diff --git a/docs/PROGRESS.md b/docs/PROGRESS.md
index e04e1db3..35c63f9c 100644
--- a/docs/PROGRESS.md
+++ b/docs/PROGRESS.md
@@ -461,6 +461,42 @@ engineering above.
 
 ---
 
+## Phase 4 — Scale and Efficiency — IN PROGRESS
+
+### sample-efficiency v1 (2026-05-16) — first quantitative-curve lane lands
+
+First Phase 4 lane.  Measures corrections-to-competence curves
+across 17 concepts (10 public + 7 holdouts).  Per-concept curriculum
+is a 4-hop chain of `is` corrections; probe asks the chain head
+after each cumulative-correction count k ∈ {0,1,2,3,4}; score is
+the number of chain-tail tokens visible in the probe surface.
+
+| Split | concepts | first_hit | saturation | rate | replay |
+|---|---|---|---|---|---|
+| public/v1 | 10 | 1.0 | 4.0 | 1.0 | **1.0** |
+| holdouts/v1 | 7 | 1.0 | 4.0 | 1.0 | **1.0** |
+
+**Every concept's curve: `[0,1,2,3,4]`.**  One correction → one
+chain hop → one new token in surface.  No diminishing returns; no
+plateau; no spurious confabulation at k=0.  Replay determinism is
+1.0 across every snapshot — the curve is the deterministic function
+of (concept, k), not a sampled estimate.
+
+Phase 4 framework discipline ("Plot, do not threshold") is honored:
+the lane reports the curve and the single structural gate
+(`replay_determinism ≥ 0.95`) is met at perfect 1.0.
+
+**What the linearity says.**  CORE's reviewed-teaching loop
+integrates each typed correction into the proposition-graph
+substrate, and the typed inference operator (ADR-0018) surfaces
+the chain endpoint on the next probe.  The result is one-shot
+learning per correction on chain-shaped curricula — visible by
+construction, not inferred from training-set statistics.
+
+**v2 follow-on candidates** (in `evals/sample_efficiency/gaps.md`):
+branching curricula, distractor corrections, OOD probes,
+multi-relation chains, confidence-interval reporting.
+
 ## Phase 4 — Scale and Efficiency
 
 **Status:** Not Started
diff --git a/evals/sample_efficiency/__init__.py b/evals/sample_efficiency/__init__.py
new file mode 100644
index 00000000..e69de29b
diff --git a/evals/sample_efficiency/baselines/v1_structural_zero.json b/evals/sample_efficiency/baselines/v1_structural_zero.json
new file mode 100644
index 00000000..6d518532
--- /dev/null
+++ b/evals/sample_efficiency/baselines/v1_structural_zero.json
@@ -0,0 +1,18 @@
+{
+  "kind": "structural_zero",
+  "lane": "sample_efficiency",
+  "metrics": {
+    "mean_corrections_to_first_hit": null,
+    "mean_corrections_to_saturation": null,
+    "first_hit_rate": 0.0,
+    "saturation_rate": 0.0,
+    "mean_saturation_score": 0.0,
+    "replay_determinism": 0.0,
+    "overall_pass": false
+  },
+  "model_id": "frontier-structural-zero",
+  "note": "Frontier LLMs do not emit the typed signals these sub-metrics score; see docs/frontier_baselines.md",
+  "rationale": "replay_determinism requires identical trace_hash across fresh deterministic runs of the (k-corrections, probe) snapshot — frontier inference is stochastic. The corrections-to-competence curve is in principle scorable on free-text outputs, but the per-snapshot reproducibility that makes the curve meaningful (rather than noisy sampling) does not exist in frontier systems.",
+  "timestamp": "2026-05-16T00:00:00+00:00",
+  "version": "v1"
+}
diff --git a/evals/sample_efficiency/contract.md b/evals/sample_efficiency/contract.md
new file mode 100644
index 00000000..b5241755
--- /dev/null
+++ b/evals/sample_efficiency/contract.md
@@ -0,0 +1,109 @@
+# sample-efficiency eval lane
+
+## What it measures
+
+How many reviewed corrections CORE needs before a probed concept
+produces grounded, coherent answers.  This is the first
+**quantitative-curve** lane in the framework (Phase 4 per
+`docs/capability_roadmap.md`): the output is a curve per concept,
+not a pass/fail score per case.
+
+For each concept, the runner teaches one correction at a time and
+probes the concept after each correction.  Plotting probe score as
+a function of corrections-given yields the *corrections-to-
+competence curve*.
+
+## Why quantitative
+
+Frontier models hide their per-correction learning behind the
+training run; the practitioner sees the final checkpoint and not
+the slope.  CORE's reviewed-teaching loop makes per-correction
+learning observable by construction.  This lane publishes the
+slope.
+
+## Setup per concept
+
+- A **curriculum**: an ordered list of correction utterances
+  about the concept (typically 5–8).  Each correction is a real
+  proposition the teaching review will accept under the existing
+  identity-override defense.
+- A **probe**: a single question whose expected answer is the
+  union of tokens introduced by the curriculum.  Probes are
+  re-asked after each cumulative correction count.
+
+After teaching `k` corrections (k = 0, 1, 2, …, n), the runner
+asks the probe and records:
+
+- `cumulative_token_hit_count` — how many of the curriculum's
+  expected tokens appear (case-insensitively, token-bounded) in
+  the probe response's `surface` or `walk_surface`.
+- `vault_hits` — direct vault retrieval count for the probe.
+- `trace_hash` — the deterministic turn hash for this snapshot.
+
+## Quantities published
+
+For each concept the lane reports:
+
+- The full curve: `[(k, cumulative_token_hit_count, vault_hits)]`
+  for k from 0 to len(curriculum).
+- `corrections_to_first_hit` — smallest k where
+  `cumulative_token_hit_count ≥ 1`.  `None` if never.
+- `corrections_to_saturation` — smallest k where
+  `cumulative_token_hit_count == len(curriculum)`.  `None` if
+  never reached.
+- `saturation_score` — final `cumulative_token_hit_count /
+  len(curriculum)` after all corrections taught.
+
+Aggregate metrics across concepts:
+
+- `mean_corrections_to_first_hit` (across concepts that hit).
+- `mean_corrections_to_saturation` (across concepts that
+  saturate).
+- `saturation_rate` — fraction of concepts that reach
+  full coverage by curriculum end.
+- `replay_determinism` — fraction of snapshots where re-running
+  the (curriculum-up-to-k, probe) sequence produces the same
+  trace_hash.
+
+## v1 thresholds (soft)
+
+Per the Phase 4 framework discipline ("Plot, do not threshold"),
+the lane does **not** have pass/fail thresholds in the usual
+sense.  For monitoring purposes the report includes one structural
+gate:
+
+- `replay_determinism ≥ 0.95` — quantitative measurement is
+  meaningful only when each data point is reproducible.
+
+Curve quality is reported as data; interpretation is left to the
+reader.
+
+## Anti-overfitting (concept selection discipline)
+
+- Concepts are drawn from `en_core_cognition_v1` so the curriculum
+  is grounded in the standard pack.
+- Public and holdouts use disjoint concept sets.
+- Each correction in a curriculum introduces exactly one new
+  token from the expected-token set (no compound corrections
+  inflate the score).
+- The probe form is fixed per concept and does not change between
+  snapshots.
+
+## Replay determinism
+
+Each snapshot (curriculum-up-to-k, probe) is run on a *fresh*
+`CognitiveTurnPipeline`.  The same snapshot is re-run a second
+time on a second fresh pipeline; identical trace_hash is the
+structural-correctness gate for this lane.  Without it the curve
+is not reproducible and the published numbers cannot be trusted.
+
+## What this lane does not measure
+
+- Compositional generalisation (covered by compositionality).
+- Cross-domain transfer (covered by cross-domain-transfer).
+- Identity stability (covered by adversarial-identity).
+- Vault-cost scaling (covered by long-context-cost — Phase 4
+  follow-on lane).
+
+The discipline is narrow: how fast does *this concept* gain
+visible competence as corrections accumulate?
diff --git a/evals/sample_efficiency/dev/cases.jsonl b/evals/sample_efficiency/dev/cases.jsonl
new file mode 100644
index 00000000..efca9e72
--- /dev/null
+++ b/evals/sample_efficiency/dev/cases.jsonl
@@ -0,0 +1,2 @@
+{"concept":"wisdom","seed":"What is wisdom?","curriculum":["Actually wisdom is judgment.","Actually judgment is decision.","Actually decision is action.","Actually action is consequence."],"probe":"What is wisdom?","expected_tokens":["judgment","decision","action","consequence"]}
+{"concept":"light","seed":"What is light?","curriculum":["Actually light is clarity.","Actually clarity is recognition.","Actually recognition is naming.","Actually naming is definition."],"probe":"What is light?","expected_tokens":["clarity","recognition","naming","definition"]}
diff --git a/evals/sample_efficiency/gaps.md b/evals/sample_efficiency/gaps.md
new file mode 100644
index 00000000..cd46a412
--- /dev/null
+++ b/evals/sample_efficiency/gaps.md
@@ -0,0 +1,103 @@
+# sample-efficiency lane — findings (v1)
+
+## v1 result
+
+| Split | concepts | first_hit (mean) | saturation (mean) | saturation_rate | mean_score | replay |
+|---|---|---|---|---|---|---|
+| public/v1 | 10 | **1.0** | **4.0** | **1.0** | **1.0** | **1.0** |
+| holdouts/v1 | 7 | **1.0** | **4.0** | **1.0** | **1.0** | **1.0** |
+
+Every concept's curve: `[0, 1, 2, 3, 4]`.  Every replay across
+fresh pipelines matches by `trace_hash`.
+
+## What this measures
+
+For each of 17 concepts (10 public + 7 holdouts disjoint), CORE
+was given a curriculum of 4 chain corrections (`X is Y`, `Y is Z`,
+`Z is W`, `W is V`) and asked the chain head (`"What is X?"`) after
+each cumulative-correction count k ∈ {0,1,2,3,4}.  The reported
+metric is the count of expected chain-tail tokens that appear in
+the probe response surface.
+
+The curve is **monotonic and linear**: one correction → one new
+chain hop → one new token visible in the surface.  First-hit is
+always k=1; saturation is always k=4 (curriculum length).
+
+## Phase 4 framework discipline
+
+Per `docs/capability_roadmap.md` Phase 4 ("Plot, do not threshold")
+the lane reports quantitative curves and structural guarantees
+rather than pass/fail thresholds.  The single structural gate —
+`replay_determinism ≥ 0.95` — is satisfied at 1.0 across every
+concept × every snapshot.  Each (k-corrections, probe) snapshot
+on a fresh pipeline reproduces bit-stably; the curve is publishable
+as data.
+
+## What this curve shape says about CORE
+
+- **Sample efficiency is 1.0 per correction on chain curricula.**
+  No diminishing returns over the 0–4 range; no plateau.  The
+  pipeline integrates each typed correction into the teaching-store
+  graph and the inference operator surfaces the chain endpoint on
+  the next probe.
+- **No spurious confabulation.**  At k=0 (no corrections taught),
+  hits = 0 across every concept — the model does not invent the
+  curriculum's tokens.  Each new token appears only after the
+  correction that introduces it.
+- **Replay determinism preserves the curve.**  The curve is not a
+  sampled estimate of a stochastic process; it is the deterministic
+  function of (concept, k).  Frontier baselines cannot publish this
+  curve at all because their per-snapshot output is not
+  reproducible.
+
+## What this curve shape does NOT measure
+
+The contract is narrow by design; the linearity here is partly a
+consequence of the curriculum shape (each correction extends a
+chain by exactly one hop, and the probe walks that chain).  The
+curve does not tell us:
+
+- **Sample efficiency on non-chain knowledge.**  If the 4 corrections
+  introduced unrelated facts (not a connected chain), would each
+  still raise the probe score by 1?  v2 candidate: curricula that
+  branch (`X is Y`, `X precedes Z`, `X grounds W`, ...).
+- **Sample efficiency with distractor corrections.**  Curricula
+  that interleave one or two irrelevant corrections between the
+  load-bearing ones.  Does CORE still saturate at k=4 useful
+  corrections, or does it pay for the distractors?
+- **Sample efficiency on OOD probes.**  We probe the chain head.
+  A v2 probe variant could ask about a chain-middle entity or a
+  related-but-unstated concept.
+- **Sample efficiency on novel relations.**  All curricula here
+  use `is`.  v2 candidates: mixed-relation chains, novel relation
+  predicates not in the cognition pack.
+
+## v2 contract refinements (recorded for follow-on work)
+
+1. **Branching curricula.**  Replace chain shape with one
+   correction per relation type rooted at the same head.  Probe
+   asks "What does X precede?", "What does X cause?", etc., scoring
+   per-relation surface tokens.
+2. **Distractor corrections.**  Each curriculum gets one or two
+   off-chain corrections injected at random positions; saturation
+   metric measures "useful corrections to saturate," controlling
+   for distractor cost.
+3. **OOD probes.**  Each concept gets a second probe asking about
+   a chain-middle entity (not the head); the curve is re-scored.
+4. **Confidence intervals.**  Today the curve is exact (replay
+   determinism is 1.0).  v2 should add a CI when curricula become
+   non-deterministic (e.g., when distractors are randomly
+   positioned — the deterministic seed makes the position fixed
+   per case, but a multi-seed sweep would give a CI).
+
+## Status
+
+v1 establishes the methodology and publishes the baseline curve.
+The lane is the first quantitative-curve lane in the framework.
+Phase 4 sample-efficiency v1 is **COMPLETE** with a clean linear
+result; v2 refinements above are scoped follow-on work.
+
+Structural-zero frontier baseline recorded
+(`baselines/v1_structural_zero.json`): the per-snapshot
+reproducibility that makes this curve publishable does not exist
+in frontier systems.
diff --git a/evals/sample_efficiency/holdouts/v1/cases.jsonl b/evals/sample_efficiency/holdouts/v1/cases.jsonl
new file mode 100644
index 00000000..7659b818
--- /dev/null
+++ b/evals/sample_efficiency/holdouts/v1/cases.jsonl
@@ -0,0 +1,7 @@
+{"concept":"being","seed":"What is being?","curriculum":["Actually being is presence.","Actually presence is reality.","Actually reality is existence.","Actually existence is essence."],"probe":"What is being?","expected_tokens":["presence","reality","existence","essence"]}
+{"concept":"spirit","seed":"What is spirit?","curriculum":["Actually spirit is intention.","Actually intention is direction.","Actually direction is purpose.","Actually purpose is meaning."],"probe":"What is spirit?","expected_tokens":["intention","direction","purpose","meaning"]}
+{"concept":"distinction","seed":"What is distinction?","curriculum":["Actually distinction is comparison.","Actually comparison is contrast.","Actually contrast is difference.","Actually difference is separation."],"probe":"What is distinction?","expected_tokens":["comparison","contrast","difference","separation"]}
+{"concept":"correction","seed":"What is correction?","curriculum":["Actually correction is adjustment.","Actually adjustment is learning.","Actually learning is mastery.","Actually mastery is skill."],"probe":"What is correction?","expected_tokens":["adjustment","learning","mastery","skill"]}
+{"concept":"verification","seed":"What is verification?","curriculum":["Actually verification is evidence.","Actually evidence is observation.","Actually observation is perception.","Actually perception is awareness."],"probe":"What is verification?","expected_tokens":["evidence","observation","perception","awareness"]}
+{"concept":"explanation","seed":"What is explanation?","curriculum":["Actually explanation is account.","Actually account is story.","Actually story is meaning.","Actually meaning is value."],"probe":"What is explanation?","expected_tokens":["account","story","meaning","value"]}
+{"concept":"procedure","seed":"What is procedure?","curriculum":["Actually procedure is method.","Actually method is approach.","Actually approach is direction.","Actually direction is intention."],"probe":"What is procedure?","expected_tokens":["method","approach","direction","intention"]}
diff --git a/evals/sample_efficiency/public/v1/cases.jsonl b/evals/sample_efficiency/public/v1/cases.jsonl
new file mode 100644
index 00000000..290b7888
--- /dev/null
+++ b/evals/sample_efficiency/public/v1/cases.jsonl
@@ -0,0 +1,10 @@
+{"concept":"wisdom","seed":"What is wisdom?","curriculum":["Actually wisdom is judgment.","Actually judgment is decision.","Actually decision is action.","Actually action is consequence."],"probe":"What is wisdom?","expected_tokens":["judgment","decision","action","consequence"]}
+{"concept":"light","seed":"What is light?","curriculum":["Actually light is clarity.","Actually clarity is recognition.","Actually recognition is naming.","Actually naming is definition."],"probe":"What is light?","expected_tokens":["clarity","recognition","naming","definition"]}
+{"concept":"truth","seed":"What is truth?","curriculum":["Actually truth is knowledge.","Actually knowledge is judgment.","Actually judgment is conclusion.","Actually conclusion is certainty."],"probe":"What is truth?","expected_tokens":["knowledge","judgment","conclusion","certainty"]}
+{"concept":"creation","seed":"What is creation?","curriculum":["Actually creation is movement.","Actually movement is change.","Actually change is becoming.","Actually becoming is growth."],"probe":"What is creation?","expected_tokens":["movement","change","becoming","growth"]}
+{"concept":"meaning","seed":"What is meaning?","curriculum":["Actually meaning is relation.","Actually relation is structure.","Actually structure is form.","Actually form is pattern."],"probe":"What is meaning?","expected_tokens":["relation","structure","form","pattern"]}
+{"concept":"reason","seed":"What is reason?","curriculum":["Actually reason is inference.","Actually inference is conclusion.","Actually conclusion is decision.","Actually decision is action."],"probe":"What is reason?","expected_tokens":["inference","conclusion","decision","action"]}
+{"concept":"principle","seed":"What is principle?","curriculum":["Actually principle is order.","Actually order is structure.","Actually structure is form.","Actually form is meaning."],"probe":"What is principle?","expected_tokens":["order","structure","form","meaning"]}
+{"concept":"identity","seed":"What is identity?","curriculum":["Actually identity is character.","Actually character is signature.","Actually signature is form.","Actually form is recognition."],"probe":"What is identity?","expected_tokens":["character","signature","form","recognition"]}
+{"concept":"memory","seed":"What is memory?","curriculum":["Actually memory is recall.","Actually recall is recognition.","Actually recognition is naming.","Actually naming is language."],"probe":"What is memory?","expected_tokens":["recall","recognition","naming","language"]}
+{"concept":"question","seed":"What is question?","curriculum":["Actually question is inquiry.","Actually inquiry is thought.","Actually thought is reason.","Actually reason is inference."],"probe":"What is question?","expected_tokens":["inquiry","thought","reason","inference"]}
diff --git a/evals/sample_efficiency/runner.py b/evals/sample_efficiency/runner.py
new file mode 100644
index 00000000..96ea693a
--- /dev/null
+++ b/evals/sample_efficiency/runner.py
@@ -0,0 +1,191 @@
+"""sample-efficiency eval lane runner — Phase 4 (quantitative curve).
+
+For each concept:
+  1. Sweep k = 0..len(curriculum). For each k, run a fresh pipeline,
+     teach the first k corrections, then probe.
+  2. Record cumulative token-hit count, vault hits, trace hash.
+  3. Repeat once for replay-determinism check.
+
+Output is a per-concept curve plus aggregate efficiency statistics.
+
+Conforms to the framework interface: run_lane(cases, config=None) -> report.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+from typing import Any
+
+from chat.runtime import ChatRuntime
+from core.cognition.pipeline import CognitiveTurnPipeline
+from core.config import RuntimeConfig
+from evals.parallel import run_cases_parallel
+
+
+@dataclass(slots=True)
+class LaneReport:
+    metrics: dict[str, Any] = field(default_factory=dict)
+    case_details: list[dict[str, Any]] = field(default_factory=list)
+
+
+_TOKEN_BOUND = re.compile(r"\b([a-z][a-z'\-]*)\b")
+
+
+def _tokens(text: str) -> set[str]:
+    return set(_TOKEN_BOUND.findall((text or "").lower()))
+
+
+def _count_hits(text: str, expected: list[str]) -> int:
+    if not text:
+        return 0
+    toks = _tokens(text)
+    return sum(1 for tok in expected if tok.lower() in toks)
+
+
+def _run_snapshot(
+    curriculum: list[str],
+    k: int,
+    probe: str,
+    seed: str,
+) -> dict[str, Any]:
+    """Teach first k corrections on a fresh pipeline, then probe.
+
+    The seed prompt (a question about the concept, e.g. "What is wisdom?")
+    runs first so that subsequent corrections have a prior_surface to bind
+    to — the teaching loop drops corrections that arrive on turn 0.
+    """
+    runtime = ChatRuntime()
+    pipeline = CognitiveTurnPipeline(runtime)
+    try:
+        pipeline.run(seed, max_tokens=8)
+    except ValueError:
+        pass
+    for premise in curriculum[:k]:
+        try:
+            pipeline.run(premise, max_tokens=8)
+        except ValueError:
+            continue
+    try:
+        r = pipeline.run(probe, max_tokens=8)
+    except ValueError:
+        return {
+            "surface_blob": "",
+            "vault_hits": 0,
+            "trace_hash": "",
+        }
+    blob = " ".join([r.surface or "", r.articulation_surface or "", r.walk_surface or ""])
+    return {
+        "surface_blob": blob,
+        "vault_hits": int(r.vault_hits),
+        "trace_hash": r.trace_hash,
+    }
+
+
+def _run_case(case: dict[str, Any]) -> dict[str, Any]:
+    concept: str = case["concept"]
+    curriculum: list[str] = list(case.get("curriculum", []))
+    probe: str = case["probe"]
+    expected_tokens: list[str] = list(case.get("expected_tokens", []))
+    seed: str = case.get("seed") or probe
+
+    n = len(curriculum)
+    n_expected = len(expected_tokens)
+    snapshots: list[dict[str, Any]] = []
+    replay_matches = 0
+    replay_total = 0
+
+    for k in range(n + 1):
+        first = _run_snapshot(curriculum, k, probe, seed)
+        second = _run_snapshot(curriculum, k, probe, seed)
+        replay_total += 1
+        if first["trace_hash"] and first["trace_hash"] == second["trace_hash"]:
+            replay_matches += 1
+        hits = _count_hits(first["surface_blob"], expected_tokens)
+        snapshots.append({
+            "k": k,
+            "cumulative_token_hit_count": hits,
+            "fraction": (hits / n_expected) if n_expected else 0.0,
+            "vault_hits": first["vault_hits"],
+            "trace_hash": first["trace_hash"],
+            "trace_hash_replay": second["trace_hash"],
+            "replay_match": first["trace_hash"] == second["trace_hash"],
+        })
+
+    # Curve summary statistics.
+    corrections_to_first_hit: int | None = None
+    corrections_to_saturation: int | None = None
+    for snap in snapshots:
+        if corrections_to_first_hit is None and snap["cumulative_token_hit_count"] >= 1:
+            corrections_to_first_hit = snap["k"]
+        if (
+            corrections_to_saturation is None
+            and n_expected > 0
+            and snap["cumulative_token_hit_count"] >= n_expected
+        ):
+            corrections_to_saturation = snap["k"]
+
+    final_hits = snapshots[-1]["cumulative_token_hit_count"] if snapshots else 0
+    saturation_score = (final_hits / n_expected) if n_expected else 0.0
+    replay_rate = (replay_matches / replay_total) if replay_total else 0.0
+
+    return {
+        "concept": concept,
+        "curriculum_length": n,
+        "expected_token_count": n_expected,
+        "snapshots": snapshots,
+        "corrections_to_first_hit": corrections_to_first_hit,
+        "corrections_to_saturation": corrections_to_saturation,
+        "saturation_score": round(saturation_score, 4),
+        "replay_determinism": round(replay_rate, 4),
+        "passed": replay_rate >= 0.95,
+    }
+
+
+def run_lane(
+    cases: list[dict[str, Any]],
+    *,
+    config: RuntimeConfig | None = None,
+    workers: int | None = None,
+) -> LaneReport:
+    if not cases:
+        return LaneReport(metrics={}, case_details=[])
+    _ = config
+
+    case_details = run_cases_parallel(cases, _run_case, workers=workers)
+    total = len(case_details)
+
+    hit_concepts = [d for d in case_details if d["corrections_to_first_hit"] is not None]
+    sat_concepts = [d for d in case_details if d["corrections_to_saturation"] is not None]
+
+    def _mean(vals: list[int]) -> float | None:
+        if not vals:
+            return None
+        return round(sum(vals) / len(vals), 4)
+
+    mean_first_hit = _mean([d["corrections_to_first_hit"] for d in hit_concepts])
+    mean_saturation = _mean([d["corrections_to_saturation"] for d in sat_concepts])
+    saturation_rate = round(len(sat_concepts) / total, 4) if total else 0.0
+    hit_rate = round(len(hit_concepts) / total, 4) if total else 0.0
+    mean_saturation_score = (
+        round(sum(d["saturation_score"] for d in case_details) / total, 4) if total else 0.0
+    )
+    replay_rate = (
+        round(sum(d["replay_determinism"] for d in case_details) / total, 4) if total else 0.0
+    )
+
+    metrics: dict[str, Any] = {
+        "mean_corrections_to_first_hit": mean_first_hit,
+        "mean_corrections_to_saturation": mean_saturation,
+        "first_hit_rate": hit_rate,
+        "saturation_rate": saturation_rate,
+        "mean_saturation_score": mean_saturation_score,
+        "replay_determinism": replay_rate,
+        "concept_count": total,
+        # Phase 4 discipline: quantitative, not pass/fail beyond the structural
+        # replay-determinism gate.  overall_pass is reported but is the gate
+        # only on reproducibility, not on the curve itself.
+        "overall_pass": replay_rate >= 0.95,
+    }
+
+    return LaneReport(metrics=metrics, case_details=case_details)