diff --git a/.amplifier/evaluation/.gitignore b/.amplifier/evaluation/.gitignore
index a4c12a55..d11aabeb 100644
--- a/.amplifier/evaluation/.gitignore
+++ b/.amplifier/evaluation/.gitignore
@@ -2,6 +2,20 @@
 # Never source controlled (matches the existing per-harness runs/ convention).
 runs/
 
+# Personal analysis tooling and scratch. Useful locally, not part of the
+# harness contract, so it stays out of the tracked tree.
+local/
+
+# Scratch the AI User is instructed to write when a prompt contains a double
+# quote (see agents/*/invocation.md). It belongs at /workspace/eval-prompt.txt
+# INSIDE the DTU, but the AI User drives from this directory on the host, so a
+# missing path prefix drops it here instead. swe_bench_lifecycle.DRIVER_SCRATCH
+# already cleans these in-DTU; this keeps a host-side leak out of the repo.
+eval-prompt.txt
+eval-run.out
+eval-run.done
+answer.txt
+
 # Python / uv
 .venv/
 __pycache__/
diff --git a/.amplifier/evaluation/agents/amplifier-agent-local/install.yaml b/.amplifier/evaluation/agents/amplifier-agent-local/install.yaml
index 26fdabb6..90227d22 100644
--- a/.amplifier/evaluation/agents/amplifier-agent-local/install.yaml
+++ b/.amplifier/evaluation/agents/amplifier-agent-local/install.yaml
@@ -23,9 +23,17 @@ setup_cmds:
   # Write the host-config JSON that every task invocation passes via --config.
   # It sets headless approval, pins the provider/model, and adds an extra skills
   # source location (/root/extra-skills) used by the skill-config-location task.
+  #
+  # `debug.rawLlmPayloads: true` folds into the provider config as `raw: true`,
+  # which makes the provider attach the FULL outbound request kwargs to
+  # `llm:request` and the accumulated final response to `llm:response`. Those
+  # events are written to the session events.jsonl that extract.yaml already
+  # pulls (glob **/events.jsonl), so no extraction change is needed. This is the
+  # analysis surface for the eval: the actual prompts and responses, not just
+  # token counts. Contract covered by tests/e2e/suites/raw_capture.
   - |
     cat > /root/host-config.json <<'JSON'
-    {"approval":{"mode":"yes"},"provider":{"module":"anthropic","config":{"default_model":"claude-sonnet-5"}},"skills":{"skills":["/root/extra-skills"]}}
+    {"approval":{"mode":"yes"},"provider":{"module":"anthropic","config":{"default_model":"claude-sonnet-5"}},"skills":{"skills":["/root/extra-skills"]},"debug":{"rawLlmPayloads":true}}
     JSON
   - mkdir -p /root/extra-skills
   # install setup_cmds run under a login shell. Writing this profile.d file
diff --git a/.amplifier/evaluation/agents/amplifier-foundation/install.yaml b/.amplifier/evaluation/agents/amplifier-foundation/install.yaml
index 4b052484..d352f41c 100644
--- a/.amplifier/evaluation/agents/amplifier-foundation/install.yaml
+++ b/.amplifier/evaluation/agents/amplifier-foundation/install.yaml
@@ -13,9 +13,19 @@
 # `requires.env` is verified on the host before launch and merged into the task
 # profile's passthrough services, so ANTHROPIC_API_KEY reaches the agent.
 #
-# Opus 4.8 is pinned via the opus48 routing matrix (routes all model_role
-# mappings to claude-opus-4-8) so this variant runs the SAME model as the
-# opencode + amplifier-agent stack; the only difference is the agent stack.
+# Sonnet 5 is pinned via the provider's `default_model` so this variant runs the
+# SAME model as the opencode + amplifier-agent stack; the only difference is the
+# agent stack.
+#
+# NO `routing.matrix` is set, deliberately. A previous revision named a matrix
+# (`opus48`) that does not exist in the routing-matrix bundle; the reference
+# resolved to nothing and the pin silently came from `default_model` alone. That
+# was verified against a real run: every one of the 84 LLM requests across three
+# tasks -- including the delegated sub-agent sessions -- carried the
+# `default_model` value. Naming a matrix here would only re-introduce a
+# role-based fan-out (`reasoning`/`creative`/`writing` -> opus under the stock
+# `anthropic` matrix), which is precisely what a single-model pin must avoid.
+# `default_model` is the whole mechanism; keep it that way.
 requires:
   env:
     - ANTHROPIC_API_KEY
@@ -37,12 +47,10 @@ setup_cmds:
           source: git+https://github.com/microsoft/amplifier-module-provider-anthropic@main
           config:
             api_key: $ANTHROPIC_API_KEY
-            default_model: claude-opus-4-8
+            default_model: claude-sonnet-5
             enable_1m_context: 'true'
             enable_prompt_caching: 'true'
             priority: 1
-      routing:
-        matrix: opus48
     EOF
 
   # Compose the anchors bundle as an app bundle (same mechanism the foundation
@@ -56,8 +64,20 @@ setup_cmds:
   # `includes:` the FULL amplifier-foundation bundle, which would drag all of
   # foundation on top of the lean anchors variant and defeat its purpose. The
   # behaviors keep the composition lean: anchors stays the active bundle and only
-  # the two CI agents are added -- no full foundation, no design mode, no logging
-  # hook (that lives only in the separate -logging behavior).
+  # the two CI agents are added -- no full foundation, no design mode, and no
+  # ADDITIONAL logging hook (that lives in the separate -logging behavior, which
+  # we do not add here).
+  #
+  # To be precise about what IS logging, because it caught us out once: anchors
+  # itself already composes TWO loggers upstream --
+  # `foundation:behaviors/logging` (hooks-logging -> <session>/events.jsonl) and
+  # `context-intelligence:behaviors/context-intelligence-logging`
+  # (hook-context-intelligence -> <session>/context-intelligence/events.jsonl).
+  # So every LLM call in this variant is written to disk TWICE, in two different
+  # envelope shapes. That is anchors' own design and is deliberately NOT patched
+  # here -- this agent evaluates anchors as published. The harness is what has to
+  # cope: `metrics.parse_events` de-duplicates by response identity. Before that
+  # fix this variant reported exactly double its real calls, tokens and cost.
   #
   # We add BOTH behavior files explicitly and on purpose:
   #   - navigation (LAYER 1) registers session-navigator (reads local session
diff --git a/.amplifier/evaluation/agents/amplifier-foundation/invocation.md b/.amplifier/evaluation/agents/amplifier-foundation/invocation.md
index 66491b6c..3dd3a2e9 100644
--- a/.amplifier/evaluation/agents/amplifier-foundation/invocation.md
+++ b/.amplifier/evaluation/agents/amplifier-foundation/invocation.md
@@ -98,10 +98,10 @@ took. Do NOT judge correctness yourself -- the grader does that.
 
 ## Model pinning
 
-Opus 4.8 is pinned at install time via the provider's `default_model:
-claude-opus-4-8` plus the `opus48` routing matrix in
-`/root/.amplifier/settings.yaml`. That is the single source of truth for which
-model runs, so `amplifier run` needs no model flag -- just invoke it as shown.
+Sonnet 5 is pinned at install time via the provider's `default_model:
+claude-sonnet-5` in `/root/.amplifier/settings.yaml`, with no routing matrix
+set. That is the single source of truth for which model runs, so `amplifier run`
+needs no model flag -- just invoke it as shown.
 
 ## Notes
 
diff --git a/.amplifier/evaluation/agents/amplifier-foundation/meta.yaml b/.amplifier/evaluation/agents/amplifier-foundation/meta.yaml
index 607ebf9e..8366716f 100644
--- a/.amplifier/evaluation/agents/amplifier-foundation/meta.yaml
+++ b/.amplifier/evaluation/agents/amplifier-foundation/meta.yaml
@@ -12,8 +12,8 @@ id: amplifier-foundation
 description: >
   The Amplifier CLI with the anchors bundle (amplifier-foundation@main,
   bundles/anchors) plus the context-intelligence bundle (agents only, no server)
-  composed, driven via `amplifier run` on Claude Opus 4.8. A general-agent variant
+  composed, driven via `amplifier run` on Claude Sonnet 5. A general-agent variant
   under test, run on the same tasks as the opencode + amplifier-agent stack for
-  head-to-head comparison. Opus is pinned via the opus48 routing matrix so the
-  model matches the opencode agent.
-model: claude-opus-4-8
+  head-to-head comparison. Sonnet 5 is pinned via the provider's default_model
+  so the model matches every other agent-under-test.
+model: claude-sonnet-5
diff --git a/.amplifier/evaluation/agents/opencode-amplifier-agent/invocation.md b/.amplifier/evaluation/agents/opencode-amplifier-agent/invocation.md
index 6d054e61..2e87d187 100644
--- a/.amplifier/evaluation/agents/opencode-amplifier-agent/invocation.md
+++ b/.amplifier/evaluation/agents/opencode-amplifier-agent/invocation.md
@@ -8,7 +8,7 @@ already on PATH once you export the two tool directories.
 Run each turn with a single command. Note the PATH export (opencode and uv
 tools live in `$HOME/.opencode/bin` and `$HOME/.local/bin`) and the model pin:
 
-    export PATH="$HOME/.opencode/bin:$HOME/.local/bin:$PATH"; cd /workspace && amplifier-opencode launch -- run --auto --model amplifier/claude-opus-4-8 "<your message>"
+    export PATH="$HOME/.opencode/bin:$HOME/.local/bin:$PATH"; cd /workspace && amplifier-opencode launch -- run --auto --model amplifier/claude-sonnet-5 "<your message>"
 
 The CLI prints the agent's final response to stdout and exits. Capture the
 response from stdout.
@@ -25,7 +25,7 @@ on every invocation.
 
 ## Model pinning
 
-opus-4-8 is pinned via opencode's `--model amplifier/claude-opus-4-8` flag. The
+sonnet-5 is pinned via opencode's `--model amplifier/claude-sonnet-5` flag. The
 amplifier adapter writes no default model, so the flag is the single source of
 truth for which model runs. Keep it on every invocation.
 
diff --git a/.amplifier/evaluation/agents/opencode-amplifier-agent/meta.yaml b/.amplifier/evaluation/agents/opencode-amplifier-agent/meta.yaml
index 3c47ec18..23380787 100644
--- a/.amplifier/evaluation/agents/opencode-amplifier-agent/meta.yaml
+++ b/.amplifier/evaluation/agents/opencode-amplifier-agent/meta.yaml
@@ -5,7 +5,7 @@ id: opencode-amplifier-agent
 description: >
   opencode driven by amplifier-agent (OpenAI-compatible backend), running
   Anthropic models. The agent-under-test for evaluation runs.
-model: claude-opus-4-8
+model: claude-sonnet-5
 
 # Agent-owned timing hints for clean agent-only wall clock. The AI User drives
 # this agent by shelling `amplifier-digital-twin
diff --git a/.amplifier/evaluation/agents/opencode-vanilla/install.yaml b/.amplifier/evaluation/agents/opencode-vanilla/install.yaml
index ba3ef95e..71fd1371 100644
--- a/.amplifier/evaluation/agents/opencode-vanilla/install.yaml
+++ b/.amplifier/evaluation/agents/opencode-vanilla/install.yaml
@@ -8,7 +8,7 @@
 # opencode's anthropic provider reads ANTHROPIC_API_KEY from the environment
 # automatically -- no `opencode auth login` is needed.
 #
-# Opus 4.8 (`claude-opus-4-8`) is not in opencode's models.dev catalog, so we
+# Sonnet 5 (`claude-sonnet-5`) is not in opencode's models.dev catalog, so we
 # register it explicitly under the built-in anthropic provider (@ai-sdk/anthropic)
 # in ~/.config/opencode/opencode.json. opencode sends the id to the Anthropic API
 # verbatim.
@@ -30,19 +30,19 @@ setup_cmds:
     export PATH="$HOME/.opencode/bin:$PATH"; opencode --version
   - echo 'export PATH="$HOME/.opencode/bin:$PATH"' > /etc/profile.d/opencode.sh
   - export PATH="$HOME/.opencode/bin:$PATH"; opencode --version
-  # Register Opus 4.8 on the anthropic provider and make it the default model.
+  # Register Sonnet 5 on the anthropic provider and make it the default model.
   - mkdir -p "$HOME/.config/opencode"
   - |
     cat > "$HOME/.config/opencode/opencode.json" <<'JSON'
     {
       "$schema": "https://opencode.ai/config.json",
-      "model": "anthropic/claude-opus-4-8",
+      "model": "anthropic/claude-sonnet-5",
       "provider": {
         "anthropic": {
           "npm": "@ai-sdk/anthropic",
           "models": {
-            "claude-opus-4-8": {
-              "name": "Claude Opus 4.8"
+            "claude-sonnet-5": {
+              "name": "Claude Sonnet 5"
             }
           }
         }
@@ -51,4 +51,4 @@ setup_cmds:
     JSON
   # Best-effort warm-up: exercises the full path once (auth + model resolution)
   # so the operator's first real run is fast. Bounded and non-fatal.
-  - 'export PATH="$HOME/.opencode/bin:$PATH"; timeout 240 opencode run --model anthropic/claude-opus-4-8 --auto "reply with exactly: ok" || true'
+  - 'export PATH="$HOME/.opencode/bin:$PATH"; timeout 240 opencode run --model anthropic/claude-sonnet-5 --auto "reply with exactly: ok" || true'
diff --git a/.amplifier/evaluation/agents/opencode-vanilla/invocation.md b/.amplifier/evaluation/agents/opencode-vanilla/invocation.md
index 0c736288..9cecfba5 100644
--- a/.amplifier/evaluation/agents/opencode-vanilla/invocation.md
+++ b/.amplifier/evaluation/agents/opencode-vanilla/invocation.md
@@ -8,7 +8,7 @@ a one-shot per turn.
 Run each turn with a single command. Note the PATH export (opencode lives in
 `$HOME/.opencode/bin`), the model pin, and `--auto`:
 
-    export PATH="$HOME/.opencode/bin:$PATH"; cd /workspace && opencode run --model anthropic/claude-opus-4-8 --auto "<your message>"
+    export PATH="$HOME/.opencode/bin:$PATH"; cd /workspace && opencode run --model anthropic/claude-sonnet-5 --auto "<your message>"
 
 The CLI prints the agent's final response to stdout and exits. Capture the
 response from stdout.
@@ -21,7 +21,7 @@ agent cannot complete web or file tasks. Always pass `--auto`.
 
 ## Model pinning
 
-opus-4-8 is pinned via `--model anthropic/claude-opus-4-8`. It is also the
+sonnet-5 is pinned via `--model anthropic/claude-sonnet-5`. It is also the
 default model in opencode's config, but keep the flag on every invocation as the
 single source of truth for which model runs.
 
diff --git a/.amplifier/evaluation/agents/opencode-vanilla/meta.yaml b/.amplifier/evaluation/agents/opencode-vanilla/meta.yaml
index cb98c826..81a83ed9 100644
--- a/.amplifier/evaluation/agents/opencode-vanilla/meta.yaml
+++ b/.amplifier/evaluation/agents/opencode-vanilla/meta.yaml
@@ -4,9 +4,9 @@
 id: opencode-vanilla
 description: >
   Vanilla opencode talking directly to the Anthropic API (no amplifier-agent /
-  amplifier-app-opencode bridge), pinned to Claude Opus 4.8. A baseline
+  amplifier-app-opencode bridge), pinned to Claude Sonnet 5. A baseline
   agent-under-test for comparison against opencode-amplifier-agent.
-model: claude-opus-4-8
+model: claude-sonnet-5
 
 # Agent-owned timing hints for clean agent-only wall clock. The AI User drives
 # this agent by shelling `amplifier-digital-twin exec <dtu> -- bash -c '...
diff --git a/.amplifier/evaluation/pyproject.toml b/.amplifier/evaluation/pyproject.toml
index 58568540..9a9045af 100644
--- a/.amplifier/evaluation/pyproject.toml
+++ b/.amplifier/evaluation/pyproject.toml
@@ -40,6 +40,16 @@ amplifier-bundle-evaluation = { git = "https://github.com/microsoft/amplifier-bu
 amplifier-core = { git = "https://github.com/microsoft/amplifier-core", branch = "main" }
 amplifier-foundation = { git = "https://github.com/microsoft/amplifier-foundation", branch = "main" }
 
+# Test-only dependencies. pytest-asyncio is required not by these tests but by
+# amplifier-core's pytest plugin, which loads via a pytest11 entry point as soon
+# as pytest starts in this venv; without it collection fails before any test
+# runs. Install with `uv sync --group dev`.
+[dependency-groups]
+dev = [
+    "pytest>=8.0",
+    "pytest-asyncio>=0.23",
+]
+
 [tool.pyright]
 include = ["src"]
 extraPaths = ["src"]
diff --git a/.amplifier/evaluation/src/eval/metrics.py b/.amplifier/evaluation/src/eval/metrics.py
index 2bb0d742..d597ba43 100644
--- a/.amplifier/evaluation/src/eval/metrics.py
+++ b/.amplifier/evaluation/src/eval/metrics.py
@@ -53,6 +53,7 @@
 
 import datetime
 import glob
+import hashlib
 import json
 import re
 import sqlite3
@@ -278,6 +279,61 @@ def parse_opencode_db(db_paths: list[str], workspace_dir: str = "/workspace") ->
     }
 
 
+def _response_identity(obj: dict) -> str | None:
+    """Return a stable identity for an `llm:response` event, or None.
+
+    WHY THIS EXISTS. A single logical LLM call is written to disk more than once
+    whenever a session composes more than one logging hook. The published
+    `anchors` bundle does exactly this: it includes BOTH
+    `foundation:behaviors/logging` (hooks-logging -> `<session>/events.jsonl`)
+    and `context-intelligence:behaviors/context-intelligence-logging`
+    (hook-context-intelligence -> `<session>/context-intelligence/events.jsonl`).
+    Both files are legitimate telemetry and both are pulled by extraction, so
+    summing across them counted every call, token and dollar TWICE.
+
+    The two copies are not detectable by comparing bytes: the loggers use
+    different envelope shapes (`ts` vs `timestamp`, metadata at the top level vs
+    nested under `data`), so the files share zero identical lines. They are only
+    recognisable by what they DESCRIBE. Hence identity, not equality:
+
+      strong    the provider's own response id (`data.raw.id`). Globally unique
+                per call, present whenever raw payload capture is on.
+      fallback  session id + event timestamp + usage, hashed. Used when raw
+                capture is off. The timestamp is sub-microsecond and is byte
+                identical across both loggers (verified on real run artifacts),
+                so it discriminates distinct calls reliably.
+      None      neither is available -- see the caller. We do NOT guess.
+
+    Returning None on a weak signal is deliberate. Over-de-duplicating would
+    silently UNDERSTATE cost, which is a worse failure than the overcount this
+    function exists to fix: an inflated number invites scrutiny, a deflated one
+    does not.
+    """
+    # Bind before narrowing: `x.get(k) if isinstance(x.get(k), dict) else {}`
+    # calls get() twice, and the isinstance on the first call does not narrow
+    # the second, so the result stays `Any | None`.
+    raw_data = obj.get("data")
+    data: dict[str, Any] = raw_data if isinstance(raw_data, dict) else {}
+
+    raw_payload = data.get("raw")
+    if isinstance(raw_payload, dict):
+        rid = raw_payload.get("id")
+        if isinstance(rid, str) and rid:
+            return f"id:{rid}"
+
+    # Fallback fingerprint. Require a timestamp: without one, two genuinely
+    # distinct calls that happened to use the same token counts would collide.
+    ts = data.get("timestamp") or obj.get("ts") or obj.get("timestamp")
+    if not isinstance(ts, (str, int, float)) or isinstance(ts, bool):
+        return None
+
+    session_id = data.get("session_id") or obj.get("session_id")
+    raw_usage = data.get("usage")
+    usage: dict[str, Any] = raw_usage if isinstance(raw_usage, dict) else {}
+    fingerprint = json.dumps([session_id, str(ts), usage], sort_keys=True, default=str)
+    return "fp:" + hashlib.sha1(fingerprint.encode()).hexdigest()
+
+
 def parse_events(events_paths: list[str]) -> dict[str, Any]:
     """Sum token/cost usage and compute wallclock from Amplifier events.jsonl.
 
@@ -295,7 +351,9 @@ def parse_events(events_paths: list[str]) -> dict[str, Any]:
         (count of files that existed and were readable), had_timestamps (bool),
         agent_wallclock_s (float, 0.0 when no timestamps were found -- callers
         must consult had_timestamps to distinguish that from a genuine 0-length
-        span).
+        span), duplicate_responses (count of duplicate `llm:response` events
+        dropped), unidentified_responses (count that carried no usable identity
+        and were therefore counted without de-duplication).
     """
     input_tokens = 0
     output_tokens = 0
@@ -305,6 +363,9 @@ def parse_events(events_paths: list[str]) -> dict[str, Any]:
     saw_cost = False
     llm_responses = 0
     files_read = 0
+    duplicate_responses = 0
+    unidentified_responses = 0
+    seen_responses: set[str] = set()
 
     min_ts: float | None = None
     max_ts: float | None = None
@@ -338,6 +399,20 @@ def parse_events(events_paths: list[str]) -> dict[str, Any]:
             if obj.get("event") != "llm:response":
                 continue
 
+            # De-duplicate by identity. The same call reaches this loop once per
+            # logging hook the session composed (see _response_identity), and
+            # summing every copy is what inflated cost, tokens and call counts.
+            identity = _response_identity(obj)
+            if identity is None:
+                # No usable identity: count it rather than risk collapsing two
+                # genuinely distinct calls. Surfaced in the notes.
+                unidentified_responses += 1
+            elif identity in seen_responses:
+                duplicate_responses += 1
+                continue
+            else:
+                seen_responses.add(identity)
+
             data = obj.get("data")
             usage = data.get("usage") if isinstance(data, dict) else None
             if not isinstance(usage, dict):
@@ -375,6 +450,8 @@ def parse_events(events_paths: list[str]) -> dict[str, Any]:
         "files_read": files_read,
         "had_timestamps": had_timestamps,
         "agent_wallclock_s": agent_wallclock_s,
+        "duplicate_responses": duplicate_responses,
+        "unidentified_responses": unidentified_responses,
     }
 
 
@@ -384,6 +461,13 @@ def find_events_files(output_dir: Path | str) -> list[str]:
     The extractor pulls session logs under `output_dir/sessions/`, preserving
     each session's `context-intelligence/events.jsonl`. We glob the whole tree
     so nested session directories are all found, sorted for determinism.
+
+    This deliberately returns EVERY events.jsonl, including the several a single
+    session can produce when it composes more than one logging hook. Narrowing
+    the glob to pick a "primary" file would be fragile -- the layout differs per
+    agent and would silently yield zero on any change. Duplicates are handled
+    where they can be handled correctly: `parse_events` de-duplicates by
+    response identity.
     """
     root = Path(output_dir).expanduser().resolve()
     return sorted(str(p) for p in root.rglob("events.jsonl"))
@@ -508,6 +592,10 @@ def _finalize(
     Shared by both the events.jsonl and opencode.db branches: they produce the
     same `parsed` shape, so the `not_available` discipline, rounding, and note
     assembly live here once. Branch-specific wording is passed in.
+
+    `duplicate_responses` / `unidentified_responses` are read with `.get()`
+    because they are meaningful only for the events.jsonl branch -- the opencode
+    branch reads one row per session from SQLite and cannot double-count.
     """
     files_read = parsed["files_read"]
     notes_parts: list[str] = []
@@ -534,6 +622,22 @@ def _finalize(
             f"Parsed {parsed['llm_responses']} {response_label}(s) across "
             f"{files_read} {source_label}(s). {parse_note_suffix}"
         )
+        # State the de-duplication explicitly. Without this the corrected figure
+        # is indistinguishable from a run that simply made fewer calls.
+        dupes = parsed.get("duplicate_responses") or 0
+        if dupes:
+            notes_parts.append(
+                f"Dropped {dupes} duplicate {response_label}(s): this session composes more "
+                f"than one logging hook, so each call was written to disk more than once. "
+                f"Counted once each by response identity."
+            )
+        unknown = parsed.get("unidentified_responses") or 0
+        if unknown:
+            notes_parts.append(
+                f"{unknown} {response_label}(s) carried no response id and no timestamp, so "
+                f"they could not be de-duplicated and are counted as-is; if this session "
+                f"composes multiple logging hooks these may be overcounted."
+            )
         notes_parts.append(cost_absent_note if cost_val == NOT_AVAILABLE else cost_present_note)
         if agent_wc == NOT_AVAILABLE:
             notes_parts.append("agent_wallclock_s is not_available: no timestamps found.")
diff --git a/.amplifier/evaluation/tests/test_metrics_dedup.py b/.amplifier/evaluation/tests/test_metrics_dedup.py
new file mode 100644
index 00000000..cc74da8d
--- /dev/null
+++ b/.amplifier/evaluation/tests/test_metrics_dedup.py
@@ -0,0 +1,214 @@
+"""Regression tests for `llm:response` de-duplication in the metrics pass.
+
+WHY THIS FILE EXISTS. A session that composes more than one logging hook writes
+every LLM call to disk more than once. The published `anchors` bundle does this:
+it includes both `foundation:behaviors/logging` (-> `<session>/events.jsonl`)
+and `context-intelligence:behaviors/context-intelligence-logging` (->
+`<session>/context-intelligence/events.jsonl`). Extraction pulls both files and
+the metrics pass summed across them, so the amplifier-foundation agent reported
+exactly DOUBLE its real calls, tokens and cost -- 20 calls at $2.98 for a trial
+that actually made 10 calls at $1.49.
+
+The bug was invisible to every cheap check. The two files share zero identical
+lines, because the loggers use different envelope shapes (`ts` vs `timestamp`,
+metadata at the top level vs nested under `data`). Only the payload identity
+gives it away. These tests pin that behaviour:
+
+  - the two real envelope shapes, carrying one call, count as one
+  - distinct calls are still counted separately (the fix must not over-collapse)
+  - de-duplication works with raw capture OFF, via the timestamp fingerprint
+  - an event with no usable identity is counted rather than silently dropped
+  - the correction is stated in `notes`, not applied silently
+
+Run:  uv run python -m pytest tests/ -q
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "src"))
+
+from eval.metrics import normalize_metrics, parse_events
+
+# One LLM call, as each of the two loggers actually writes it. Shapes copied
+# from real run artifacts (runs/ab-3x3/.../amplifier-foundation__*).
+RESPONSE_ID = "msg_011CdjqAotUMQC2DjuMppaYo"
+TIMESTAMP = "2026-08-05T15:23:15.602291443+00:00"
+SESSION = "fa3b1b70-e043-406a-8c8f-1fbc3edd24f0"
+USAGE = {
+    "input_tokens": 2,
+    "output_tokens": 398,
+    "cache_write_tokens": 16534,
+    "cost_usd": "0.1132975",
+}
+
+
+def ci_shape(
+    response_id: str = RESPONSE_ID, ts: str = TIMESTAMP, usage: dict | None = None
+) -> dict:
+    """hook-context-intelligence: `timestamp` inside `data`, `workspace` on top."""
+    return {
+        "event": "llm:response",
+        "timestamp": ts,
+        "workspace": "-workspace",
+        "data": {
+            "session_id": SESSION,
+            "timestamp": ts,
+            "model": "claude-sonnet-5",
+            "provider": "anthropic",
+            "status": "ok",
+            "duration_ms": 8812,
+            "usage": dict(usage or USAGE),
+            "raw": {"id": response_id, "role": "assistant", "content": []},
+        },
+    }
+
+
+def logging_shape(
+    response_id: str = RESPONSE_ID, ts: str = TIMESTAMP, usage: dict | None = None
+) -> dict:
+    """foundation hooks-logging: `ts` on top, metadata hoisted out of `data`."""
+    return {
+        "event": "llm:response",
+        "ts": ts,
+        "lvl": "INFO",
+        "status": "ok",
+        "duration_ms": 8812,
+        "session_id": SESSION,
+        "schema": {"name": "amplifier.log", "ver": "1.0.0"},
+        "data": {
+            "model": "claude-sonnet-5",
+            "provider": "anthropic",
+            "usage": dict(usage or USAGE),
+            "raw": {"id": response_id, "role": "assistant", "content": []},
+        },
+    }
+
+
+def write_events(path: Path, events: list[dict]) -> str:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text("".join(json.dumps(e) + "\n" for e in events))
+    return str(path)
+
+
+def test_two_loggers_one_call_counts_once(tmp_path):
+    """The exact production bug: same call, two files, two envelope shapes."""
+    a = write_events(tmp_path / "s" / "context-intelligence" / "events.jsonl", [ci_shape()])
+    b = write_events(tmp_path / "s" / "events.jsonl", [logging_shape()])
+
+    parsed = parse_events([a, b])
+
+    assert parsed["llm_responses"] == 1
+    assert parsed["duplicate_responses"] == 1
+    assert parsed["output_tokens"] == 398
+    assert parsed["cost_usd"] == pytest.approx(0.1132975)
+
+
+def test_the_two_shapes_share_no_bytes(tmp_path):
+    """Guards the reason identity is required: equality cannot find these.
+
+    If this ever fails, the loggers have converged and a cheaper de-duplication
+    would work -- but until then, byte comparison provably cannot catch the
+    duplicate, which is why `_response_identity` exists.
+    """
+    assert json.dumps(ci_shape()) != json.dumps(logging_shape())
+    assert set(ci_shape()) != set(logging_shape())
+
+
+def test_distinct_calls_are_not_collapsed(tmp_path):
+    """The fix must not over-collapse: understating cost is the worse failure."""
+    events = [
+        ci_shape(response_id="msg_A", ts="2026-08-05T15:23:15.000000001+00:00"),
+        ci_shape(response_id="msg_B", ts="2026-08-05T15:23:16.000000002+00:00"),
+        ci_shape(response_id="msg_C", ts="2026-08-05T15:23:17.000000003+00:00"),
+    ]
+    path = write_events(tmp_path / "events.jsonl", events)
+
+    parsed = parse_events([path])
+
+    assert parsed["llm_responses"] == 3
+    assert parsed["duplicate_responses"] == 0
+    assert parsed["output_tokens"] == 398 * 3
+
+
+def test_dedup_without_raw_capture_uses_timestamp_fingerprint(tmp_path):
+    """Raw capture is opt-in; de-duplication must still hold when it is off."""
+
+    def strip_raw(event: dict) -> dict:
+        event["data"].pop("raw", None)
+        return event
+
+    a = write_events(tmp_path / "a" / "events.jsonl", [strip_raw(ci_shape())])
+    b = write_events(tmp_path / "b" / "events.jsonl", [strip_raw(logging_shape())])
+
+    parsed = parse_events([a, b])
+
+    assert parsed["llm_responses"] == 1
+    assert parsed["duplicate_responses"] == 1
+
+
+def test_distinct_calls_without_raw_still_distinct(tmp_path):
+    """Fingerprint must discriminate on timestamp, not collapse on equal usage."""
+    first = ci_shape(ts="2026-08-05T15:23:15.000000001+00:00")
+    second = ci_shape(ts="2026-08-05T15:23:16.000000002+00:00")
+    for e in (first, second):
+        e["data"].pop("raw")
+    path = write_events(tmp_path / "events.jsonl", [first, second])
+
+    parsed = parse_events([path])
+
+    assert parsed["llm_responses"] == 2
+    assert parsed["duplicate_responses"] == 0
+
+
+def test_unidentifiable_event_is_counted_not_dropped(tmp_path):
+    """No id and no timestamp: count it, flag it, never silently discard it.
+
+    Dropping here would understate cost invisibly. Counting may overstate, which
+    is the failure mode that gets noticed and investigated.
+    """
+    event = {"event": "llm:response", "data": {"usage": {"output_tokens": 10}}}
+    path = write_events(tmp_path / "events.jsonl", [event, dict(event)])
+
+    parsed = parse_events([path])
+
+    assert parsed["llm_responses"] == 2
+    assert parsed["unidentified_responses"] == 2
+    assert parsed["duplicate_responses"] == 0
+
+
+def test_notes_state_the_correction(tmp_path):
+    """A silent correction is one nobody can audit. It must appear in notes."""
+    a = write_events(tmp_path / "a" / "events.jsonl", [ci_shape()])
+    b = write_events(tmp_path / "b" / "events.jsonl", [logging_shape()])
+
+    record = normalize_metrics([a, b], total_wallclock_s=1.0, source="test")
+
+    assert record["llm_responses"] == 1
+    assert "Dropped 1 duplicate" in record["notes"]
+    assert "more than one logging hook" in record["notes"]
+
+
+def test_wallclock_unaffected_by_duplicates(tmp_path):
+    """Duplicates share timestamps, so the span must not change."""
+    early, late = "2026-08-05T15:23:15.000000+00:00", "2026-08-05T15:23:45.000000+00:00"
+    single = write_events(
+        tmp_path / "one" / "events.jsonl",
+        [ci_shape(response_id="a", ts=early), ci_shape(response_id="b", ts=late)],
+    )
+    dupe = write_events(
+        tmp_path / "two" / "events.jsonl",
+        [logging_shape(response_id="a", ts=early), logging_shape(response_id="b", ts=late)],
+    )
+
+    one_logger = parse_events([single])
+    two_loggers = parse_events([single, dupe])
+
+    assert two_loggers["llm_responses"] == one_logger["llm_responses"] == 2
+    assert two_loggers["agent_wallclock_s"] == pytest.approx(one_logger["agent_wallclock_s"])
+    assert two_loggers["agent_wallclock_s"] == pytest.approx(30.0)
diff --git a/.amplifier/evaluation/uv.lock b/.amplifier/evaluation/uv.lock
index 176a1e15..113a804e 100644
--- a/.amplifier/evaluation/uv.lock
+++ b/.amplifier/evaluation/uv.lock
@@ -76,6 +76,12 @@ dependencies = [
     { name = "pyyaml" },
 ]
 
+[package.dev-dependencies]
+dev = [
+    { name = "pytest" },
+    { name = "pytest-asyncio" },
+]
+
 [package.metadata]
 requires-dist = [
     { name = "amplifier-bundle-evaluation", git = "https://github.com/microsoft/amplifier-bundle-evaluation?rev=ec9f4ca3a23e74a519fced2fe0e2a9ec39e5764d" },
@@ -85,6 +91,12 @@ requires-dist = [
     { name = "pyyaml", specifier = ">=6.0" },
 ]
 
+[package.metadata.requires-dev]
+dev = [
+    { name = "pytest", specifier = ">=8.0" },
+    { name = "pytest-asyncio", specifier = ">=0.23" },
+]
+
 [[package]]
 name = "duckdb"
 version = "1.5.4"
@@ -121,6 +133,33 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/dc/82/4d52f3f9f9703a226b26b80bdae3f6905aeefe5221bf1815fc93ff02ca25/duckdb-1.5.4-cp314-cp314-win_arm64.whl", hash = "sha256:0f8722346024e5d9f02b58bf7b0491a629f97fdc8a04a10e432940f471ee387a", size = 14449863, upload-time = "2026-06-17T10:48:50.18Z" },
 ]
 
+[[package]]
+name = "iniconfig"
+version = "2.3.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/72/34/14ca021ce8e5dfedc35312d08ba8bf51fdd999c576889fc2c24cb97f4f10/iniconfig-2.3.0.tar.gz", hash = "sha256:c76315c77db068650d49c5b56314774a7804df16fee4402c1f19d6d15d8c4730", size = 20503, upload-time = "2025-10-18T21:55:43.219Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/cb/b1/3846dd7f199d53cb17f49cba7e651e9ce294d8497c8c150530ed11865bb8/iniconfig-2.3.0-py3-none-any.whl", hash = "sha256:f631c04d2c48c52b84d0d0549c99ff3859c98df65b3101406327ecc7d53fbf12", size = 7484, upload-time = "2025-10-18T21:55:41.639Z" },
+]
+
+[[package]]
+name = "packaging"
+version = "26.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/7d/fa/3944b40b07da9ce895c0e6303a5ab7d53da063554f534556b134a54d6093/packaging-26.3.tar.gz", hash = "sha256:94edc256424af38762eb31306eed28beb9f0efc50a8837492c9d6fd6004aed79", size = 313412, upload-time = "2026-08-04T18:15:28.737Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/63/34/ba1c580383c9eada3711951fef0795c80b829a078d72188184bcab9dd527/packaging-26.3-py3-none-any.whl", hash = "sha256:d7193f7c8e4e93f444fde0262bf90af30e16fa0ad0ad44cb553c87339b23cd1c", size = 129956, upload-time = "2026-08-04T18:15:27.159Z" },
+]
+
+[[package]]
+name = "pluggy"
+version = "1.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/f9/e2/3e91f31a7d2b083fe6ef3fa267035b518369d9511ffab804f839851d2779/pluggy-1.6.0.tar.gz", hash = "sha256:7dcc130b76258d33b90f61b658791dede3486c3e6bfb003ee5c9bfb396dd22f3", size = 69412, upload-time = "2025-05-15T12:30:07.975Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/54/20/4d324d65cc6d9205fabedc306948156824eb9f0ee1633355a8f7ec5c66bf/pluggy-1.6.0-py3-none-any.whl", hash = "sha256:e920276dd6813095e9377c0bc5566d94c932c33b27a3e3945d8389c374dd4746", size = 20538, upload-time = "2025-05-15T12:30:06.134Z" },
+]
+
 [[package]]
 name = "pydantic"
 version = "2.13.4"
@@ -238,6 +277,44 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/4b/2d/69abac8f838090bbecd5df894befb2c2619e7996a98ddb949db9f3b93225/pydantic_core-2.46.4-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:d51026d73fcfd93610abc7b27789c26b313920fcfb20e27462d74a7f8b06e983", size = 2193071, upload-time = "2026-05-06T13:38:08.682Z" },
 ]
 
+[[package]]
+name = "pygments"
+version = "2.20.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/c3/b2/bc9c9196916376152d655522fdcebac55e66de6603a76a02bca1b6414f6c/pygments-2.20.0.tar.gz", hash = "sha256:6757cd03768053ff99f3039c1a36d6c0aa0b263438fcab17520b30a303a82b5f", size = 4955991, upload-time = "2026-03-29T13:29:33.898Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f4/7e/a72dd26f3b0f4f2bf1dd8923c85f7ceb43172af56d63c7383eb62b332364/pygments-2.20.0-py3-none-any.whl", hash = "sha256:81a9e26dd42fd28a23a2d169d86d7ac03b46e2f8b59ed4698fb4785f946d0176", size = 1231151, upload-time = "2026-03-29T13:29:30.038Z" },
+]
+
+[[package]]
+name = "pytest"
+version = "9.1.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "colorama", marker = "sys_platform == 'win32'" },
+    { name = "iniconfig" },
+    { name = "packaging" },
+    { name = "pluggy" },
+    { name = "pygments" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e4/47/b9efed96c114afcfa3c9d3fe98a76a1d14c74a9e266d397cf6eb64be5e01/pytest-9.1.1.tar.gz", hash = "sha256:1088fbde8f2b49d95a549a195707afa7a76a3ce9bcadc26b6d71f0ffda5fe313", size = 1636369, upload-time = "2026-06-19T10:58:32.857Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/24/25/1de2678b631f5a49215c6c96fff41ba892b0a34df68d6d80292b1b48aa7f/pytest-9.1.1-py3-none-any.whl", hash = "sha256:37a86b45efb9a47a61a36449063e8e18d0cab3161329fc099eb21783169c4f0c", size = 386536, upload-time = "2026-06-19T10:58:31.347Z" },
+]
+
+[[package]]
+name = "pytest-asyncio"
+version = "1.4.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pytest" },
+    { name = "typing-extensions", marker = "python_full_version < '3.13'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/43/7c/d36d04db312ecf4298932ef77e6e4a9e8ad017906e24e34f0b0c361a2473/pytest_asyncio-1.4.0.tar.gz", hash = "sha256:c6c0d2259945122819f171a32ecea2c349ead889ee28176caaf492143424be42", size = 58514, upload-time = "2026-05-26T09:56:04.083Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/03/e2/08a497ef684b88559c9cc5f4ad53a37e7b99e727094a86d6ea32536d5d3c/pytest_asyncio-1.4.0-py3-none-any.whl", hash = "sha256:933ca923a23075a87fb7070c0ec272a6848489824d887c85c812670932835aa1", size = 16930, upload-time = "2026-05-26T09:56:02.576Z" },
+]
+
 [[package]]
 name = "pyyaml"
 version = "6.0.3"