
+                metrics["primary_metric"] = pm
+                edited_payload.setdefault("metrics", {}).update(metrics)
+
+                # Emit the certificate for inspection, then exit with a CI-visible error.
+                _report(
+                    run=str(edited_report),
+                    format="cert",
+                    baseline=str(baseline_report),
+                    output=cert_out,
+                )
+                err = MetricsError(
                     code="E111",
-                    message=(
-                        "Primary metric computation failed (NaN/inf). "
-                        f"Context: device={device}, adapter={adapter_name}, edit={edit_name}. "
-                        "Baseline ok; edited failed to compute ppl. "
-                        "Try: use an accelerator (mps/cuda), force float32, reduce max_modules, "
-                        "or lower the evaluation batch size."
-                    ),
+                    message=f"Primary metric degraded or non-finite ({degraded_reason}).",
                     details={
-                        "device": device,
+                        "reason": degraded_reason,
                         "adapter": adapter_name,
+                        "device": device,
                         "edit": edit_name,
-                        "pm_preview": pm_prev,
-                        "pm_final": pm_final,
-                        "pm_ratio": pm_ratio,
                     },
                 )
-                code = _resolve_exit_code(err, profile=prof)
-                console.print(f"[red]{err}[/red]")
-                # Do not emit a certificate
-                raise typer.Exit(code)
+                raise typer.Exit(_resolve_exit_code(err, profile=profile))
 
     console.print("📜 Emitting certificate")
     _report(

{invarlock-0.3.5 → invarlock-0.3.6}/src/invarlock/cli/commands/doctor.py

@@ -188,7 +188,9 @@ def doctor_command(
         None, "--config", "-c", help="Path to YAML config for preflight lints"
     ),
     profile: str | None = typer.Option(
-        None, "--profile", help="Profile to apply for preflight (ci|release)"
+        None,
+        "--profile",
+        help="Profile to apply for preflight (e.g. ci, release, ci_cpu; dev is a no-op)",
     ),
     baseline: str | None = typer.Option(
         None, "--baseline", help="Optional baseline report to check pairing readiness"

{invarlock-0.3.5 → invarlock-0.3.6}/src/invarlock/cli/commands/explain_gates.py

@@ -99,10 +99,6 @@ def explain_gates_command(
         pm = cert.get("primary_metric", {})
         ratio = pm.get("ratio_vs_baseline")
         ratio_ci = pm.get("display_ci")
-    elif isinstance(cert.get("ppl"), dict):  # legacy
-        ppl = cert.get("ppl", {})
-        ratio = ppl.get("ratio_vs_baseline")
-        ratio_ci = ppl.get("ratio_ci")
     hysteresis_applied = bool(validation.get("hysteresis_applied"))
     status = "PASS" if bool(validation.get("primary_metric_acceptable")) else "FAIL"
     console.print("[bold]Gate: Primary Metric vs Baseline[/bold]")
@@ -125,6 +121,63 @@ def explain_gates_command(
             f"  note: hysteresis applied → effective threshold = {limit_with_hyst:.3f}x"
         )
 
+    # Tail gate explanation (warn/fail; based on per-window Δlog-loss vs baseline)
+    pm_tail = (
+        cert.get("primary_metric_tail", {})
+        if isinstance(cert.get("primary_metric_tail"), dict)
+        else {}
+    )
+    if pm_tail:
+        mode = str(pm_tail.get("mode", "warn") or "warn").strip().lower()
+        evaluated = bool(pm_tail.get("evaluated", False))
+        passed = bool(pm_tail.get("passed", True))
+        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)
+        qmax = policy.get("quantile_max")
+        eps = policy.get("epsilon", stats.get("epsilon"))
+        mass = stats.get("tail_mass")
+        mmax = policy.get("mass_max")
+
+        if not evaluated:
+            status_tail = "INFO"
+        elif passed:
+            status_tail = "PASS"
+        elif mode == "fail":
+            status_tail = "FAIL"
+        else:
+            status_tail = "WARN"
+
+        console.print("\n[bold]Gate: Primary Metric Tail (ΔlogNLL)[/bold]")
+        console.print(f"  mode: {mode}")
+        console.print(f"  status: {status_tail}")
+        if isinstance(q_val, int | float):
+            console.print(f"  observed: {q_name}={float(q_val):.4f}")
+        if isinstance(mass, int | float):
+            console.print(f"  tail_mass: Pr[ΔlogNLL > ε]={float(mass):.4f}")
+        thr_parts: list[str] = []
+        if isinstance(qmax, int | float):
+            thr_parts.append(f"{q_name}≤{float(qmax):.4f}")
+        if isinstance(mmax, int | float):
+            thr_parts.append(f"mass≤{float(mmax):.4f}")
+        if isinstance(eps, int | float):
+            thr_parts.append(f"ε={float(eps):.1e}")
+        if thr_parts:
+            console.print("  threshold: " + "; ".join(thr_parts))
+
     # Dataset split visibility from report provenance
     try:
         split = (report_data.get("provenance", {}) or {}).get("dataset_split")
@@ -151,10 +204,6 @@ def explain_gates_command(
                     drift = float(final) / float(preview)
             except Exception:
                 drift = None
-    if isinstance(cert.get("ppl"), dict):  # legacy
-        ppl = cert.get("ppl", {})
-        drift = ppl.get("preview_final_ratio", drift)
-        drift_ci = ppl.get("drift_ci")
     drift_status = (
         "PASS" if bool(validation.get("preview_final_drift_acceptable")) else "FAIL"
     )

{invarlock-0.3.5 → invarlock-0.3.6}/src/invarlock/cli/commands/report.py

@@ -120,7 +120,7 @@ def _generate_reports(
             else:
                 console.print(f"  📄 {fmt.upper()}: {file_path}")
 
-        # Show key metrics (PM-first). Avoid legacy PPL wording.
+        # Show key metrics (PM-first). Avoid PPL-first wording.
         console.print("\n📈 Key Metrics:")
         console.print(f"  Model: {primary_report['meta']['model_id']}")
         console.print(f"  Edit: {primary_report['edit']['name']}")