feat: D16 adjudication — 2v2 stays hypothesis until a rule fires.

temporal_freshness then authority_pairing; unresolved keeps (not confirmed).
bin/facts/audit contradict. Gitea #29.

.github/workflows/ci.yml

@@ -45,10 +45,10 @@ jobs:
         run: |
           uv run python -m unittest discover -s bin/tools -t .
 
-      - name: Go tests (root module, no ladybug cgo)
+      - name: Go tests (root module; duckdb-go CGO via gcc, no ladybug)
         run: |
-          go vet ./...
+          CC=gcc CXX=g++ CGO_CFLAGS= CGO_LDFLAGS= go vet ./...
-          go test ./... -count=1
+          CC=gcc CXX=g++ CGO_CFLAGS= CGO_LDFLAGS= go test ./... -count=1
 
       - name: brain ranking tests (no cgo / no ladybug)
         run: go test ./internal/brain/rank -count=1

.gitignore

@@ -12,3 +12,5 @@ __pycache__/
 lib-ladybug/
 go.work.local
 models/
+# Purged from git history. Do not re-add.
+docs/crm-associations-proof.md

AGENTS.md

@@ -48,7 +48,8 @@ bin/postgres/ query.go (read-only YAML)
 bin/git/      import.go (go-git history; Python shim execs it)
 bin/web/      search.go (SearXNG; Python shim execs it)
 bin/reasoner/ bakeoff.go (D18 CPU OpenAI tool-call bake-off)
-internal/     shared Go (brain/rank is cgo-free; chats parsers; gitlog; websearch; reasoner)
+internal/     shared Go (brain/rank is cgo-free; facts D16; chats; gitlog; websearch; reasoner; duckstats)
+bin/qa/       stats.go (DuckDB quantiles / JSONL count; gcc CGO, not Zig)
 bin/watch/    corpus watcher (used by bin/brain/watch.go)
 bin/tools/    vendored python libs behind bin/* (kblib, yamlout, websearch)
 bin/cgo/      zig zcc zc++ (CGO via zig cc, not gcc)
@@ -83,7 +84,7 @@ bin/brain/index.go --rebuild --with-facts --with-chats
 ## Tools
 
 ```bash
-bin/facts/audit.go ["self"|"facts"|"info"|"stale"]  # 2-source + staleness gate
+bin/facts/audit.go ["self"|"db"|"contradict"]     # 2-source + D16 adjudication
 bin/facts/crm.go [--dry-run]                       # proof person↔company/company↔project (ooCRM × corpus SoT)
 bin/kb/search "query" [--repo X]                  # deprecated wrapper → bin/brain/search.go
 bin/brain/search.go "query" [--root facts|info]   # deduction search → YAML
@@ -101,13 +102,16 @@ bin/git/import.go [REPO] [--json] [--limit N]     # go-git history → commit le
 bin/web/search.go "query" [--json]                # SearXNG; throttled ≠ absence
 bin/reasoner/bakeoff.go [--model ID] [--json]     # D18 CPU tool-call bake-off
 bin/postgres/query.go --profile onlyoffice -c 'SELECT 1'
+bin/qa/stats.go                                  # D22 DuckDB quantiles / JSONL (gcc CGO)
 bin/mail/ocr.go <image|pdf>                      # tesseract eng+deu (scans)
 bin/md/tables                                     # what the graph holds → YAML
 bin/brain/deduce "question"                       # thinking wrapper
 ```
 
 Never start a shell command with `cd` — use the tool working-directory
-parameter. Search before reading whole files.
+parameter. Search before reading whole files. For YAML/JSON/XML/CSV/TOML/HCL
+prefer mikefarah/yq (`skills/yq/SKILL.md`). For bulk rows and quantiles use
+duckdb-go (`internal/duckstats`, `skills/duckdb/SKILL.md`), not Ladybug.
 
 ## GitHub safety rules (ABSOLUTE — never violate)
 

PLAN.md

@@ -5,8 +5,10 @@ RAG over the operational Brain/ops/eSlider stack. Built like Sherlock
 Holmes: nothing is asserted unless it has proof.
 
 Status: **v1 in** (epic [#16](https://git.produktor.io/eSlider/2dph/issues/16) closed).
-v2 board: milestone [v2](https://git.produktor.io/eSlider/2dph/milestone/13) — OCR [#6](https://git.produktor.io/eSlider/2dph/issues/6),
+v2 board: milestone [v2](https://git.produktor.io/eSlider/2dph/milestone/13) —
-[#29](https://git.produktor.io/eSlider/2dph/issues/29) OQ1, [#30](https://git.produktor.io/eSlider/2dph/issues/30) OQ3.
+OCR [#6](https://git.produktor.io/eSlider/2dph/issues/6) in,
+[#29](https://git.produktor.io/eSlider/2dph/issues/29) OQ1 in,
+[#30](https://git.produktor.io/eSlider/2dph/issues/30) OQ3 in.
 Gap: [docs/roadmap.md](docs/roadmap.md).
 
 ## What
@@ -42,12 +44,13 @@ detective method: **a fact needs ≥2 independent sources or it is
 | D13 | portfolio | start graph `(Person:eslider)-[:HAS]->(Portfolio)`, associate other natural/juristic persons later. |
 | D14 | tooling style | `bin/{subject}/{method}.go` shebang (e.g. `bin/brain/search.go`). Shared code in `internal/`. One root `go.mod` + `go.work`. No `bin/*/main.go`, no nested modules. |
 | D15 | repo | Gitea [`eSlider/2dph`](https://git.produktor.io/eSlider/2dph) is origin + [issues](https://git.produktor.io/eSlider/2dph/issues). GitHub `eSlider/2dph` is the public clone (PRs + Actions CI). No direct `main` pushes. TDD → PR → CI green → merge. |
-| D16 | contradictions | ≥2 yes vs ≥2 no → unrelated sources conflict → hypothesis → `(not confirmed)`. Resolution (authority, staleness adjudication) = **v2**, tracked as open question. |
+| D16 | contradictions | ≥2 yes vs ≥2 no → hypothesis → `(not confirmed)` until a rule fires. Order: **temporal_freshness** (fresh ≥2 vs stale minority), then **authority_pairing** (runtime/config A×B beats narrative C). Store as `a x b vs c x d` on hypothesis leafs. `bin/facts/audit contradict`. [#29](https://git.produktor.io/eSlider/2dph/issues/29). |
 | D17 | assertion gate | Fact-check every *claim* (facts → info → live → web), not every edit. `bin/brain/search.go` adds a `web` block when there is no facts hit (`throttled`/`skipped`/`refused` ≠ absence). `--root` and `--no-web` stay local. Missing graph ≠ “does not exist”. |
 | D18 | reasoner | Pluggable OpenAI-compatible URL (`REASONER_BASE_URL`). RAM: `Qwen/Qwen3.5-9B`. Quality: `prism-ml/Bonsai-27B-gguf` or `Qwen/Qwen3.6-27B`. No official Qwen3.6-9B. CPU bake-off: `bin/reasoner/bakeoff.go` + compose profile `reasoner` (`OLLAMA_NUM_GPU=0`, `:11435`). PicoClaw is compose profile `picoclaw`; tools are `search`/`get`/`audit`. Weights are not copied into the 2dph image. Agent lever/loop: [#15](https://git.produktor.io/eSlider/2dph/issues/15). |
 | D19 | git history | [go-git](https://github.com/go-git/go-git) via `bin/git/import.go`. No subprocess of the git binary. Conversion prints commit leafs; brain write is `bin/brain/index.go`. |
 | D20 | agent API | OpenAPI + MCP are generated from the same `internal/httpapi.Ops` table as `bin/brain/serve.go` handlers. `GET /openapi.json`, `POST /mcp` (JSON-RPC tools/list + tools/call). Tool names match OpenAPI paths (`search`/`get`/`stats`/`audit`/`ingest`). |
 | D21 | CGO | Ladybug/tokenizers CGO is compiled with **Zig** (`bin/cgo/zcc` → `zig cc -target …-linux-gnu`), not gcc. `bin/cgo/zig` pins Zig 0.14.1 + liblbug 0.19.1 + libtokenizers 1.27.0. Compose `target: api` has no CPython; write/rebuild is profile `index`. |
+| D22 | analytics | **duckdb-go** in-process (`internal/duckstats`, `bin/qa/stats.go`) for quantiles/JSONL. Links with **gcc/g++**, not Zig. Ladybug stays the graph; web-search cache stays modernc sqlite. Slice small structured docs with **mikefarah/yq**, not kislyuk/jq. [#30](https://git.produktor.io/eSlider/2dph/issues/30). |
 
 ## Architecture
 
@@ -110,20 +113,20 @@ Common props on every node/edge: `root`, `confidence`, `evidence[]`, `how`,
 
 - `bin/{subject}/{method}` — line 2 is a usage comment (mirrors `psql-yq`).
 - bash + python primary; golang via Go shebang when a compiled helper is right.
-- YAML default output, `--json` for machines. Slice with `yq`.
+- YAML default output, `--json` for machines. Slice with mikefarah/yq.
 - Everything that touches the network / DB is read-only, throttled, cached.
 - Tests (TDD) gate every commit; `gh` + CI/CD on every push.
 
 ## Open questions (v2)
 
-- OQ1: mutually-contradicting evidence — how to resolve (authority weighting,
+- OQ1: **in** — D16 adjudication: `temporal_freshness` then `authority_pairing`.
-  temporal freshness, audit adjudication). **v2**; [#29](https://git.produktor.io/eSlider/2dph/issues/29).
+  Unresolved 2v2 stays hypothesis. [#29](https://git.produktor.io/eSlider/2dph/issues/29).
 - OQ2: OCR — **in**. `pdftotext -layout` first; scans `pdftoppm` + tesseract
   `eng+deu` (`bin/mail/ocr.go`, `internal/ocr`). No gocv, no gosseract CGO
   (D21 Zig owns Ladybug CGO). Optional `OCR_ENGINE=paddle` / compose profile
   `ocr-paddle`. Docling left the default path. [#6](https://git.produktor.io/eSlider/2dph/issues/6).
-- OQ3: optional duckdb-md layer for `SELECT … FORMAT MARKDOWN` export/write-back.
+- OQ3: **in** — duckdb-go (`internal/duckstats`, `bin/qa/stats.go`) for
-  [#30](https://git.produktor.io/eSlider/2dph/issues/30).
+  quantiles / JSONL count. Not a second graph. [#30](https://git.produktor.io/eSlider/2dph/issues/30).
 - OQ4: YAML-first storage for leafs — deferred: JSON is ~10x faster to
   serialize and unambiguous; YAML only where humans edit files.
 
@@ -183,4 +186,4 @@ Narrative: [docs/roadmap.md](docs/roadmap.md).
 | 4 | [#15](https://git.produktor.io/eSlider/2dph/issues/15) | **in** — lever/loop documented (`search` → `get` → `audit`). |
 | 5 | [#19](https://git.produktor.io/eSlider/2dph/issues/19) | **in** — CI recall SoT is `bin/brain/eval.go` via Zig. Python `bin/kb/eval` stays as an explicit fallback. |
 
-Does **not** block epic close: OQ1 [#29](https://git.produktor.io/eSlider/2dph/issues/29), OQ3 [#30](https://git.produktor.io/eSlider/2dph/issues/30), OQ4. OCR [#6](https://git.produktor.io/eSlider/2dph/issues/6) is **in**.
+Does **not** block epic close: OQ4. OCR [#6](https://git.produktor.io/eSlider/2dph/issues/6), OQ1 [#29](https://git.produktor.io/eSlider/2dph/issues/29), OQ3 [#30](https://git.produktor.io/eSlider/2dph/issues/30) are **in**.

bin/facts/audit

@@ -1,14 +1,15 @@
 #!/usr/bin/env python3
 """facts/audit - evidence & lexicon checks for the 2dph brain.
 
-  bin/facts/audit self      # lexicon: every fact in db has >=2 sources
+  bin/facts/audit self         # lexicon: docs + two-source rule
-  bin/facts/audit db        # evidence gate: run against var/kb.lbug
+  bin/facts/audit db           # evidence gate against var/kb.lbug
+  bin/facts/audit contradict   # D16 adjudication (JSON claim(s) on stdin)
 
-`self` mode checks the repo itself (no network, no runtime deps). It greps
+`self` mode checks the repo itself (no network, no runtime deps).
-for known-good two-source pairings and confirms the docs are consistent.
+`db` mode loads every Leaf with root=facts. Confirmed facts need ` x `;
-`db` mode loads every Leaf with root=facts and asserts each has source_rev
+hypothesis contradictions need `a x b vs c x d` (both sides ≥2).
-and a non-empty `loc` (the "where did you see it" evidence pointer) and that
+`contradict` applies temporal_freshness then authority_pairing; ≥2 vs ≥2
-'confirmed' facts carry a two-source `source` field.
+with no rule stays hypothesis / `(not confirmed)`.
 
 Exit 0 = all checks pass, 1 = audit failures, 2 = could not evaluate.
 """
@@ -22,6 +23,8 @@ from pathlib import Path
 ROOT = Path(__file__).resolve().parents[2]
 sys.path.insert(0, str(ROOT / "bin" / "tools"))
 
+from contradict import adjudicate, check_fact_row  # noqa: E402
+
 
 def audit_db() -> list[str]:
     from kblib import connect
@@ -33,14 +36,8 @@ def audit_db() -> list[str]:
     r = conn.execute("MATCH (l:Leaf {root:'facts'}) RETURN l.id, l.source, l.loc, l.how, l.confidence")
     problems: list[str] = []
     for lid, source, loc, how, conf in r.get_all():
-        if conf != "confirmed":
+        problems.extend(check_fact_row(str(lid), str(source or ""), str(loc or ""),
-            problems.append(f"{lid}: facts require confidence='confirmed', got '{conf}'")
+                                       str(how or ""), str(conf or "")))
-        if not source or " x " not in source:
-            problems.append(f"{lid}: needs 2-source evidence in source, got '{source}'")
-        if not loc:
-            problems.append(f"{lid}: missing loc (evidence pointer)")
-        if not how:
-            problems.append(f"{lid}: missing how")
     conn.close()
     db.close()
     return problems
@@ -54,20 +51,50 @@ def audit_self() -> list[str]:
         problems.append("PLAN.md missing recall@5 gate")
     if re.search(r"(?i)facts must have.*2 sources|2.source", plan) is None:
         problems.append("PLAN.md missing the two-source evidence rule for facts")
+    if "temporal_freshness" not in plan or "authority_pairing" not in plan:
+        problems.append("PLAN.md missing D16 adjudication rules")
     if re.search(r"(?i)HNSW|BM25|deduction", (ROOT / "README.md").read_text()) is None:
         problems.append("README.md missing search/retrieval description")
     return problems
 
 
+def audit_contradict(raw: str) -> tuple[list[str], list[dict]]:
+    raw = raw.strip()
+    if not raw:
+        return ["contradict: empty stdin (JSON claim or {claims:[...]})"], []
+    try:
+        payload = json.loads(raw)
+    except json.JSONDecodeError as e:
+        return [f"contradict: invalid JSON: {e}"], []
+    if isinstance(payload, dict) and "claims" in payload:
+        claims = list(payload.get("claims") or [])
+    elif isinstance(payload, dict):
+        claims = [payload]
+    elif isinstance(payload, list):
+        claims = payload
+    else:
+        return ["contradict: expected object or list"], []
+    details = [adjudicate(c) for c in claims]
+    return [], details
+
+
 def main(argv: list[str]) -> int:
     import argparse
     p = argparse.ArgumentParser(description="evidence & lexicon audit")
-    p.add_argument("mode", choices=("self", "db"))
+    p.add_argument("mode", choices=("self", "db", "contradict"))
     p.add_argument("--json", action="store_true")
     a = p.parse_args(argv)
 
-    problems = audit_self() if a.mode == "self" else audit_db()
+    details: list[dict] = []
-    out = {"mode": a.mode, "ok": not problems, "problems": problems}
+    if a.mode == "self":
+        problems = audit_self()
+    elif a.mode == "db":
+        problems = audit_db()
+    else:
+        problems, details = audit_contradict(sys.stdin.read())
+    out: dict = {"mode": a.mode, "ok": not problems, "problems": problems}
+    if details:
+        out["contradictions"] = details
     if a.json:
         print(json.dumps(out, indent=2))
     else:

bin/facts/audit.go

@@ -5,6 +5,7 @@
 //
 //	./bin/facts/audit.go self
 //	./bin/facts/audit.go db
+//	./bin/facts/audit.go contradict --json < claim.json
 //
 // Python bin/facts/audit is the implementation (CI runs it directly).
 // NOTE: never run `gofmt -w` on this file — it breaks the shebang.

bin/qa/stats.go

@@ -0,0 +1,73 @@
+//usr/bin/env go run -tags=qa_stats "$0" "$@"; exit
+//go:build qa_stats
+//
+// bin/qa/stats.go - DuckDB quantiles over a JSON number array or JSONL count.
+//
+//	./bin/qa/stats.go <<< '[1,2,3,4,5]'
+//	./bin/qa/stats.go --jsonl rows.jsonl
+//
+// NOTE: never run `gofmt -w` on this file — it breaks the shebang.
+// DuckDB CGO needs gcc/g++ (not Zig). After eval "$(bin/cgo/zig env)":
+//   CC=gcc CXX=g++ CGO_CFLAGS= CGO_LDFLAGS= ./bin/qa/stats.go
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"os"
+	"strings"
+
+	"github.com/eSlider/2dph/internal/duckstats"
+)
+
+func main() {
+	os.Exit(run(os.Args[1:]))
+}
+
+func run(args []string) int {
+	jsonl := ""
+	for i := 0; i < len(args); i++ {
+		a := args[i]
+		switch {
+		case a == "--jsonl" && i+1 < len(args):
+			i++
+			jsonl = args[i]
+		case strings.HasPrefix(a, "--jsonl="):
+			jsonl = strings.TrimPrefix(a, "--jsonl=")
+		case a == "-h" || a == "--help":
+			fmt.Fprintln(os.Stderr, "bin/qa/stats.go [--jsonl FILE]  # stdin = JSON [float,…]")
+			return 0
+		default:
+			fmt.Fprintln(os.Stderr, "unknown arg:", a)
+			return 2
+		}
+	}
+	if jsonl != "" {
+		n, err := duckstats.CountJSONL(jsonl)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			return 1
+		}
+		fmt.Printf("n: %d\n", n)
+		return 0
+	}
+	raw, err := io.ReadAll(os.Stdin)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		return 1
+	}
+	var samples []float64
+	if err := json.Unmarshal(raw, &samples); err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		return 1
+	}
+	s, err := duckstats.Quantiles(samples)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		return 1
+	}
+	fmt.Printf("n: %d\nmin: %g\np50: %g\np95: %g\nmax: %g\navg: %g\n",
+		s.N, s.Min, s.P50, s.P95, s.Max, s.Avg)
+	return 0
+}

bin/reasoner/bakeoff.go

@@ -7,7 +7,7 @@
 //	./bin/reasoner/bakeoff.go --model MichelRosselli/bonsai-27b:Q1_0 --json
 //
 // Measures OpenAI tool_calls (search/get/audit) and RSS from Ollama /api/ps, not VRAM.
-// PicoClaw is not in this repo; the tool names match internal/httpapi MCP ops.
+// PicoClaw is compose profile picoclaw; tool names match internal/httpapi MCP ops.
 // NOTE: never run `gofmt -w` on this file — it breaks the shebang.
 package main
 
@@ -17,6 +17,7 @@ import (
 	"os"
 	"strings"
 
+	"github.com/eSlider/2dph/internal/duckstats"
 	"github.com/eSlider/2dph/internal/reasoner"
 )
 
@@ -61,6 +62,14 @@ func run(args []string) int {
 	}
 	c := reasoner.Client{BaseURL: base, Model: model, Device: device}
 	rep := reasoner.Run(c)
+	lat := make([]float64, 0, len(rep.Prompts))
+	for _, p := range rep.Prompts {
+		lat = append(lat, float64(p.LatencyMS))
+	}
+	if st, err := duckstats.Quantiles(lat); err == nil {
+		rep.LatencyP50MS = st.P50
+		rep.LatencyP95MS = st.P95
+	}
 	raw, err := json.MarshalIndent(rep, "", "  ")
 	if err != nil {
 		fmt.Fprintln(os.Stderr, err)
@@ -76,6 +85,8 @@ func run(args []string) int {
 		fmt.Printf("xml_leak: %d\n", rep.XMLLeak)
 		fmt.Printf("rss_mb: %d\n", rep.RSSMB)
 		fmt.Printf("vram_mb: %d\n", rep.VRAMMB)
+		fmt.Printf("latency_p50_ms: %g\n", rep.LatencyP50MS)
+		fmt.Printf("latency_p95_ms: %g\n", rep.LatencyP95MS)
 		for _, p := range rep.Prompts {
 			status := "fail"
 			if p.OK {

bin/tools/contradict.py

@@ -0,0 +1,103 @@
+"""D16 contradiction adjudication (same rules as internal/facts)."""
+from __future__ import annotations
+
+from typing import Any
+
+CONF_CONFIRMED = "confirmed"
+CONF_HYPOTHESIS = "hypothesis"
+
+RULE_UNRESOLVED = "unresolved"
+RULE_TEMPORAL = "temporal_freshness"
+RULE_AUTHORITY = "authority_pairing"
+RULE_TWO_SOURCE = "two_source"
+RULE_SINGLE = "single_source"
+
+KIND_RUNTIME = "runtime"
+KIND_CONFIG = "config"
+KIND_NARRATIVE = "narrative"
+
+
+def _independent(sources: list[dict]) -> int:
+    seen: set[str] = set()
+    for i, s in enumerate(sources):
+        sid = str(s.get("id") or "") or f"{s.get('kind', '')}#{i}"
+        seen.add(sid)
+    return len(seen)
+
+
+def _fresh_n(sources: list[dict]) -> int:
+    return sum(1 for s in sources if not s.get("stale"))
+
+
+def _strong_n(sources: list[dict]) -> int:
+    return sum(1 for s in sources if s.get("kind") in (KIND_RUNTIME, KIND_CONFIG))
+
+
+def adjudicate(claim: dict[str, Any]) -> dict[str, Any]:
+    yes = list(claim.get("yes") or [])
+    no = list(claim.get("no") or [])
+    yes_n, no_n = _independent(yes), _independent(no)
+    text = str(claim.get("text") or "")
+
+    def out(conf: str, rule: str, winner: str = "") -> dict[str, Any]:
+        return {
+            "text": text,
+            "confidence": conf,
+            "confirmed": conf == CONF_CONFIRMED,
+            "rule": rule,
+            "winner": winner,
+            "yes": yes_n,
+            "no": no_n,
+        }
+
+    if yes_n < 2 or no_n < 2:
+        if yes_n >= 2:
+            return out(CONF_CONFIRMED, RULE_TWO_SOURCE, "yes")
+        if no_n >= 2:
+            return out(CONF_CONFIRMED, RULE_TWO_SOURCE, "no")
+        return out(CONF_HYPOTHESIS, RULE_SINGLE)
+    yf, nf = _fresh_n(yes), _fresh_n(no)
+    if yf >= 2 and nf < 2:
+        return out(CONF_CONFIRMED, RULE_TEMPORAL, "yes")
+    if nf >= 2 and yf < 2:
+        return out(CONF_CONFIRMED, RULE_TEMPORAL, "no")
+    ys, ns = _strong_n(yes), _strong_n(no)
+    if ys >= 2 and ns < 2:
+        return out(CONF_CONFIRMED, RULE_AUTHORITY, "yes")
+    if ns >= 2 and ys < 2:
+        return out(CONF_CONFIRMED, RULE_AUTHORITY, "no")
+    return out(CONF_HYPOTHESIS, RULE_UNRESOLVED)
+
+
+def parse_source_field(source: str) -> tuple[str, str]:
+    """Split `a x b vs c x d` into (yes, no). Empty no if no ` vs `."""
+    if " vs " not in source:
+        return source, ""
+    yes, _, no = source.partition(" vs ")
+    return yes.strip(), no.strip()
+
+
+def check_fact_row(lid: str, source: str, loc: str, how: str, conf: str) -> list[str]:
+    """Lexicon checks for one facts leaf (no Ladybug)."""
+    problems: list[str] = []
+    src = source or ""
+    if conf == CONF_CONFIRMED:
+        if " vs " in src:
+            problems.append(f"{lid}: confirmed fact cannot keep a vs-contradiction")
+        if " x " not in src:
+            problems.append(f"{lid}: needs 2-source evidence in source, got '{source}'")
+    elif conf == CONF_HYPOTHESIS:
+        yes, no = parse_source_field(src)
+        if not no or " x " not in yes or " x " not in no:
+            problems.append(
+                f"{lid}: hypothesis contradiction needs 'a x b vs c x d', got '{source}'"
+            )
+    elif conf == "partial":
+        pass
+    else:
+        problems.append(f"{lid}: unknown confidence '{conf}'")
+    if not loc:
+        problems.append(f"{lid}: missing loc (evidence pointer)")
+    if not how:
+        problems.append(f"{lid}: missing how")
+    return problems

bin/tools/test_bin_layout.py

@@ -127,6 +127,22 @@ class BinLayoutTest(unittest.TestCase):
             self.assertIn("cmdbin.ExecFile", text)
             self.assertIn(f"bin/facts/{method.removesuffix('.go')}", text)
 
+    def test_d16_adjudication_is_cgo_free(self) -> None:
+        self.assertTrue((ROOT / "internal" / "facts" / "contradict.go").is_file())
+        go = (ROOT / "internal" / "facts" / "contradict.go").read_text()
+        py = (ROOT / "bin" / "tools" / "contradict.py").read_text()
+        audit = (ROOT / "bin" / "facts" / "audit").read_text()
+        for token in ("temporal_freshness", "authority_pairing", "unresolved"):
+            self.assertIn(token, go)
+            self.assertIn(token, py)
+        self.assertIn("contradict", audit)
+        self.assertIn(" vs ", py)
+        plan = (ROOT / "PLAN.md").read_text()
+        self.assertIn("temporal_freshness", plan)
+        self.assertIn("authority_pairing", plan)
+        shebang = (ROOT / "bin" / "facts" / "audit.go").read_text()
+        self.assertIn("contradict", shebang)
+
     def test_mail_import_is_shebang_not_brain_write(self) -> None:
         self._assert_shebang("bin/mail/import.go")
         index_mail = (ROOT / "bin" / "mail" / "index_mail").read_text()
@@ -217,6 +233,32 @@ class BinLayoutTest(unittest.TestCase):
             if "go-git/go-git" in line:
                 self.assertNotIn("indirect", line)
 
+    def test_duckdb_go_is_direct_require(self) -> None:
+        text = (ROOT / "go.mod").read_text()
+        first = text.split("require (")[1].split(")")[0]
+        self.assertRegex(first, r"github.com/duckdb/duckdb-go/v2\s+v")
+        for line in first.splitlines():
+            if "duckdb/duckdb-go" in line:
+                self.assertNotIn("indirect", line)
+        skill = (ROOT / "skills" / "duckdb" / "SKILL.md").read_text()
+        self.assertIn("github.com/duckdb/duckdb-go", skill)
+        self.assertIn("Ladybug", skill)
+        self.assertIn("sqlite", skill.lower())
+        self.assertIn("gcc", skill.lower())
+        self.assertIn("Zig", skill)
+        plan = (ROOT / "PLAN.md").read_text()
+        self.assertIn("D22", plan)
+        self.assertIn("duckdb-go", plan)
+        self._assert_shebang("bin/qa/stats.go")
+        reasoner = (ROOT / "internal" / "reasoner" / "client.go").read_text()
+        self.assertNotIn("duckdb", reasoner)
+        self.assertNotIn("duckstats", reasoner)
+        bakeoff = (ROOT / "bin" / "reasoner" / "bakeoff.go").read_text()
+        self.assertIn("internal/duckstats", bakeoff)
+        webcache = (ROOT / "internal" / "websearch" / "cache.go").read_text()
+        self.assertNotIn("duckdb", webcache)
+        self.assertIn("modernc.org/sqlite", webcache)
+
     def test_cgo_uses_zig_not_gcc(self) -> None:
         for rel in ("bin/cgo/zig", "bin/cgo/zcc", "bin/cgo/zc++"):
             p = ROOT / rel

bin/tools/test_contradict.py

@@ -0,0 +1,104 @@
+import os
+import sys
+import unittest
+
+sys.path.insert(0, os.path.dirname(__file__))
+
+from contradict import (  # noqa: E402
+    RULE_AUTHORITY,
+    RULE_SINGLE,
+    RULE_TEMPORAL,
+    RULE_TWO_SOURCE,
+    RULE_UNRESOLVED,
+    adjudicate,
+    check_fact_row,
+    parse_source_field,
+)
+
+
+def src(i, kind, stale=False):
+    return {"id": i, "kind": kind, "stale": stale}
+
+
+class TestContradict(unittest.TestCase):
+    def test_two_vs_two_stays_hypothesis(self):
+        r = adjudicate({
+            "text": "svc listens on 443",
+            "yes": [src("docker-ps", "runtime"), src("compose", "config")],
+            "no": [src("docker-old", "runtime"), src("compose-old", "config")],
+        })
+        self.assertFalse(r["confirmed"])
+        self.assertEqual(r["rule"], RULE_UNRESOLVED)
+        self.assertEqual(r["winner"], "")
+
+    def test_temporal_freshness(self):
+        r = adjudicate({
+            "text": "svc listens on 443",
+            "yes": [src("docker-ps", "runtime"), src("compose", "config")],
+            "no": [src("old-readme", "narrative", True), src("old-wiki", "narrative", True)],
+        })
+        self.assertTrue(r["confirmed"])
+        self.assertEqual(r["rule"], RULE_TEMPORAL)
+        self.assertEqual(r["winner"], "yes")
+
+    def test_authority_pairing(self):
+        r = adjudicate({
+            "text": "svc listens on 443",
+            "yes": [src("docker-ps", "runtime"), src("compose", "config")],
+            "no": [src("readme", "narrative"), src("wiki", "narrative")],
+        })
+        self.assertTrue(r["confirmed"])
+        self.assertEqual(r["rule"], RULE_AUTHORITY)
+        self.assertEqual(r["winner"], "yes")
+
+    def test_two_source_and_single(self):
+        two = adjudicate({
+            "text": "arc-1 runs Matrix",
+            "yes": [src("compose", "config"), src("docker-ps", "runtime")],
+        })
+        self.assertTrue(two["confirmed"])
+        self.assertEqual(two["rule"], RULE_TWO_SOURCE)
+        one = adjudicate({"text": "maybe", "yes": [src("readme", "narrative")]})
+        self.assertFalse(one["confirmed"])
+        self.assertEqual(one["rule"], RULE_SINGLE)
+
+    def test_parse_source_field(self):
+        yes, no = parse_source_field("docker ps x compose.yml vs old.md x wiki.md")
+        self.assertIn(" x ", yes)
+        self.assertIn(" x ", no)
+
+    def test_check_fact_row_allows_hypothesis_vs(self):
+        p = check_fact_row(
+            "L1", "a.md x b.md vs c.md x d.md", "var/", "audit", "hypothesis",
+        )
+        self.assertEqual(p, [])
+        p = check_fact_row("L2", "a.md x b.md", "var/", "audit", "confirmed")
+        self.assertEqual(p, [])
+        p = check_fact_row("L3", "a.md x b.md vs c.md x d.md", "var/", "audit", "confirmed")
+        self.assertTrue(any("vs-contradiction" in x for x in p))
+        p = check_fact_row("L4", "only-one.md", "var/", "audit", "hypothesis")
+        self.assertTrue(any("a x b vs" in x for x in p))
+
+    def test_audit_contradict_cli_unresolved(self):
+        import json
+        import subprocess
+        from pathlib import Path
+        root = Path(__file__).resolve().parents[2]
+        payload = json.dumps({
+            "text": "svc 443",
+            "yes": [src("a", "runtime"), src("b", "config")],
+            "no": [src("c", "runtime"), src("d", "config")],
+        })
+        proc = subprocess.run(
+            [sys.executable, str(root / "bin" / "facts" / "audit"), "contradict", "--json"],
+            input=payload, capture_output=True, text=True, check=False,
+        )
+        self.assertEqual(proc.returncode, 0, proc.stderr)
+        out = json.loads(proc.stdout)
+        self.assertTrue(out["ok"])
+        self.assertEqual(out["contradictions"][0]["rule"], RULE_UNRESOLVED)
+        self.assertFalse(out["contradictions"][0]["confirmed"])
+
+
+if __name__ == "__main__":
+    unittest.main()

bin/tools/test_published_docs.py

@@ -121,7 +121,7 @@ class PublishedDocsTest(unittest.TestCase):
         self.assertIn("D18", plan)
         self.assertIn("Qwen/Qwen3.5-9B", plan)
         compose = (ROOT / "compose.yaml").read_text()
-        self.assertIn('profiles: ["reasoner"]', compose)
+        self.assertIn('"reasoner"', compose)
         self.assertIn("OLLAMA_NUM_GPU", compose)
         self.assertIn("127.0.0.1:11435", compose)
         dockerfile = (ROOT / "Dockerfile").read_text()

bin/tools/test_skills.py

@@ -47,3 +47,17 @@ class SkillsTest(unittest.TestCase):
         self.assertIn("throttled", skill.lower())
         self.assertIn("not a negative finding", agents)
         self.assertIn("Fact-check every", agents)
+
+    def test_yq_is_mikefarah_for_structured_data(self) -> None:
+        skill = (ROOT / "skills" / "yq" / "SKILL.md").read_text()
+        self.assertIn("https://github.com/mikefarah/yq", skill)
+        for fmt in ("YAML", "JSON", "XML", "CSV", "TOML", "HCL"):
+            self.assertIn(fmt, skill)
+        self.assertIn("not kislyuk", skill.lower())
+        plan = (ROOT / "PLAN.md").read_text()
+        self.assertIn("mikefarah/yq", plan)
+        agents = (ROOT / "AGENTS.md").read_text()
+        self.assertIn("mikefarah/yq", agents)
+        web = (ROOT / "skills" / "web-search" / "SKILL.md").read_text()
+        self.assertIn("| yq ", web)
+        self.assertNotIn("| jq ", web)

bin/tools/test_system_perf.py

@@ -0,0 +1,36 @@
+"""qa/system_perf.py is an offline-gated system test (no live brain in CI)."""
+from __future__ import annotations
+
+import ast
+import unittest
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+
+
+class SystemPerfScriptTest(unittest.TestCase):
+    def test_script_compiles_and_is_read_only(self) -> None:
+        path = ROOT / "qa" / "system_perf.py"
+        src = path.read_text()
+        compile(src, str(path), "exec")
+        self.assertIn("--json", src)
+        self.assertIn("qwen3.5:9b", src)
+        self.assertIn("--picoclaw", src)
+        self.assertIn("BRAIN_URL", src)
+        self.assertIn("tools/list", src)
+        self.assertIn("tools/call", src)
+        self.assertIn("GATE_HEALTH_MS", src)
+        self.assertIn("GATE_GET_P50_MS", src)
+        self.assertNotIn("kb.lbug", src)
+        self.assertNotIn("password", src.lower())
+        self.assertNotIn("token", src.lower())
+
+    def test_script_does_not_write_ladybug(self) -> None:
+        tree = ast.parse((ROOT / "qa" / "system_perf.py").read_text())
+        writes = [
+            n.func.attr
+            for n in ast.walk(tree)
+            if isinstance(n, ast.Call) and isinstance(n.func, ast.Attribute)
+            and n.func.attr in {"write_text", "write_bytes", "dump"}
+        ]
+        self.assertEqual(writes, [], f"system_perf must not write files: {writes}")

compose.yaml

@@ -2,7 +2,7 @@
 #
 #   docker compose up -d brain                  # API (Zig CGO serve)
 #   docker compose --profile index run --rm index   # Python rebuild
-#   docker compose --profile picoclaw up brain-mcp
+#   docker compose --profile picoclaw up -d     # brain-mcp + CPU reasoner + PicoClaw gateway
 #   docker compose --profile reasoner up -d reasoner   # CPU Ollama :11435
 #   docker compose --profile searxng up -d
 #   OCR_ENGINE=paddle docker compose --profile ocr-paddle run --rm ocr-paddle
@@ -11,6 +11,14 @@
 
 name: 2dph
 
+networks:
+  default:
+    name: 2dph_sys
+    driver: bridge
+    ipam:
+      config:
+        - subnet: 10.23.42.0/24
+
 services:
   brain:
     image: ghcr.io/eslider/2dph:api
@@ -98,8 +106,8 @@ services:
       - ./deploy/searxng/limiter.toml:/etc/searxng/limiter.toml:ro
     restart: unless-stopped
 
-  # MCP endpoint for an external agent (PicoClaw is not shipped here).
+  # MCP endpoint for PicoClaw (and any MCP client).
-  #   docker compose --profile picoclaw up brain-mcp
+  #   docker compose --profile picoclaw up -d
   brain-mcp:
     profiles: ["picoclaw"]
     image: ghcr.io/eslider/2dph:api
@@ -121,7 +129,7 @@ services:
   #   docker compose --profile reasoner up -d reasoner
   #   docker compose --profile reasoner exec reasoner ollama pull qwen3.5:9b
   reasoner:
-    profiles: ["reasoner"]
+    profiles: ["reasoner", "picoclaw"]
     image: docker.io/ollama/ollama:latest
     environment:
       OLLAMA_NUM_GPU: "0"
@@ -132,6 +140,26 @@ services:
       - reasoner-ollama:/root/.ollama
     restart: unless-stopped
 
+  # Official PicoClaw gateway. Config has no secrets (Ollama + HTTP MCP).
+  # Host network: brain/reasoner bind 127.0.0.1 only, so host.docker.internal
+  # (docker0) cannot reach them. Gateway 127.0.0.1:18790 (not the 18800 launcher).
+  # If :8630/:11435 are already bound, do not start brain-mcp/reasoner:
+  #   docker compose --profile picoclaw up -d --no-deps picoclaw
+  picoclaw:
+    profiles: ["picoclaw"]
+    image: docker.io/sipeed/picoclaw:v0.3.1
+    network_mode: host
+    depends_on:
+      - brain-mcp
+      - reasoner
+    environment:
+      PICOCLAW_GATEWAY_HOST: "127.0.0.1"
+    entrypoint: ["picoclaw", "gateway"]
+    volumes:
+      - picoclaw-home:/root/.picoclaw
+      - ./deploy/picoclaw/config.json:/root/.picoclaw/config.json:ro
+    restart: unless-stopped
+
   # Optional PP-OCRv5 (not default). Default OCR is tesseract eng+deu.
   #   OCR_ENGINE=paddle docker compose --profile ocr-paddle run --rm ocr-paddle
   ocr-paddle:
@@ -145,3 +173,4 @@ volumes:
   kb-model:
   kb-var:
   reasoner-ollama:
+  picoclaw-home:

deploy/picoclaw/config.json

@@ -0,0 +1,33 @@
+{
+  "agents": {
+    "defaults": {
+      "model_name": "qwen3.5-9b",
+      "max_tool_iterations": 8,
+      "max_tokens": 512,
+      "context_window": 8192
+    }
+  },
+  "model_list": [
+    {
+      "model_name": "qwen3.5-9b",
+      "model": "ollama/qwen3.5:9b",
+      "api_base": "http://127.0.0.1:11435/v1",
+      "request_timeout": 600
+    }
+  ],
+  "tools": {
+    "web": {
+      "enabled": false
+    },
+    "mcp": {
+      "enabled": true,
+      "servers": {
+        "2dph": {
+          "enabled": true,
+          "type": "http",
+          "url": "http://127.0.0.1:8630/mcp"
+        }
+      }
+    }
+  }
+}

docs/README.md

@@ -20,7 +20,7 @@ Evidence-first knowledge graph. Facts need proof or they are
 | explanation | [roadmap](roadmap.md) — gap to v1 (epic #16) |
 | howto | [picoclaw](picoclaw.md) — MCP agent profile |
 | howto | [reasoner](reasoner.md) — CPU bake-off (D18) |
-| reference | [PLAN.md](../PLAN.md) — decisions D1–D21 |
+| reference | [PLAN.md](../PLAN.md) — decisions D1–D22 |
 
 Decisions the public face must name: **D3** SearXNG compose, **D6** Go service /
 Python write sidecar, **D14** `bin/{subject}/{method}.go`, **D15** Gitea origin,

docs/design.md

@@ -71,7 +71,9 @@ corpus HEAD.
 - C: narrative — READMEs, AGENTS.md, docs
 
 Confirmed = A×B or B×C agreement. Single source = hypothesis + `(not confirmed)`.
-Conflicting pairings (≥2 yes vs ≥2 no) = hypothesis (OQ1 → v2 resolution).
+Conflicting pairings (≥2 yes vs ≥2 no) stay hypothesis until
+`temporal_freshness` or `authority_pairing` fires (`bin/facts/audit contradict`,
+[#29](https://git.produktor.io/eSlider/2dph/issues/29)).
 
 ## Read path
 

docs/picoclaw.md

@@ -1,18 +1,34 @@
 # PicoClaw profile (reference agent)
 
-2dph is the memory/fact gate. PicoClaw (or any MCP client) is the agent loop
+2dph is the memory/fact gate. Compose profile `picoclaw` runs the official
-and is **not** shipped in this repo.
+PicoClaw gateway (`docker.io/sipeed/picoclaw:v0.3.1`) plus `brain-mcp` and the
+CPU reasoner. Default agent model is `qwen3.5:9b` (RAM path, D18). Weights stay
+in the reasoner volume, not in the 2dph image.
+No secrets in git: Ollama needs no key; MCP is local HTTP.
 
 ```bash
-docker compose --profile picoclaw up brain-mcp
+docker compose --profile picoclaw up -d
+# already serving :8630 / :11435:
+docker compose --profile picoclaw up -d --no-deps picoclaw
 ```
 
-The API listens on `127.0.0.1:8630`. Point the agent at
+Gateway: `127.0.0.1:18790`. Brain MCP: `http://127.0.0.1:8630/mcp`.
-`http://127.0.0.1:8630/mcp` using [deploy/picoclaw/mcp.json.example](../deploy/picoclaw/mcp.json.example).
+Cursor-style clients can use [deploy/picoclaw/mcp.json.example](../deploy/picoclaw/mcp.json.example).
+PicoClaw itself uses [deploy/picoclaw/config.json](../deploy/picoclaw/config.json)
+(`127.0.0.1` + host network — loopback publishes are not reachable via docker0).
 
 OpenAPI: `GET http://127.0.0.1:8630/openapi.json`.
 
 Before a factual reply: `search` → `get` → `audit`. `throttled` is not a
 negative finding. See `skills/picoclaw/SKILL.md`.
 
-No Cursor required. A live PicoClaw binary/image is an operator choice.
+System performance (MCP gates + qwen3.5:9b tool_call + PicoClaw gateway):
+
+```bash
+./qa/system_perf.py --json | yq '.gates'
+REASONER_MODEL=qwen3.5:9b ./qa/system_perf.py --reasoner --picoclaw --json | yq '.reasoner'
+```
+
+The default agent model is `qwen3.5:9b`. PicoClaw `context_window` is 8192
+(heuristic `max_tokens*4` at 512 is 2048, too small for MCP tool schemas).
+`request_timeout` is 600s for a CPU turn (tool_call + MCP search + answer).

docs/reasoner.md

@@ -1,8 +1,8 @@
 # Reasoner bake-off (D18)
 
 Pluggable OpenAI-compatible URL. 2dph does not ship weights. PicoClaw is
-not in this repo; the bake-off hits the same tool names PicoClaw would
+compose profile `picoclaw` (`sipeed/picoclaw`); the bake-off hits the same
-(`search` → `get` → `audit` from `internal/httpapi.Ops`).
+tool names (`search` → `get` → `audit` from `internal/httpapi.Ops`).
 
 ```bash
 docker compose --profile reasoner up -d reasoner
@@ -11,6 +11,8 @@ REASONER_BASE_URL=http://127.0.0.1:11435/v1 REASONER_MODEL=qwen3.5:9b \
   ./bin/reasoner/bakeoff.go --json
 ```
 
+JSON includes `latency_p50_ms` / `latency_p95_ms` from DuckDB (`internal/duckstats`, D22).
+
 Host Ollama on `:11434` is left alone. This sidecar binds `127.0.0.1:11435`
 with `OLLAMA_NUM_GPU=0` (CPU). Measure RSS (`/api/ps` `size`), not VRAM.
 

docs/roadmap.md

@@ -36,14 +36,14 @@ Epic [#16](https://git.produktor.io/eSlider/2dph/issues/16) closed.
 
 ## v2
 
-[#6](https://git.produktor.io/eSlider/2dph/issues/6) OCR — `pdftotext` then
+[#6](https://git.produktor.io/eSlider/2dph/issues/6) OCR — **in**.
-`pdftoppm` + tesseract `eng+deu`. Optional `ocr-paddle`.
+[#30](https://git.produktor.io/eSlider/2dph/issues/30) OQ3 duckdb-go — **in**.
 [#29](https://git.produktor.io/eSlider/2dph/issues/29) OQ1 contradiction
-resolution. [#30](https://git.produktor.io/eSlider/2dph/issues/30) OQ3 duckdb-md.
+resolution — **in** (`temporal_freshness`, `authority_pairing`).
 
 ## Blockers
 
-None for epic #16 (closed). Remaining v2: OQ1, OQ3, OQ4.
+None for epic #16 (closed). Remaining v2: OQ4.
 
 ```
 question
@@ -58,8 +58,8 @@ question
 
 ## Not v1
 
-OQ1 contradiction resolution, OQ3 duckdb-md export, OQ4 YAML-first leafs.
+OQ4 YAML-first leafs.
-OCR (OQ2) is in: tesseract, not docling.
+OCR (OQ2), duckdb-go (OQ3/D22), and D16 adjudication (OQ1) are in.
 
 ## Close epic #16 when
 

go.mod

@@ -7,6 +7,7 @@ require (
 	github.com/arran4/golang-ical v0.3.5
 	github.com/chewxy/math32 v1.11.2
 	github.com/daulet/tokenizers v1.27.0
+	github.com/duckdb/duckdb-go/v2 v2.10505.0
 	github.com/go-git/go-git/v5 v5.19.2
 	golang.org/x/sys v0.47.0
 	golang.org/x/text v0.40.0
@@ -20,10 +21,17 @@ require (
 	github.com/apache/arrow-go/v18 v18.6.0 // indirect
 	github.com/cloudflare/circl v1.6.3 // indirect
 	github.com/cyphar/filepath-securejoin v0.6.1 // indirect
+	github.com/duckdb/duckdb-go-bindings v0.10505.0 // indirect
+	github.com/duckdb/duckdb-go-bindings/lib/darwin-amd64 v0.10505.0 // indirect
+	github.com/duckdb/duckdb-go-bindings/lib/darwin-arm64 v0.10505.0 // indirect
+	github.com/duckdb/duckdb-go-bindings/lib/linux-amd64 v0.10505.0 // indirect
+	github.com/duckdb/duckdb-go-bindings/lib/linux-arm64 v0.10505.0 // indirect
+	github.com/duckdb/duckdb-go-bindings/lib/windows-amd64 v0.10505.0 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/emirpasic/gods v1.18.1 // indirect
 	github.com/go-git/gcfg v1.5.1-0.20230307220236-3a3c6141e376 // indirect
 	github.com/go-git/go-billy/v5 v5.9.0 // indirect
+	github.com/go-viper/mapstructure/v2 v2.5.0 // indirect
 	github.com/goccy/go-json v0.10.6 // indirect
 	github.com/golang/groupcache v0.0.0-20241129210726-2c02b8208cf8 // indirect
 	github.com/google/flatbuffers v25.12.19+incompatible // indirect

go.sum

@@ -31,6 +31,20 @@ github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSs
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc h1:U9qPSI2PIWSS1VwoXQT9A3Wy9MM3WgvqSxFWenqJduM=
 github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/duckdb/duckdb-go-bindings v0.10505.0 h1:/0pPsTLrcCsTGxT0VrHgJWnOcPe1tQL1vrki1v3jbAI=
+github.com/duckdb/duckdb-go-bindings v0.10505.0/go.mod h1:HoD5xePkDj3VZbBnVVfxVVYIljZ9khCprWA7FgwIiC4=
+github.com/duckdb/duckdb-go-bindings/lib/darwin-amd64 v0.10505.0 h1:FrMqquFBQlMsi34h2KZgCku54rqA8xEbXZ0NLVDKwYs=
+github.com/duckdb/duckdb-go-bindings/lib/darwin-amd64 v0.10505.0/go.mod h1:EnAvZh1kNJHp5yF+M1ZHNEvapnmt6anq1xXHVrAGqMo=
+github.com/duckdb/duckdb-go-bindings/lib/darwin-arm64 v0.10505.0 h1:lbRbpQwT1MmUhh/VTwukV9K8bxKByV3UghAP3MvsbBo=
+github.com/duckdb/duckdb-go-bindings/lib/darwin-arm64 v0.10505.0/go.mod h1:IGLSeEcFhNeZF16aVjQCULD7TsFZKG5G7SyKJAXKp5c=
+github.com/duckdb/duckdb-go-bindings/lib/linux-amd64 v0.10505.0 h1:nrsaVYj3XYCRbS2FpdOMD/KHE7egRMr+/NR1IHmjT84=
+github.com/duckdb/duckdb-go-bindings/lib/linux-amd64 v0.10505.0/go.mod h1:KAIynZ0GHCS7X5fRyuFnQMg/SZBPK/bS9OCOVojClxw=
+github.com/duckdb/duckdb-go-bindings/lib/linux-arm64 v0.10505.0 h1:qM6oGDgwXBILJGbTY4fCy6QOczLpucUA6yn6g3ORjh4=
+github.com/duckdb/duckdb-go-bindings/lib/linux-arm64 v0.10505.0/go.mod h1:81SGOYoEUs8qaAfSk1wRfM5oobrIJ5KI7AzYhK6/bvQ=
+github.com/duckdb/duckdb-go-bindings/lib/windows-amd64 v0.10505.0 h1:DjqZl9rYreHkSOqnqLmkrqH5T8UdQNcxZLJVZzGmXXA=
+github.com/duckdb/duckdb-go-bindings/lib/windows-amd64 v0.10505.0/go.mod h1:K25pJL26ARblGDeuAkrdblFvUen92+CwksLtPEHRqqQ=
+github.com/duckdb/duckdb-go/v2 v2.10505.0 h1:SWwvLn2Qx/RQSnQNupwgIF8VbnJ5A6OQU9lYb/mDETI=
+github.com/duckdb/duckdb-go/v2 v2.10505.0/go.mod h1:m0PW4J4FG9hlFlVdXi6Ds9owpyIDaBdE2jyce00fGcE=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
 github.com/elazarl/goproxy v1.7.2 h1:Y2o6urb7Eule09PjlhQRGNsqRfPmYI3KKQLFpCAV3+o=
@@ -47,6 +61,8 @@ github.com/go-git/go-git-fixtures/v4 v4.3.2-0.20231010084843-55a94097c399 h1:eMj
 github.com/go-git/go-git-fixtures/v4 v4.3.2-0.20231010084843-55a94097c399/go.mod h1:1OCfN199q1Jm3HZlxleg+Dw/mwps2Wbk9frAWm+4FII=
 github.com/go-git/go-git/v5 v5.19.2 h1:wkfn7vOlUBu8ivAWKBWisTiwJK4jYHzTF8Ndv1LyGqY=
 github.com/go-git/go-git/v5 v5.19.2/go.mod h1:QqCBE1EFN5ddFmrliLQ3/ntRCUjZU3EJuwuB/jWEHjk=
+github.com/go-viper/mapstructure/v2 v2.5.0 h1:vM5IJoUAy3d7zRSVtIwQgBj7BiWtMPfmPEgAXnvj1Ro=
+github.com/go-viper/mapstructure/v2 v2.5.0/go.mod h1:oJDH3BJKyqBA2TXFhDsKDGDTlndYOZ6rGS0BRZIxGhM=
 github.com/goccy/go-json v0.10.6 h1:p8HrPJzOakx/mn/bQtjgNjdTcN+/S6FcG2CTtQOrHVU=
 github.com/goccy/go-json v0.10.6/go.mod h1:oq7eo15ShAhp70Anwd5lgX2pLfOS3QCiwU/PULtXL6M=
 github.com/golang/groupcache v0.0.0-20241129210726-2c02b8208cf8 h1:f+oWsMOmNPc8JmEHVZIycC7hBoQxHH9pNKQORJNozsQ=

internal/brain/rank/escalate.go

@@ -19,20 +19,35 @@ type SecondSourceHit struct {
 
 type WebFn func(query string) SecondSource
 
-// ShouldEscalate is true when the default deduction path has no facts hit.
+// ShouldEscalate is true when the default deduction path has no confirmed
+// facts hit. Hypothesis/partial facts are `(not confirmed)` (D16).
 // `--root facts|info` is a single-root ask: do not mix in the web.
 func ShouldEscalate(hits []Hit, rootFilter string) bool {
 	if rootFilter != "" {
 		return false
 	}
 	for _, h := range hits {
-		if h.Root == "facts" {
+		if ConfirmedFact(h) {
 			return false
 		}
 	}
 	return true
 }
 
+// ConfirmedFact is a facts-root hit that is not hypothesis/partial.
+// Empty confidence is treated as confirmed (legacy leafs).
+func ConfirmedFact(h Hit) bool {
+	if h.Root != "facts" {
+		return false
+	}
+	switch h.Confidence {
+	case "hypothesis", "partial":
+		return false
+	default:
+		return true
+	}
+}
+
 // Deduce returns the second-source block, or nil when web must not run.
 func Deduce(hits []Hit, query, rootFilter string, noWeb bool, web WebFn) *SecondSource {
 	if noWeb || web == nil || !ShouldEscalate(hits, rootFilter) {

internal/brain/rank/escalate_test.go

@@ -14,6 +14,16 @@ func TestShouldEscalateWhenNoFacts(t *testing.T) {
 	}
 }
 
+func TestShouldEscalateWhenHypothesisFacts(t *testing.T) {
+	hyp := Hit{ID: "c", Root: "facts", Confidence: "hypothesis", Source: "a x b vs c x d"}
+	if !ShouldEscalate([]Hit{hyp}, "") {
+		t.Fatal("hypothesis facts are (not confirmed); escalate")
+	}
+	if ConfirmedFact(hyp) {
+		t.Fatal("hypothesis is not confirmed")
+	}
+}
+
 func TestShouldNotEscalateWhenFactsConfirm(t *testing.T) {
 	hits := []Hit{h("f", "facts", "docker ps x compose"), h("i", "info", "docs/a.md")}
 	if ShouldEscalate(hits, "") {

internal/brain/rank/query.go

@@ -3,10 +3,10 @@ package rank
 // BM25 ranks best-first, so the top hits are the *highest* scores; cosine
 // distance ranks best-first ascending. Both mirror kblib.py.
 const FTSStmt = "CALL QUERY_FTS_INDEX('Leaf', 'id', $q) " +
-	"RETURN node.id, node.text, node.root, node.source, score ORDER BY score DESC LIMIT $n"
+	"RETURN node.id, node.text, node.root, node.source, score, node.confidence ORDER BY score DESC LIMIT $n"
 
 const VecStmt = "CALL QUERY_VECTOR_INDEX('Leaf', 'Leaf_vec', $q, $n) " +
-	"RETURN node.id, node.text, node.root, node.source, distance ORDER BY distance LIMIT $n"
+	"RETURN node.id, node.text, node.root, node.source, distance, node.confidence ORDER BY distance LIMIT $n"
 
 // HopStmt is the Cypher walk from a search hit. Depth 1 = File, 2 = Commit, 3 = Person.
 func HopStmt(depth int) string {

internal/brain/rank/rank.go

@@ -19,6 +19,7 @@ type Hit struct {
 	ID         string    `json:"id"`
 	Text       string    `json:"text"`
 	Root       string    `json:"root"`
+	Confidence string    `json:"confidence,omitempty"`
 	Source     string    `json:"-"`
 	Score      float64   `json:"score"`
 	Snippet    string    `json:"snippet,omitempty"`

internal/brain/rank/rank_test.go

@@ -172,4 +172,7 @@ func TestFTSQueryOrdersByScoreDescending(t *testing.T) {
 	if !strings.Contains(FTSStmt, "ORDER BY score DESC") {
 		t.Fatalf("FTS query must order by score DESC, got:\n%s", FTSStmt)
 	}
+	if !strings.Contains(FTSStmt, "node.confidence") {
+		t.Fatal("FTS must return confidence for D16")
+	}
 }

internal/brain/search.go

@@ -212,7 +212,11 @@ func rowsToHits(res *lbug.QueryResult) ([]Hit, error) {
 		root := fmt.Sprint(vals[2])
 		source := fmt.Sprint(vals[3])
 		score := float64(vals[4].(float64))
-		hits = append(hits, Hit{ID: id, Text: text, Root: root, Source: source, Score: score})
+		conf := ""
+		if len(vals) >= 6 {
+			conf = fmt.Sprint(vals[5])
+		}
+		hits = append(hits, Hit{ID: id, Text: text, Root: root, Source: source, Score: score, Confidence: conf})
 	}
 	return hits, nil
 }
@@ -230,6 +234,7 @@ type jsonHit struct {
 	ID         string         `json:"id"`
 	Text       string         `json:"text"`
 	Root       string         `json:"root"`
+	Confidence string         `json:"confidence,omitempty"`
 	Score      float64        `json:"score"`
 	Snippet    string         `json:"snippet,omitempty"`
 	Hops       []rank.HopNode `json:"hops,omitempty"`
@@ -242,6 +247,7 @@ func toJSONOut(hits []Hit, query, rootFilter string, web *rank.SecondSource) *js
 			ID:         h.ID,
 			Text:       h.Text,
 			Root:       h.Root,
+			Confidence: h.Confidence,
 			Score:      h.Score,
 			Snippet:    h.Snippet,
 			Hops:       h.Hops,
@@ -265,6 +271,9 @@ func resultsToDicts(hits []Hit) []any {
 			{"root", h.Root},
 			{"score", h.Score},
 		}
+		if h.Confidence != "" {
+			d = append(d, KV{"confidence", h.Confidence})
+		}
 		if h.Snippet != "" {
 			d = append(d, KV{"snippet", h.Snippet})
 		}

internal/duckstats/duckstats.go

@@ -0,0 +1,47 @@
+// Package duckstats runs in-process DuckDB for columnar aggregates.
+// Graph facts stay in Ladybug. Web-search KV cache stays modernc sqlite.
+package duckstats
+
+import (
+	"database/sql"
+	"fmt"
+
+	_ "github.com/duckdb/duckdb-go/v2"
+)
+
+type Stats struct {
+	N   int     `json:"n"`
+	Min float64 `json:"min"`
+	P50 float64 `json:"p50"`
+	P95 float64 `json:"p95"`
+	Max float64 `json:"max"`
+	Avg float64 `json:"avg"`
+}
+
+func Quantiles(samples []float64) (Stats, error) {
+	if len(samples) == 0 {
+		return Stats{}, fmt.Errorf("duckstats: empty samples")
+	}
+	db, err := sql.Open("duckdb", "")
+	if err != nil {
+		return Stats{}, err
+	}
+	defer db.Close()
+	var s Stats
+	err = db.QueryRow(`
+SELECT count(v), min(v), quantile_cont(v, 0.5), quantile_cont(v, 0.95), max(v), avg(v)
+FROM (SELECT unnest(?) AS v)`, samples).Scan(
+		&s.N, &s.Min, &s.P50, &s.P95, &s.Max, &s.Avg)
+	return s, err
+}
+
+func CountJSONL(path string) (int64, error) {
+	db, err := sql.Open("duckdb", "")
+	if err != nil {
+		return 0, err
+	}
+	defer db.Close()
+	var n int64
+	err = db.QueryRow(`SELECT count(*) FROM read_json_auto(?)`, path).Scan(&n)
+	return n, err
+}

internal/duckstats/duckstats_test.go

@@ -0,0 +1,51 @@
+package duckstats
+
+import (
+	"os"
+	"testing"
+)
+
+func TestQuantilesEmpty(t *testing.T) {
+	_, err := Quantiles(nil)
+	if err == nil {
+		t.Fatal("empty slice must error")
+	}
+}
+
+func TestQuantilesOdd(t *testing.T) {
+	s, err := Quantiles([]float64{1, 2, 3, 4, 5})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if s.N != 5 {
+		t.Fatalf("n=%d", s.N)
+	}
+	if s.Min != 1 || s.Max != 5 {
+		t.Fatalf("min=%v max=%v", s.Min, s.Max)
+	}
+	if s.P50 != 3 {
+		t.Fatalf("p50=%v want 3", s.P50)
+	}
+	if s.Avg != 3 {
+		t.Fatalf("avg=%v want 3", s.Avg)
+	}
+	if s.P95 < 4.5 || s.P95 > 5 {
+		t.Fatalf("p95=%v want in [4.5,5]", s.P95)
+	}
+}
+
+func TestCountJSONL(t *testing.T) {
+	dir := t.TempDir()
+	p := dir + "/rows.jsonl"
+	body := "{\"ms\":1}\n{\"ms\":2}\n{\"ms\":3}\n"
+	if err := os.WriteFile(p, []byte(body), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	n, err := CountJSONL(p)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if n != 3 {
+		t.Fatalf("count=%d want 3", n)
+	}
+}

internal/facts/contradict.go

@@ -0,0 +1,120 @@
+// Package facts is cgo-free evidence rules (D16 contradictions).
+package facts
+
+import "strconv"
+
+const (
+	ConfConfirmed  = "confirmed"
+	ConfHypothesis = "hypothesis"
+
+	RuleUnresolved        = "unresolved"
+	RuleTemporalFreshness = "temporal_freshness"
+	RuleAuthorityPairing  = "authority_pairing"
+	RuleTwoSource         = "two_source"
+	RuleSingleSource      = "single_source"
+
+	KindRuntime   = "runtime"
+	KindConfig    = "config"
+	KindNarrative = "narrative"
+)
+
+// Source is one independent pointer on a yes or no side.
+type Source struct {
+	ID    string `json:"id"`
+	Kind  string `json:"kind"`
+	When  string `json:"when,omitempty"`
+	Stale bool   `json:"stale,omitempty"`
+}
+
+// Claim is one assertion with yes/no evidence lists.
+type Claim struct {
+	Text string   `json:"text"`
+	Yes  []Source `json:"yes"`
+	No   []Source `json:"no"`
+}
+
+// Result is audit output. Confirmed=false means `(not confirmed)`.
+type Result struct {
+	Text       string `json:"text"`
+	Confidence string `json:"confidence"`
+	Confirmed  bool   `json:"confirmed"`
+	Rule       string `json:"rule"`
+	Winner     string `json:"winner,omitempty"`
+	YesN       int    `json:"yes"`
+	NoN        int    `json:"no"`
+}
+
+func independent(ss []Source) int {
+	seen := map[string]struct{}{}
+	for i, s := range ss {
+		id := s.ID
+		if id == "" {
+			id = s.Kind + "#" + strconv.Itoa(i)
+		}
+		seen[id] = struct{}{}
+	}
+	return len(seen)
+}
+
+func freshN(ss []Source) int {
+	n := 0
+	for _, s := range ss {
+		if !s.Stale {
+			n++
+		}
+	}
+	return n
+}
+
+func strongN(ss []Source) int {
+	n := 0
+	for _, s := range ss {
+		if s.Kind == KindRuntime || s.Kind == KindConfig {
+			n++
+		}
+	}
+	return n
+}
+
+func out(c Claim, conf, rule, winner string) Result {
+	return Result{
+		Text:       c.Text,
+		Confidence: conf,
+		Confirmed:  conf == ConfConfirmed,
+		Rule:       rule,
+		Winner:     winner,
+		YesN:       independent(c.Yes),
+		NoN:        independent(c.No),
+	}
+}
+
+// Adjudicate applies D16: ≥2 yes vs ≥2 no stays hypothesis until a rule fires.
+// Order: temporal_freshness, then authority_pairing (A/B beats narrative C).
+func Adjudicate(c Claim) Result {
+	yesN := independent(c.Yes)
+	noN := independent(c.No)
+	if yesN < 2 || noN < 2 {
+		if yesN >= 2 {
+			return out(c, ConfConfirmed, RuleTwoSource, "yes")
+		}
+		if noN >= 2 {
+			return out(c, ConfConfirmed, RuleTwoSource, "no")
+		}
+		return out(c, ConfHypothesis, RuleSingleSource, "")
+	}
+	yf, nf := freshN(c.Yes), freshN(c.No)
+	if yf >= 2 && nf < 2 {
+		return out(c, ConfConfirmed, RuleTemporalFreshness, "yes")
+	}
+	if nf >= 2 && yf < 2 {
+		return out(c, ConfConfirmed, RuleTemporalFreshness, "no")
+	}
+	ys, ns := strongN(c.Yes), strongN(c.No)
+	if ys >= 2 && ns < 2 {
+		return out(c, ConfConfirmed, RuleAuthorityPairing, "yes")
+	}
+	if ns >= 2 && ys < 2 {
+		return out(c, ConfConfirmed, RuleAuthorityPairing, "no")
+	}
+	return out(c, ConfHypothesis, RuleUnresolved, "")
+}

internal/facts/contradict_test.go

@@ -0,0 +1,86 @@
+package facts
+
+import "testing"
+
+func src(id, kind string, stale bool) Source {
+	return Source{ID: id, Kind: kind, Stale: stale}
+}
+
+func TestTwoVsTwoStaysHypothesis(t *testing.T) {
+	c := Claim{
+		Text: "svc listens on 443",
+		Yes: []Source{
+			src("docker-ps", KindRuntime, false),
+			src("compose", KindConfig, false),
+		},
+		No: []Source{
+			src("docker-ps-old", KindRuntime, false),
+			src("compose-old", KindConfig, false),
+		},
+	}
+	r := Adjudicate(c)
+	if r.Confirmed || r.Confidence != ConfHypothesis || r.Rule != RuleUnresolved {
+		t.Fatalf("2v2 must stay (not confirmed): %+v", r)
+	}
+	if r.Winner != "" {
+		t.Fatalf("unresolved must not name a winner: %+v", r)
+	}
+}
+
+func TestTemporalFreshnessResolvesStaleSide(t *testing.T) {
+	c := Claim{
+		Text: "svc listens on 443",
+		Yes: []Source{
+			src("docker-ps", KindRuntime, false),
+			src("compose", KindConfig, false),
+		},
+		No: []Source{
+			src("old-readme", KindNarrative, true),
+			src("old-wiki", KindNarrative, true),
+		},
+	}
+	r := Adjudicate(c)
+	if !r.Confirmed || r.Rule != RuleTemporalFreshness || r.Winner != "yes" {
+		t.Fatalf("fresh yes vs stale no: %+v", r)
+	}
+}
+
+func TestAuthorityPairingBeatsNarrative(t *testing.T) {
+	c := Claim{
+		Text: "svc listens on 443",
+		Yes: []Source{
+			src("docker-ps", KindRuntime, false),
+			src("compose", KindConfig, false),
+		},
+		No: []Source{
+			src("readme", KindNarrative, false),
+			src("wiki", KindNarrative, false),
+		},
+	}
+	r := Adjudicate(c)
+	if !r.Confirmed || r.Rule != RuleAuthorityPairing || r.Winner != "yes" {
+		t.Fatalf("A×B vs C×C: %+v", r)
+	}
+}
+
+func TestTwoSourceYesIsConfirmed(t *testing.T) {
+	c := Claim{
+		Text: "arc-1 runs Matrix",
+		Yes: []Source{
+			src("compose", KindConfig, false),
+			src("docker-ps", KindRuntime, false),
+		},
+	}
+	r := Adjudicate(c)
+	if !r.Confirmed || r.Rule != RuleTwoSource || r.Winner != "yes" {
+		t.Fatalf("%+v", r)
+	}
+}
+
+func TestSingleSourceIsHypothesis(t *testing.T) {
+	c := Claim{Text: "maybe", Yes: []Source{src("readme", KindNarrative, false)}}
+	r := Adjudicate(c)
+	if r.Confirmed || r.Rule != RuleSingleSource {
+		t.Fatalf("%+v", r)
+	}
+}

internal/reasoner/client.go

@@ -113,6 +113,8 @@ type Report struct {
 	XMLLeak      int      `json:"xml_leak"`
 	RSSMB        int      `json:"rss_mb"`
 	VRAMMB       int      `json:"vram_mb"`
+	LatencyP50MS float64  `json:"latency_p50_ms,omitempty"`
+	LatencyP95MS float64  `json:"latency_p95_ms,omitempty"`
 	Prompts      []Result `json:"prompts"`
 }
 

qa/system_perf.py

@@ -0,0 +1,257 @@
+#!/usr/bin/env python3
+"""System performance test: PicoClaw surface (brain MCP) + optional reasoner.
+
+  BRAIN_URL=http://127.0.0.1:8630 ./qa/system_perf.py --json
+  REASONER_BASE_URL=http://127.0.0.1:11435/v1 REASONER_MODEL=qwen3.5:9b \\
+    ./qa/system_perf.py --reasoner --picoclaw --json
+
+Does not write Ladybug. Search includes web (D17); expect ~10s+ per search.
+Exit 1 if health/get/audit gates fail. Reasoner is measured, not gated.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import statistics
+import sys
+import time
+import urllib.error
+import urllib.request
+from concurrent.futures import ThreadPoolExecutor
+
+DEFAULT_BRAIN = "http://127.0.0.1:8630"
+DEFAULT_REASONER = "http://127.0.0.1:11435/v1"
+DEFAULT_MODEL = "qwen3.5:9b"
+DEFAULT_PICOCLAW = "http://127.0.0.1:18790"
+
+GATE_HEALTH_MS = 500
+GATE_GET_P50_MS = 50
+GATE_AUDIT_P50_MS = 50
+
+
+def _req(url: str, data: bytes | None = None, timeout: float = 90) -> bytes:
+    headers = {"Content-Type": "application/json"} if data is not None else {}
+    req = urllib.request.Request(url, data=data, headers=headers)
+    with urllib.request.urlopen(req, timeout=timeout) as res:
+        return res.read()
+
+
+def timed(fn):
+    t0 = time.perf_counter()
+    out = fn()
+    return (time.perf_counter() - t0) * 1000.0, out
+
+
+def stats(samples: list[float]) -> dict:
+    s = sorted(samples)
+    n = len(s)
+    return {
+        "n": n,
+        "min_ms": round(s[0], 1),
+        "p50_ms": round(s[n // 2], 1),
+        "p95_ms": round(s[min(n - 1, int(n * 0.95))], 1),
+        "max_ms": round(s[-1], 1),
+        "avg_ms": round(statistics.mean(s), 1),
+    }
+
+
+def mcp(brain: str, method: str, params=None, timeout: float = 90) -> dict:
+    payload: dict = {"jsonrpc": "2.0", "id": 1, "method": method}
+    if params is not None:
+        payload["params"] = params
+    raw = _req(brain.rstrip("/") + "/mcp", json.dumps(payload).encode(), timeout=timeout)
+    return json.loads(raw.decode())
+
+
+def mcp_call(brain: str, name: str, arguments: dict, timeout: float = 90) -> tuple[bool, str]:
+    d = mcp(brain, "tools/call", {"name": name, "arguments": arguments}, timeout=timeout)
+    res = d.get("result") or {}
+    text = ((res.get("content") or [{}])[0].get("text") or "")
+    return (not res.get("isError")), text
+
+
+def reasoner_tool_call(base: str, model: str, user: str) -> str:
+    payload = {
+        "model": model,
+        "messages": [
+            {"role": "system", "content": "You are PicoClaw. Always call search before answering."},
+            {"role": "user", "content": user},
+        ],
+        "tools": [
+            {
+                "type": "function",
+                "function": {
+                    "name": "search",
+                    "description": "deduction search",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {"q": {"type": "string"}},
+                        "required": ["q"],
+                    },
+                },
+            }
+        ],
+        "tool_choice": "required",
+    }
+    raw = _req(
+        base.rstrip("/") + "/chat/completions",
+        json.dumps(payload).encode(),
+        timeout=600,
+    )
+    chat = json.loads(raw.decode())
+    tcs = chat["choices"][0]["message"].get("tool_calls") or []
+    if not tcs:
+        return ""
+    return tcs[0]["function"]["name"]
+
+
+def run(args: argparse.Namespace) -> dict:
+    brain = args.brain.rstrip("/")
+    report: dict = {
+        "brain": brain,
+        "device": "cpu",
+        "ok": True,
+        "gates": {},
+        "mcp": {},
+    }
+    ms, _ = timed(lambda: _req(brain + "/health", timeout=5))
+    report["mcp"]["health"] = {"n": 1, "avg_ms": round(ms, 1)}
+    report["gates"]["health"] = ms <= GATE_HEALTH_MS
+    if ms > GATE_HEALTH_MS:
+        report["ok"] = False
+
+    list_ms = []
+    for _ in range(args.n):
+        ms, d = timed(lambda: mcp(brain, "tools/list", timeout=10))
+        names = [t["name"] for t in ((d.get("result") or {}).get("tools") or [])]
+        if "search" not in names:
+            report["ok"] = False
+        list_ms.append(ms)
+    report["mcp"]["tools_list"] = stats(list_ms)
+
+    audit_ms = []
+    for _ in range(args.n):
+        ms, (ok, _) = timed(lambda: mcp_call(brain, "audit", {}))
+        if not ok:
+            report["ok"] = False
+        audit_ms.append(ms)
+    report["mcp"]["audit"] = stats(audit_ms)
+    report["gates"]["audit_p50"] = report["mcp"]["audit"]["p50_ms"] <= GATE_AUDIT_P50_MS
+    if not report["gates"]["audit_p50"]:
+        report["ok"] = False
+
+    ok, text = mcp_call(brain, "search", {"q": "LadybugDB", "n": 2}, timeout=90)
+    inner = json.loads(text) if ok else {}
+    hits = inner.get("results") or []
+    leaf_id = hits[0]["id"] if hits else ""
+    report["mcp"]["search_seed"] = {
+        "ok": ok,
+        "count": inner.get("count"),
+        "web": (inner.get("web") or {}).get("status"),
+    }
+
+    get_ms = []
+    if leaf_id:
+        for _ in range(args.n):
+            ms, (ok, _) = timed(lambda: mcp_call(brain, "get", {"id": leaf_id, "body": True}))
+            if not ok:
+                report["ok"] = False
+            get_ms.append(ms)
+        report["mcp"]["get"] = stats(get_ms)
+        report["gates"]["get_p50"] = report["mcp"]["get"]["p50_ms"] <= GATE_GET_P50_MS
+        if not report["gates"]["get_p50"]:
+            report["ok"] = False
+
+        def one_get() -> float:
+            t0 = time.perf_counter()
+            mcp_call(brain, "get", {"id": leaf_id, "body": True})
+            return (time.perf_counter() - t0) * 1000.0
+
+        t0 = time.perf_counter()
+        with ThreadPoolExecutor(max_workers=8) as ex:
+            conc = list(ex.map(lambda _: one_get(), range(8)))
+        wall = (time.perf_counter() - t0) * 1000.0
+        report["mcp"]["get_concurrent_8"] = {**stats(conc), "wall_ms": round(wall, 1)}
+
+    search_ms = []
+    for q in ("LadybugDB", "model2vec"):
+        ms, (ok, text) = timed(lambda q=q: mcp_call(brain, "search", {"q": q, "n": 3}, timeout=90))
+        inner = json.loads(text) if ok else {}
+        search_ms.append(ms)
+        report.setdefault("mcp", {}).setdefault("search_samples", []).append(
+            {
+                "q": q,
+                "ms": round(ms, 1),
+                "ok": ok,
+                "count": inner.get("count"),
+                "web": (inner.get("web") or {}).get("status"),
+            }
+        )
+    if search_ms:
+        report["mcp"]["search"] = stats(search_ms)
+
+    if args.reasoner:
+        base = args.reasoner_url
+        model = args.model
+        report["reasoner"] = {"base_url": base, "model": model, "calls": []}
+        for user in (
+            "Use tools. Search the 2dph brain for LadybugDB. Call search.",
+            "Use tools. Search the 2dph brain for model2vec. Call search.",
+        ):
+            ms, name = timed(lambda user=user: reasoner_tool_call(base, model, user))
+            report["reasoner"]["calls"].append({"ms": round(ms, 1), "tool": name})
+        tools = [c["tool"] for c in report["reasoner"]["calls"]]
+        report["gates"]["reasoner_tool_call"] = bool(tools) and all(t == "search" for t in tools)
+        if not report["gates"]["reasoner_tool_call"]:
+            report["ok"] = False
+
+    if args.picoclaw:
+        gw = args.picoclaw_url.rstrip("/")
+        ms, raw = timed(lambda: _req(gw + "/health", timeout=5))
+        body = json.loads(raw.decode())
+        report["picoclaw"] = {
+            "url": gw,
+            "health_ms": round(ms, 1),
+            "status": body.get("status"),
+        }
+        report["gates"]["picoclaw_health"] = body.get("status") == "ok" and ms <= GATE_HEALTH_MS
+        if not report["gates"]["picoclaw_health"]:
+            report["ok"] = False
+    return report
+
+
+def main(argv: list[str]) -> int:
+    p = argparse.ArgumentParser(description="2dph system performance (MCP + optional reasoner)")
+    p.add_argument("--brain", default=os.environ.get("BRAIN_URL", DEFAULT_BRAIN))
+    p.add_argument("--n", type=int, default=20)
+    p.add_argument("--json", action="store_true")
+    p.add_argument("--reasoner", action="store_true")
+    p.add_argument("--picoclaw", action="store_true")
+    p.add_argument("--picoclaw-url", default=os.environ.get("PICOCLAW_URL", DEFAULT_PICOCLAW))
+    p.add_argument("--reasoner-url", default=os.environ.get("REASONER_BASE_URL", DEFAULT_REASONER))
+    p.add_argument("--model", default=os.environ.get("REASONER_MODEL", DEFAULT_MODEL))
+    args = p.parse_args(argv)
+    try:
+        report = run(args)
+    except (urllib.error.URLError, TimeoutError, OSError) as e:
+        print(f"system_perf: {e}", file=sys.stderr)
+        return 1
+    if args.json:
+        print(json.dumps(report, indent=2))
+    else:
+        print(f"ok={report['ok']} brain={report['brain']}")
+        for name, block in report.get("mcp", {}).items():
+            if isinstance(block, dict) and "p50_ms" in block:
+                print(f"  {name}: p50={block['p50_ms']} p95={block['p95_ms']} n={block['n']}")
+            elif name == "health":
+                print(f"  health: {block.get('avg_ms')} ms")
+        for k, v in report.get("gates", {}).items():
+            print(f"  gate {k}: {v}")
+        for c in (report.get("reasoner") or {}).get("calls") or []:
+            print(f"  reasoner {c['tool']}: {c['ms']} ms")
+    return 0 if report["ok"] else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main(sys.argv[1:]))

skills/brain/SKILL.md

@@ -45,6 +45,8 @@ bin/brain/eval.go                                      # recall@5 >= 0.95 gate (
   are not evidence of absence. `--root facts|info` and `--no-web` skip the web.
 - If recall looks wrong, run `bin/brain/eval.go`; it gates control questions and
   should stay at or above 95% recall@5.
+- Contradictions (≥2 yes vs ≥2 no) stay `(not confirmed)` until
+  `bin/facts/audit contradict` fires `temporal_freshness` or `authority_pairing`.
 - Agents: `GET /openapi.json` and `POST /mcp` on `bin/brain/serve.go` (same
   handlers; tool names match paths `search`/`get`/`stats`/`audit`). Generated
   list: [tools.md](tools.md).

skills/duckdb/SKILL.md

@@ -0,0 +1,38 @@
+---
+name: duckdb
+description: >-
+  Use https://github.com/duckdb/duckdb-go in-process for columnar analytics
+  (quantiles, GROUP BY, JSON/CSV/Parquet/JSONL scans) when that is faster than
+  nested Go loops. Not Ladybug. Not the web-search sqlite cache. Use when
+  aggregating samples, counting JSONL, or SQL over tabular files.
+---
+
+# duckdb-go
+
+Use https://github.com/duckdb/duckdb-go where it makes sense to get better performance in code.
+
+In-process DuckDB (`internal/duckstats`, `database/sql` driver `duckdb`).
+Vectorized SQL over tables, JSONL, CSV, Parquet. CGO with bundled libs
+(linux/darwin amd64/arm64). Links with **gcc/g++** (libstdc++), not Zig.
+D21 Zig (`bin/cgo/zcc`) is Ladybug/tokenizers only. After
+`eval "$(bin/cgo/zig env)"`:
+
+```bash
+CC=gcc CXX=g++ CGO_CFLAGS= CGO_LDFLAGS= ./bin/qa/stats.go <<< '[1,2,3,4,5]'
+CC=gcc CXX=g++ CGO_CFLAGS= CGO_LDFLAGS= go test ./internal/duckstats
+```
+
+| Store | Job |
+|-------|-----|
+| Ladybug | graph + FTS + HNSW (facts/info) |
+| modernc sqlite | web-search KV cache + throttle |
+| duckdb-go | OLAP: quantiles, counts, scans of many rows/files |
+| mikefarah/yq | small YAML/JSON/XML/CSV/TOML/HCL slice, not bulk |
+
+```bash
+./bin/qa/stats.go <<< '[1,2,3,4,5]'
+./bin/qa/stats.go --jsonl path/to/rows.jsonl
+```
+
+Do not open Ladybug through DuckDB. Do not put secrets or client PII into
+DuckDB files under the repo.

skills/picoclaw/SKILL.md

@@ -1,15 +1,15 @@
 ---
 name: picoclaw
 description: >-
-  2dph is the memory/fact gate, not the agent loop. Use when wiring PicoClaw
+  2dph is the memory/fact gate. Compose runs the official PicoClaw gateway.
-  or any MCP client: call brain search/get/audit before a factual reply.
+  Use when wiring PicoClaw or any MCP client: call brain search/get/audit
-  throttled is not a negative finding.
+  before a factual reply. throttled is not a negative finding.
 ---
 
 # PicoClaw — fact-check before assert
 
-PicoClaw (or any agent) speaks MCP at `POST /mcp` on `bin/brain/serve.go`.
+PicoClaw speaks MCP at `POST /mcp` on `bin/brain/serve.go`. Compose profile
-2dph does not run the agent loop. Compose: `docker compose --profile picoclaw up brain-mcp`
+`picoclaw` runs the official `sipeed/picoclaw` gateway plus `brain-mcp`
 (see [docs/picoclaw.md](../../docs/picoclaw.md)).
 
 ## Tool order (before a factual reply)

skills/postgres/SKILL.md

@@ -9,7 +9,7 @@ description: >-
 # postgres
 
 `bin/postgres/query.go` wraps vendored `bin/db/psql-yq`. Output is YAML
-(cheaper than psql ASCII, easy to slice with `yq`).
+(cheaper than psql ASCII, easy to slice with mikefarah/yq).
 
 ```bash
 bin/postgres/query.go --profile onlyoffice -s document_asset   # column list

skills/web-search/SKILL.md

@@ -11,7 +11,7 @@ description: >-
 ```bash
 bin/web/search.go "LadybugDB vector index"
 bin/web/search.go "model2vec multilingual" --category it
-bin/web/search.go "hypervisor" --site example.com --json | jq -r '.results[].url'
+bin/web/search.go "hypervisor" --site example.com --json | yq -r '.results[].url'
 bin/web/search.go "postgres partial index" --lang en --fresh year
 ```
 

skills/web-search/reference/instance-tuning.md

@@ -43,7 +43,7 @@ found" unless the client refuses to call it absence.
 
 ```bash
 for i in $(seq 10); do
-  bin/web/search.go "test $i" -n 1 --refresh --json | jq -r .status
+  bin/web/search.go "test $i" -n 1 --refresh --json | yq -r '.status'
 done
 ```
 

skills/yq/SKILL.md

@@ -0,0 +1,33 @@
+---
+name: yq
+description: >-
+  Use https://github.com/mikefarah/yq to work with YAML, JSON, XML, CSV,
+  TOML, HCL where it's efficient and less code. Use when slicing compose,
+  config, --json tool output, CSV/TOML/HCL/XML, or converting between those
+  formats. Not kislyuk Python yq. Not jq when yq already does the job.
+---
+
+# yq (mikefarah)
+
+Use https://github.com/mikefarah/yq to work with YAML, JSON, XML, CSV, TOML, HCL where it's efficient and less code.
+
+This is the Go `yq` (`yq --version` contains `mikefarah`). It is not
+kislyuk/yq (Python, jq-syntax, YAML-only wrapper). `bin/db/psql-yq` already
+calls this binary.
+
+Prefer `yq` over `python3 -c`, `jq`, or ad-hoc parsers when one expression
+reads or converts the file. Keep Python/Go for HTTP, binary protocols, and
+in-process tests.
+
+```bash
+yq '.services.picoclaw.image' compose.yaml
+yq -P . deploy/picoclaw/config.json          # JSON → YAML
+yq -o=json '.gates'                           # JSON stdin (qa/system_perf.py --json)
+yq -p=csv -o=json .
+yq -p=xml -o=json .
+yq -p=toml '.package.name' file.toml
+bin/brain/search.go "LadybugDB" --json | yq '.[].ref'
+bin/web/search.go "hypervisor" --json | yq -r '.results[].url'
+```
+
+Do not print secrets, PII, or `$HOME/.config/brain/` through `yq`.