+
+
+J7Scope · Contrast Lab
+Compare matched traces, inspect one concept family at a time,
+ calibrate the observation against a paired permutation null, check answer-identity
+ confounds, and optionally run one causal activation patch on a connected GPU sidecar.
+
+
+
+
+
+
+ All calibration values are baked into align.json by
+ j7scope.contrast. The browser renders them but does not recompute
+ sharedness. A green result is evidence above the configured null, not proof of a
+ complete or universal representation.
+
+
+
+
+
+
diff --git a/apps/site/index.html b/apps/site/index.html
index 6a20655..5988557 100644
--- a/apps/site/index.html
+++ b/apps/site/index.html
@@ -72,10 +72,10 @@ J7Scope J-Space Gallery · 双生工作
}
function groupCard(g, members) {
var langs = members.map(function (m) { return m.language; }).join(" ↔ ");
- return '' +
- '' + J.esc(g) + ' side-by-side ' +
+ return ' ' +
+ '' + J.esc(g) + ' Contrast Lab ' +
'' + members.length + ' aligned traces ' + J.esc(langs) + '
' +
- '' + tag("compare zh ↔ en") + (members[0].is_demo ? tag("DEMO", "demo") : "") + '
';
+ '' + tag("paired evidence") + tag("null + ceiling") + (members[0].is_demo ? tag("DEMO", "demo") : "") + '
';
}
})();
diff --git a/apps/web/package-lock.json b/apps/web/package-lock.json
index dc3bd04..4a24563 100644
--- a/apps/web/package-lock.json
+++ b/apps/web/package-lock.json
@@ -3480,9 +3480,9 @@
"peer": true
},
"node_modules/nanoid": {
- "version": "3.3.16",
- "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.16.tgz",
- "integrity": "sha512-bzlKTyNJ7+LdGIIwy8ijFpIqEQIvafahV7eYykJ8Cvh42EdJeODoJ6gUJXpQJvej1BddH8OqTXZNE/KfbWAu8Q==",
+ "version": "3.3.18",
+ "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.18.tgz",
+ "integrity": "sha512-DTg4MJbGMWkfi6VZFdNt2/caMbQy4Ou+Op/hJQvGEWcnVfoA1QA+xzRKAzw9jD6+GVOOeYr/mIcuDSdug6F6+w==",
"dev": true,
"funding": [
{
diff --git a/experiments/build_contrast.py b/experiments/build_contrast.py
new file mode 100644
index 0000000..b0f0c3c
--- /dev/null
+++ b/experiments/build_contrast.py
@@ -0,0 +1,79 @@
+"""Build a Contrast Lab record for two existing Trace v1 directories."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+if str(ROOT) not in sys.path:
+ sys.path.insert(0, str(ROOT))
+
+from j7scope.artifacts import write_json # noqa: E402
+from j7scope.contrast import build_contrast # noqa: E402
+from j7scope.data import load_parallel_pairs # noqa: E402
+from j7scope.rigor import build_lexicon # noqa: E402
+from j7scope.trace import read_trace # noqa: E402
+
+
+def main() -> None:
+ parser = argparse.ArgumentParser()
+ parser.add_argument("trace_a")
+ parser.add_argument("trace_b")
+ parser.add_argument("--trace-root", type=Path, default=ROOT / "results" / "traces")
+ parser.add_argument("--data-dir", type=Path, default=ROOT / "data")
+ parser.add_argument(
+ "--mode",
+ choices=("cross_language", "intervention", "correctness"),
+ default="cross_language",
+ )
+ parser.add_argument("--position-map", type=Path)
+ parser.add_argument("--permutations", type=int, default=1000)
+ parser.add_argument("--bootstrap", type=int, default=1000)
+ parser.add_argument("--seed", type=int, default=0)
+ args = parser.parse_args()
+
+ trace_a = read_trace(args.trace_root / args.trace_a)
+ trace_b = read_trace(args.trace_root / args.trace_b)
+ position_map = None
+ if args.position_map:
+ raw = json.loads(args.position_map.read_text(encoding="utf-8"))
+ position_map = raw.get("position_map", raw) if isinstance(raw, dict) else raw
+
+ pairs = load_parallel_pairs(args.data_dir) if args.data_dir.exists() else None
+ lexicon = build_lexicon(pairs=pairs)
+ contrast = build_contrast(
+ trace_a,
+ trace_b,
+ lexicon,
+ position_map=position_map,
+ mode=args.mode,
+ n_permutations=args.permutations,
+ n_boot=args.bootstrap,
+ seed=args.seed,
+ )
+ align = {
+ "parallel_group": (
+ trace_a["manifest"].get("parallel_group")
+ or trace_b["manifest"].get("parallel_group")
+ or f"{args.trace_a}--{args.trace_b}"
+ ),
+ "members": {"a": args.trace_a, "b": args.trace_b},
+ "position_map": contrast["position_map"],
+ "pair_rigor": contrast["pair_rigor"],
+ "contrast": contrast,
+ }
+ for trace_id in (args.trace_a, args.trace_b):
+ write_json(args.trace_root / trace_id / "align.json", align)
+ print(json.dumps({
+ "trace_a": args.trace_a,
+ "trace_b": args.trace_b,
+ "observed": contrast["calibration"]["observed"],
+ "sharedness": contrast["calibration"]["sharedness"],
+ }, ensure_ascii=False, indent=2))
+
+
+if __name__ == "__main__":
+ main()
diff --git a/experiments/build_demo_trace.py b/experiments/build_demo_trace.py
index 19c657e..18281d1 100644
--- a/experiments/build_demo_trace.py
+++ b/experiments/build_demo_trace.py
@@ -18,7 +18,6 @@
import argparse
import sys
-from datetime import datetime, timezone
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
@@ -26,6 +25,7 @@
sys.path.insert(0, str(ROOT))
from j7scope import rigor
+from j7scope.contrast import build_contrast
from j7scope.artifacts import write_json
from j7scope.data import load_parallel_pairs
from j7scope.trace import write_trace
@@ -40,6 +40,7 @@
}
FILLER_ZH = ["的", "一种", "行为", "这", "属于"]
FILLER_EN = ["a", "kind", "of", "this", "is"]
+DEMO_CREATED_AT = "2026-07-22T00:00:00+00:00"
# A session narrative: (surface_token, concept_or_None). Abstract concepts are
# interspersed with fillers; the concrete entity shows weaker cross-lingual tie.
@@ -144,7 +145,7 @@ def build_trace(narrative, language: str, lexicon, *, trace_id: str,
"sha1": None,
},
"capture": {"tool": "build_demo_trace.py",
- "created_at": datetime.now(timezone.utc).isoformat()},
+ "created_at": DEMO_CREATED_AT},
"is_demo": True,
"doi": None,
"parallel_group": parallel_group,
@@ -187,13 +188,22 @@ def main() -> None:
m_zh, t_zh, mt_zh = build_trace(ZH_NARRATIVE, "zh", lexicon,
trace_id=f"{group}-zh", parallel_group=group)
position_map = [[i, i] for i in range(len(NARRATIVE))] # identical structure
- # Cross-trace rigor: A = en session, B = zh session, at aligned positions.
- pair_rigor = rigor.compute_pair_rigor(t_en, t_zh, position_map, lexicon)
+ # Cross-trace rigor + pair-level Contrast Lab evidence. The aggregate
+ # permutation null and trajectories are baked here; the browser only renders.
+ contrast = build_contrast(
+ {"manifest": m_en, "tokens": t_en, "metrics": mt_en},
+ {"manifest": m_zh, "tokens": t_zh, "metrics": mt_zh},
+ lexicon,
+ position_map=position_map,
+ n_permutations=1000,
+ n_boot=1000,
+ )
align = {
"parallel_group": group,
"members": {"en": m_en["trace_id"], "zh": m_zh["trace_id"]},
"position_map": position_map,
- "pair_rigor": pair_rigor,
+ "pair_rigor": contrast["pair_rigor"],
+ "contrast": contrast,
}
write_trace(args.out / m_en["trace_id"], manifest=m_en, tokens=t_en,
metrics=mt_en, align=align)
diff --git a/j7scope/contrast.py b/j7scope/contrast.py
new file mode 100644
index 0000000..7709225
--- /dev/null
+++ b/j7scope/contrast.py
@@ -0,0 +1,313 @@
+"""Research-grade paired evidence for the J-Space Contrast Lab.
+
+The browser is intentionally a renderer. This module turns two Trace v1
+sessions into a precomputed contrast record containing a paired permutation
+null, a same-language ceiling, a bootstrap interval, concept-rank trajectories,
+an answer-identity confound check, and exact provenance.
+"""
+
+from __future__ import annotations
+
+from typing import Dict, List, Mapping, Optional, Sequence
+
+import numpy as np
+
+from .rigor import (
+ SHAREDNESS_DEFINITION,
+ compute_pair_rigor,
+ concepts_of,
+ estimate_same_lang_baseline,
+ overlap_coef,
+)
+
+CONTRAST_SCHEMA_VERSION = 1
+PAIR_SHAREDNESS_DEFINITION = (
+ "(paired_observed_mean - permutation_null.mean) / "
+ "(same_lang_ceiling - permutation_null.mean)"
+)
+_EPS = 1e-9
+
+
+def _row_concept(row: Mapping, lexicon: Mapping[str, str]) -> Optional[str]:
+ concept = row.get("concept")
+ if isinstance(concept, str) and concept:
+ return concept
+ token = row.get("token")
+ if isinstance(token, str):
+ return lexicon.get(token)
+ return None
+
+
+def token_concepts(token: Mapping, lexicon: Mapping[str, str]) -> List[str]:
+ """Ordered concepts in a token's complete bilingual readout.
+
+ Captured traces can either bake ``row.concept`` into each readout or rely on
+ the corpus lexicon. Supporting both makes imported JSONL traces portable.
+ """
+ out: List[str] = []
+ seen = set()
+ readout = token.get("readout", {})
+ for bucket in ("zh", "en", "other"):
+ for row in readout.get(bucket, []):
+ concept = _row_concept(row, lexicon)
+ if concept and concept not in seen:
+ seen.add(concept)
+ out.append(concept)
+ return out
+
+
+def _rank_for_concept(
+ token: Mapping, concept: str, lexicon: Mapping[str, str]
+) -> Optional[int]:
+ ranks = []
+ for bucket in ("zh", "en", "other"):
+ for fallback_rank, row in enumerate(token.get("readout", {}).get(bucket, [])):
+ if _row_concept(row, lexicon) == concept:
+ ranks.append(int(row.get("rank", fallback_rank)) + 1)
+ return min(ranks) if ranks else None
+
+
+def concept_rank_trajectories(
+ tokens_a: Sequence[Mapping],
+ tokens_b: Sequence[Mapping],
+ position_map: Sequence[Sequence[int]],
+ lexicon: Mapping[str, str],
+) -> List[dict]:
+ concepts = sorted({
+ concept
+ for token in (*tokens_a, *tokens_b)
+ for concept in token_concepts(token, lexicon)
+ })
+ rows = []
+ for concept in concepts:
+ rank_a = [_rank_for_concept(tokens_a[ia], concept, lexicon)
+ for ia, _ in position_map]
+ rank_b = [_rank_for_concept(tokens_b[ib], concept, lexicon)
+ for _, ib in position_map]
+ coverage = sum(rank is not None for rank in rank_a + rank_b)
+ if coverage:
+ rows.append({
+ "concept": concept,
+ "rank_a": rank_a,
+ "rank_b": rank_b,
+ "coverage": round(coverage / max(1, 2 * len(position_map)), 4),
+ })
+ return sorted(rows, key=lambda row: (-row["coverage"], row["concept"]))
+
+
+def _permutation_null(
+ concepts_a: Sequence[Sequence[str]],
+ concepts_b: Sequence[Sequence[str]],
+ *,
+ n_permutations: int,
+ seed: int,
+) -> tuple[float, dict]:
+ observed_values = [overlap_coef(a, b) for a, b in zip(concepts_a, concepts_b)]
+ observed = float(np.mean(observed_values)) if observed_values else 0.0
+ if len(concepts_a) < 2 or n_permutations < 1:
+ return observed, {
+ "metric": "paired_mean_concept_overlap",
+ "mean": 0.0,
+ "p05": 0.0,
+ "p95": 0.0,
+ "n": 0,
+ "seed": seed,
+ }
+
+ rng = np.random.default_rng(seed)
+ values = np.empty(n_permutations, dtype=np.float64)
+ base = np.arange(len(concepts_b))
+ for i in range(n_permutations):
+ permuted = rng.permutation(base)
+ values[i] = np.mean([
+ overlap_coef(concepts_a[j], concepts_b[int(permuted[j])])
+ for j in range(len(concepts_a))
+ ])
+ return observed, {
+ "metric": "paired_mean_concept_overlap",
+ "mean": round(float(values.mean()), 4),
+ "p05": round(float(np.percentile(values, 5)), 4),
+ "p95": round(float(np.percentile(values, 95)), 4),
+ "n": int(n_permutations),
+ "seed": int(seed),
+ }
+
+
+def _paired_bootstrap_ci(
+ overlaps: Sequence[float],
+ *,
+ null_mean: float,
+ ceiling: float,
+ n_boot: int,
+ seed: int,
+) -> List[float]:
+ if not overlaps or n_boot < 1:
+ return [0.0, 0.0]
+ values = np.asarray(overlaps, dtype=np.float64)
+ rng = np.random.default_rng(seed)
+ denominator = max(_EPS, ceiling - null_mean)
+ samples = np.empty(n_boot, dtype=np.float64)
+ for i in range(n_boot):
+ picked = values[rng.integers(0, len(values), len(values))]
+ samples[i] = (float(picked.mean()) - null_mean) / denominator
+ lo, hi = np.percentile(samples, [2.5, 97.5])
+ return [round(float(lo), 4), round(float(hi), 4)]
+
+
+def detect_answer_identity_confound(
+ manifest_a: Mapping, manifest_b: Mapping
+) -> dict:
+ """Classify whether a correctness contrast controls answer identity.
+
+ This is deliberately conservative: absent structured ``correct`` and
+ ``answer_identity`` metadata, the tool reports ``not_assessable`` instead
+ of inferring correctness from free text.
+ """
+ correct_a, correct_b = manifest_a.get("correct"), manifest_b.get("correct")
+ answer_a, answer_b = manifest_a.get("answer_identity"), manifest_b.get("answer_identity")
+ if not isinstance(correct_a, bool) or not isinstance(correct_b, bool):
+ return {
+ "status": "not_assessable",
+ "controlled": None,
+ "reason": "manifests need boolean correct and string answer_identity fields",
+ }
+ if correct_a == correct_b:
+ return {
+ "status": "not_applicable",
+ "controlled": None,
+ "reason": "the pair does not contrast correctness",
+ }
+ if not isinstance(answer_a, str) or not isinstance(answer_b, str):
+ return {
+ "status": "not_assessable",
+ "controlled": None,
+ "reason": "answer_identity is missing from one or both manifests",
+ }
+ controlled = answer_a == answer_b
+ return {
+ "status": "controlled" if controlled else "confounded",
+ "controlled": controlled,
+ "answer_identity_a": answer_a,
+ "answer_identity_b": answer_b,
+ "reason": (
+ "correctness differs while answer identity is held constant"
+ if controlled else
+ "correctness and answer identity both change in this pair"
+ ),
+ }
+
+
+def _trace_provenance(trace: Mapping) -> dict:
+ manifest = trace["manifest"]
+ jacobian = manifest.get("jacobian", {})
+ capture = manifest.get("capture", {})
+ return {
+ "trace_id": manifest.get("trace_id"),
+ "model": manifest.get("model"),
+ "model_revision": manifest.get("revision"),
+ "layer": manifest.get("layer"),
+ "language": manifest.get("language"),
+ "lens_sha1": jacobian.get("sha1"),
+ "corpus_id": jacobian.get("corpus_id"),
+ "estimator": jacobian.get("estimator"),
+ "capture_tool": capture.get("tool"),
+ "created_at": capture.get("created_at"),
+ "is_demo": bool(manifest.get("is_demo", False)),
+ }
+
+
+def build_contrast(
+ trace_a: Mapping,
+ trace_b: Mapping,
+ lexicon: Mapping[str, str],
+ *,
+ position_map: Optional[Sequence[Sequence[int]]] = None,
+ mode: str = "cross_language",
+ n_permutations: int = 1000,
+ n_boot: int = 1000,
+ seed: int = 0,
+) -> dict:
+ """Build a complete, serializable Contrast Lab evidence record."""
+ tokens_a, tokens_b = trace_a["tokens"], trace_b["tokens"]
+ if position_map is None:
+ position_map = [[i, i] for i in range(min(len(tokens_a), len(tokens_b)))]
+ position_map = [[int(ia), int(ib)] for ia, ib in position_map]
+ if not position_map:
+ raise ValueError("position_map must contain at least one aligned position")
+ if any(ia < 0 or ia >= len(tokens_a) or ib < 0 or ib >= len(tokens_b)
+ for ia, ib in position_map):
+ raise IndexError("position_map references a token outside one of the traces")
+
+ aligned_a = [token_concepts(tokens_a[ia], lexicon) for ia, _ in position_map]
+ aligned_b = [token_concepts(tokens_b[ib], lexicon) for _, ib in position_map]
+ # Empty/empty filler positions contain no hypothesis-bearing observation.
+ # Including them would dilute both observed overlap and the permutation null
+ # according to narrative length rather than representation quality.
+ informative = [
+ i for i, (concepts_a, concepts_b) in enumerate(zip(aligned_a, aligned_b))
+ if concepts_a or concepts_b
+ ]
+ evidence_a = [aligned_a[i] for i in informative]
+ evidence_b = [aligned_b[i] for i in informative]
+ if not evidence_a:
+ raise ValueError("the aligned pair contains no concepts in either trace")
+ observed, null = _permutation_null(
+ evidence_a, evidence_b, n_permutations=n_permutations, seed=seed
+ )
+ ceiling = float(np.mean([
+ estimate_same_lang_baseline(evidence_a),
+ estimate_same_lang_baseline(evidence_b),
+ ]))
+ denominator = max(_EPS, ceiling - null["mean"])
+ sharedness = (observed - null["mean"]) / denominator
+ overlaps = [overlap_coef(a, b) for a, b in zip(evidence_a, evidence_b)]
+ ci = _paired_bootstrap_ci(
+ overlaps,
+ null_mean=null["mean"],
+ ceiling=ceiling,
+ n_boot=n_boot,
+ seed=seed + 17,
+ )
+ pair_rigor = compute_pair_rigor(
+ tokens_a, tokens_b, position_map, dict(lexicon), seed=seed
+ )
+
+ return {
+ "schema_version": CONTRAST_SCHEMA_VERSION,
+ "mode": mode,
+ "trace_a": trace_a["manifest"].get("trace_id"),
+ "trace_b": trace_b["manifest"].get("trace_id"),
+ "position_map": position_map,
+ "pair_rigor": pair_rigor,
+ "calibration": {
+ "n_aligned_positions": len(position_map),
+ "n_informative_positions": len(informative),
+ "informative_positions": informative,
+ "observed": round(observed, 4),
+ "null": null,
+ "same_lang_ceiling": round(ceiling, 4),
+ "sharedness": {
+ "value": round(float(sharedness), 4),
+ "ci95": ci,
+ "definition": PAIR_SHAREDNESS_DEFINITION,
+ },
+ "token_sharedness_definition": SHAREDNESS_DEFINITION,
+ },
+ "concept_trajectories": concept_rank_trajectories(
+ tokens_a, tokens_b, position_map, lexicon
+ ),
+ "answer_identity_check": detect_answer_identity_confound(
+ trace_a["manifest"], trace_b["manifest"]
+ ),
+ "provenance": {
+ "trace_a": _trace_provenance(trace_a),
+ "trace_b": _trace_provenance(trace_b),
+ "analysis": {
+ "tool": "j7scope.contrast.build_contrast",
+ "schema_version": CONTRAST_SCHEMA_VERSION,
+ "n_permutations": int(n_permutations),
+ "n_boot": int(n_boot),
+ "seed": int(seed),
+ },
+ },
+ }
diff --git a/j7scope/rigor.py b/j7scope/rigor.py
index 634ea5e..463b15b 100644
--- a/j7scope/rigor.py
+++ b/j7scope/rigor.py
@@ -201,9 +201,15 @@ def _combined_concepts(tok: dict, lexicon: Dict[str, str]) -> List[str]:
The cross-trace metric asks whether two sessions reach the same concept, so
it pools zh and en read-out tokens into one concept set per token.
"""
- toks = ([t["token"] for t in tok["readout"].get("zh", [])]
- + [t["token"] for t in tok["readout"].get("en", [])])
- return concepts_of(toks, lexicon)
+ out: List[str] = []
+ seen = set()
+ for bucket in ("zh", "en", "other"):
+ for row in tok["readout"].get(bucket, []):
+ concept = row.get("concept") or lexicon.get(row.get("token"))
+ if concept is not None and concept not in seen:
+ seen.add(concept)
+ out.append(concept)
+ return out
def compute_pair_rigor(
diff --git a/j7scope/trace.py b/j7scope/trace.py
index b47e463..db414e9 100644
--- a/j7scope/trace.py
+++ b/j7scope/trace.py
@@ -276,4 +276,39 @@ def validate_trace_gallery(trace_root: PathLike) -> List[str]:
problems.append(
f"{trace_id}: align.position_map[{pair_index}] is invalid"
)
+ contrast = align.get("contrast")
+ if contrast is not None:
+ from .contrast import (
+ CONTRAST_SCHEMA_VERSION,
+ PAIR_SHAREDNESS_DEFINITION,
+ )
+
+ if contrast.get("schema_version") != CONTRAST_SCHEMA_VERSION:
+ problems.append(
+ f"{trace_id}: contrast schema_version must be "
+ f"{CONTRAST_SCHEMA_VERSION}"
+ )
+ calibration = contrast.get("calibration", {})
+ for key in ("observed", "null", "same_lang_ceiling", "sharedness"):
+ if key not in calibration:
+ problems.append(
+ f"{trace_id}: contrast.calibration missing {key}"
+ )
+ if calibration.get("sharedness", {}).get("definition") != PAIR_SHAREDNESS_DEFINITION:
+ problems.append(
+ f"{trace_id}: contrast sharedness definition drifted"
+ )
+ if contrast.get("position_map") != align.get("position_map"):
+ problems.append(
+ f"{trace_id}: contrast.position_map differs from align"
+ )
+ if not isinstance(contrast.get("concept_trajectories"), list):
+ problems.append(
+ f"{trace_id}: contrast.concept_trajectories must be a list"
+ )
+ provenance = contrast.get("provenance", {})
+ if not all(key in provenance for key in ("trace_a", "trace_b", "analysis")):
+ problems.append(
+ f"{trace_id}: contrast provenance is incomplete"
+ )
return problems
diff --git a/results/traces/demo-narrative-en/manifest.json b/results/traces/demo-narrative-en/manifest.json
index 46d68c7..cec68d5 100644
--- a/results/traces/demo-narrative-en/manifest.json
+++ b/results/traces/demo-narrative-en/manifest.json
@@ -1,6 +1,6 @@
{
"capture": {
- "created_at": "2026-07-22T12:31:46.828887+00:00",
+ "created_at": "2026-07-22T00:00:00+00:00",
"tool": "build_demo_trace.py"
},
"concept": null,
@@ -8,6 +8,11 @@
"is_demo": true,
"jacobian": {
"corpus_id": "demo-synthetic",
+ "estimator": "synthetic",
+ "n_probes": null,
+ "n_prompts": 0,
+ "position": null,
+ "seed": 0,
"sha1": null
},
"kind": "single",
diff --git a/results/traces/demo-parallel-en/align.json b/results/traces/demo-parallel-en/align.json
index 4f17905..1a5677c 100644
--- a/results/traces/demo-parallel-en/align.json
+++ b/results/traces/demo-parallel-en/align.json
@@ -1,4 +1,1063 @@
{
+ "contrast": {
+ "answer_identity_check": {
+ "controlled": null,
+ "reason": "manifests need boolean correct and string answer_identity fields",
+ "status": "not_assessable"
+ },
+ "calibration": {
+ "informative_positions": [
+ 3,
+ 6,
+ 13,
+ 18,
+ 23
+ ],
+ "n_aligned_positions": 25,
+ "n_informative_positions": 5,
+ "null": {
+ "mean": 0.3708,
+ "metric": "paired_mean_concept_overlap",
+ "n": 1000,
+ "p05": 0.0,
+ "p95": 0.8,
+ "seed": 0
+ },
+ "observed": 1.0,
+ "same_lang_ceiling": 1.0,
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(paired_observed_mean - permutation_null.mean) / (same_lang_ceiling - permutation_null.mean)",
+ "value": 1.0
+ },
+ "token_sharedness_definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)"
+ },
+ "concept_trajectories": [
+ {
+ "concept": "deception",
+ "coverage": 0.08,
+ "rank_a": [
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 3,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 3,
+ null
+ ]
+ },
+ {
+ "concept": "emotion",
+ "coverage": 0.08,
+ "rank_a": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 3,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 3,
+ null
+ ]
+ },
+ {
+ "concept": "concession",
+ "coverage": 0.04,
+ "rank_a": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null
+ ]
+ },
+ {
+ "concept": "entity",
+ "coverage": 0.04,
+ "rank_a": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null
+ ]
+ },
+ {
+ "concept": "manipulation",
+ "coverage": 0.04,
+ "rank_a": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null
+ ]
+ }
+ ],
+ "mode": "cross_language",
+ "pair_rigor": [
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 0,
+ "ib": 0,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 0,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 1,
+ "ib": 1,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 1,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 2,
+ "ib": 2,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 2,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 3,
+ "ib": 3,
+ "null": {
+ "mean": 0.0417,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 3,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "deception"
+ ],
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 4,
+ "ib": 4,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 4,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 5,
+ "ib": 5,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 5,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 6,
+ "ib": 6,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 6,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "manipulation"
+ ],
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 7,
+ "ib": 7,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 7,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 8,
+ "ib": 8,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 8,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 9,
+ "ib": 9,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 9,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 10,
+ "ib": 10,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 10,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 11,
+ "ib": 11,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 11,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 12,
+ "ib": 12,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 12,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 13,
+ "ib": 13,
+ "null": {
+ "mean": 0.0417,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 13,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "emotion"
+ ],
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 14,
+ "ib": 14,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 14,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 15,
+ "ib": 15,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 15,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 16,
+ "ib": 16,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 16,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 17,
+ "ib": 17,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 17,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 18,
+ "ib": 18,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 18,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "concession"
+ ],
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 19,
+ "ib": 19,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 19,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 20,
+ "ib": 20,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 20,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 21,
+ "ib": 21,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 21,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 22,
+ "ib": 22,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 22,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 23,
+ "ib": 23,
+ "null": {
+ "mean": 0.0833,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.85
+ },
+ "position": 23,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "deception",
+ "emotion",
+ "entity"
+ ],
+ "sharedness": {
+ "ci95": [
+ -0.0909,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 24,
+ "ib": 24,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 24,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ }
+ ],
+ "position_map": [
+ [
+ 0,
+ 0
+ ],
+ [
+ 1,
+ 1
+ ],
+ [
+ 2,
+ 2
+ ],
+ [
+ 3,
+ 3
+ ],
+ [
+ 4,
+ 4
+ ],
+ [
+ 5,
+ 5
+ ],
+ [
+ 6,
+ 6
+ ],
+ [
+ 7,
+ 7
+ ],
+ [
+ 8,
+ 8
+ ],
+ [
+ 9,
+ 9
+ ],
+ [
+ 10,
+ 10
+ ],
+ [
+ 11,
+ 11
+ ],
+ [
+ 12,
+ 12
+ ],
+ [
+ 13,
+ 13
+ ],
+ [
+ 14,
+ 14
+ ],
+ [
+ 15,
+ 15
+ ],
+ [
+ 16,
+ 16
+ ],
+ [
+ 17,
+ 17
+ ],
+ [
+ 18,
+ 18
+ ],
+ [
+ 19,
+ 19
+ ],
+ [
+ 20,
+ 20
+ ],
+ [
+ 21,
+ 21
+ ],
+ [
+ 22,
+ 22
+ ],
+ [
+ 23,
+ 23
+ ],
+ [
+ 24,
+ 24
+ ]
+ ],
+ "provenance": {
+ "analysis": {
+ "n_boot": 1000,
+ "n_permutations": 1000,
+ "schema_version": 1,
+ "seed": 0,
+ "tool": "j7scope.contrast.build_contrast"
+ },
+ "trace_a": {
+ "capture_tool": "build_demo_trace.py",
+ "corpus_id": "demo-synthetic",
+ "created_at": "2026-07-22T00:00:00+00:00",
+ "estimator": "synthetic",
+ "is_demo": true,
+ "language": "en",
+ "layer": 18,
+ "lens_sha1": null,
+ "model": "demo-synthetic",
+ "model_revision": null,
+ "trace_id": "demo-parallel-en"
+ },
+ "trace_b": {
+ "capture_tool": "build_demo_trace.py",
+ "corpus_id": "demo-synthetic",
+ "created_at": "2026-07-22T00:00:00+00:00",
+ "estimator": "synthetic",
+ "is_demo": true,
+ "language": "zh",
+ "layer": 18,
+ "lens_sha1": null,
+ "model": "demo-synthetic",
+ "model_revision": null,
+ "trace_id": "demo-parallel-zh"
+ }
+ },
+ "schema_version": 1,
+ "trace_a": "demo-parallel-en",
+ "trace_b": "demo-parallel-zh"
+ },
"members": {
"en": "demo-parallel-en",
"zh": "demo-parallel-zh"
diff --git a/results/traces/demo-parallel-en/manifest.json b/results/traces/demo-parallel-en/manifest.json
index 61d6af3..9a8ef25 100644
--- a/results/traces/demo-parallel-en/manifest.json
+++ b/results/traces/demo-parallel-en/manifest.json
@@ -1,6 +1,6 @@
{
"capture": {
- "created_at": "2026-07-22T12:31:46.837569+00:00",
+ "created_at": "2026-07-22T00:00:00+00:00",
"tool": "build_demo_trace.py"
},
"concept": null,
@@ -8,6 +8,11 @@
"is_demo": true,
"jacobian": {
"corpus_id": "demo-synthetic",
+ "estimator": "synthetic",
+ "n_probes": null,
+ "n_prompts": 0,
+ "position": null,
+ "seed": 0,
"sha1": null
},
"kind": "parallel_member",
diff --git a/results/traces/demo-parallel-zh/align.json b/results/traces/demo-parallel-zh/align.json
index 4f17905..1a5677c 100644
--- a/results/traces/demo-parallel-zh/align.json
+++ b/results/traces/demo-parallel-zh/align.json
@@ -1,4 +1,1063 @@
{
+ "contrast": {
+ "answer_identity_check": {
+ "controlled": null,
+ "reason": "manifests need boolean correct and string answer_identity fields",
+ "status": "not_assessable"
+ },
+ "calibration": {
+ "informative_positions": [
+ 3,
+ 6,
+ 13,
+ 18,
+ 23
+ ],
+ "n_aligned_positions": 25,
+ "n_informative_positions": 5,
+ "null": {
+ "mean": 0.3708,
+ "metric": "paired_mean_concept_overlap",
+ "n": 1000,
+ "p05": 0.0,
+ "p95": 0.8,
+ "seed": 0
+ },
+ "observed": 1.0,
+ "same_lang_ceiling": 1.0,
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(paired_observed_mean - permutation_null.mean) / (same_lang_ceiling - permutation_null.mean)",
+ "value": 1.0
+ },
+ "token_sharedness_definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)"
+ },
+ "concept_trajectories": [
+ {
+ "concept": "deception",
+ "coverage": 0.08,
+ "rank_a": [
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 3,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 3,
+ null
+ ]
+ },
+ {
+ "concept": "emotion",
+ "coverage": 0.08,
+ "rank_a": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 3,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 3,
+ null
+ ]
+ },
+ {
+ "concept": "concession",
+ "coverage": 0.04,
+ "rank_a": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null
+ ]
+ },
+ {
+ "concept": "entity",
+ "coverage": 0.04,
+ "rank_a": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null
+ ]
+ },
+ {
+ "concept": "manipulation",
+ "coverage": 0.04,
+ "rank_a": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null
+ ],
+ "rank_b": [
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ 1,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null,
+ null
+ ]
+ }
+ ],
+ "mode": "cross_language",
+ "pair_rigor": [
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 0,
+ "ib": 0,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 0,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 1,
+ "ib": 1,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 1,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 2,
+ "ib": 2,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 2,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 3,
+ "ib": 3,
+ "null": {
+ "mean": 0.0417,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 3,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "deception"
+ ],
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 4,
+ "ib": 4,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 4,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 5,
+ "ib": 5,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 5,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 6,
+ "ib": 6,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 6,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "manipulation"
+ ],
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 7,
+ "ib": 7,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 7,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 8,
+ "ib": 8,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 8,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 9,
+ "ib": 9,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 9,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 10,
+ "ib": 10,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 10,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 11,
+ "ib": 11,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 11,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 12,
+ "ib": 12,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 12,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 13,
+ "ib": 13,
+ "null": {
+ "mean": 0.0417,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 13,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "emotion"
+ ],
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 14,
+ "ib": 14,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 14,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 15,
+ "ib": 15,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 15,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 16,
+ "ib": 16,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 16,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 17,
+ "ib": 17,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 17,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 18,
+ "ib": 18,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 18,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "concession"
+ ],
+ "sharedness": {
+ "ci95": [
+ 1.0,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 19,
+ "ib": 19,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 19,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 20,
+ "ib": 20,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 20,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 21,
+ "ib": 21,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 21,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 22,
+ "ib": 22,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 22,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ },
+ {
+ "cross_lang_overlap": 1.0,
+ "ia": 23,
+ "ib": 23,
+ "null": {
+ "mean": 0.0833,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.85
+ },
+ "position": 23,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [
+ "deception",
+ "emotion",
+ "entity"
+ ],
+ "sharedness": {
+ "ci95": [
+ -0.0909,
+ 1.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 1.0
+ }
+ },
+ {
+ "cross_lang_overlap": 0.0,
+ "ia": 24,
+ "ib": 24,
+ "null": {
+ "mean": 0.0,
+ "metric": "concept_overlap",
+ "n": 24,
+ "p05": 0.0,
+ "p95": 0.0
+ },
+ "position": 24,
+ "same_lang_baseline": 1.0,
+ "shared_concepts": [],
+ "sharedness": {
+ "ci95": [
+ 0.0,
+ 0.0
+ ],
+ "definition": "(cross_lang_overlap - null.mean) / (same_lang_baseline - null.mean)",
+ "value": 0.0
+ }
+ }
+ ],
+ "position_map": [
+ [
+ 0,
+ 0
+ ],
+ [
+ 1,
+ 1
+ ],
+ [
+ 2,
+ 2
+ ],
+ [
+ 3,
+ 3
+ ],
+ [
+ 4,
+ 4
+ ],
+ [
+ 5,
+ 5
+ ],
+ [
+ 6,
+ 6
+ ],
+ [
+ 7,
+ 7
+ ],
+ [
+ 8,
+ 8
+ ],
+ [
+ 9,
+ 9
+ ],
+ [
+ 10,
+ 10
+ ],
+ [
+ 11,
+ 11
+ ],
+ [
+ 12,
+ 12
+ ],
+ [
+ 13,
+ 13
+ ],
+ [
+ 14,
+ 14
+ ],
+ [
+ 15,
+ 15
+ ],
+ [
+ 16,
+ 16
+ ],
+ [
+ 17,
+ 17
+ ],
+ [
+ 18,
+ 18
+ ],
+ [
+ 19,
+ 19
+ ],
+ [
+ 20,
+ 20
+ ],
+ [
+ 21,
+ 21
+ ],
+ [
+ 22,
+ 22
+ ],
+ [
+ 23,
+ 23
+ ],
+ [
+ 24,
+ 24
+ ]
+ ],
+ "provenance": {
+ "analysis": {
+ "n_boot": 1000,
+ "n_permutations": 1000,
+ "schema_version": 1,
+ "seed": 0,
+ "tool": "j7scope.contrast.build_contrast"
+ },
+ "trace_a": {
+ "capture_tool": "build_demo_trace.py",
+ "corpus_id": "demo-synthetic",
+ "created_at": "2026-07-22T00:00:00+00:00",
+ "estimator": "synthetic",
+ "is_demo": true,
+ "language": "en",
+ "layer": 18,
+ "lens_sha1": null,
+ "model": "demo-synthetic",
+ "model_revision": null,
+ "trace_id": "demo-parallel-en"
+ },
+ "trace_b": {
+ "capture_tool": "build_demo_trace.py",
+ "corpus_id": "demo-synthetic",
+ "created_at": "2026-07-22T00:00:00+00:00",
+ "estimator": "synthetic",
+ "is_demo": true,
+ "language": "zh",
+ "layer": 18,
+ "lens_sha1": null,
+ "model": "demo-synthetic",
+ "model_revision": null,
+ "trace_id": "demo-parallel-zh"
+ }
+ },
+ "schema_version": 1,
+ "trace_a": "demo-parallel-en",
+ "trace_b": "demo-parallel-zh"
+ },
"members": {
"en": "demo-parallel-en",
"zh": "demo-parallel-zh"
diff --git a/results/traces/demo-parallel-zh/manifest.json b/results/traces/demo-parallel-zh/manifest.json
index 1de791d..c6c1009 100644
--- a/results/traces/demo-parallel-zh/manifest.json
+++ b/results/traces/demo-parallel-zh/manifest.json
@@ -1,6 +1,6 @@
{
"capture": {
- "created_at": "2026-07-22T12:31:46.844261+00:00",
+ "created_at": "2026-07-22T00:00:00+00:00",
"tool": "build_demo_trace.py"
},
"concept": null,
@@ -8,6 +8,11 @@
"is_demo": true,
"jacobian": {
"corpus_id": "demo-synthetic",
+ "estimator": "synthetic",
+ "n_probes": null,
+ "n_prompts": 0,
+ "position": null,
+ "seed": 0,
"sha1": null
},
"kind": "parallel_member",
diff --git a/tests/test_contrast.py b/tests/test_contrast.py
new file mode 100644
index 0000000..9c83081
--- /dev/null
+++ b/tests/test_contrast.py
@@ -0,0 +1,82 @@
+from copy import deepcopy
+
+from j7scope.contrast import (
+ PAIR_SHAREDNESS_DEFINITION,
+ build_contrast,
+ detect_answer_identity_confound,
+)
+
+
+def _token(seq, concept, surface):
+ return {
+ "seq": seq,
+ "token": surface,
+ "readout": {
+ "zh": [{"token": f"中-{surface}", "rank": 0, "score": 4.0,
+ "concept": concept}],
+ "en": [{"token": f"en-{surface}", "rank": 1, "score": 3.0,
+ "concept": concept}],
+ "other": [],
+ },
+ }
+
+
+def _trace(trace_id, language, concepts, **manifest_fields):
+ manifest = {
+ "trace_id": trace_id,
+ "model": "test-model",
+ "revision": "abc123",
+ "layer": 12,
+ "language": language,
+ "jacobian": {"sha1": "lens-hash", "corpus_id": "corpus-hash",
+ "estimator": "test"},
+ "capture": {"tool": "pytest", "created_at": "2026-01-01T00:00:00Z"},
+ "is_demo": False,
+ }
+ manifest.update(manifest_fields)
+ return {
+ "manifest": manifest,
+ "tokens": [_token(i, concept, f"t{i}") for i, concept in enumerate(concepts)],
+ "metrics": {},
+ }
+
+
+def test_build_contrast_is_deterministic_and_bakes_research_controls():
+ trace_a = _trace("a", "en", ["alpha", "beta", "gamma", "delta"])
+ trace_b = _trace("b", "zh", ["alpha", "beta", "gamma", "delta"])
+
+ first = build_contrast(
+ trace_a, trace_b, {}, n_permutations=200, n_boot=200, seed=9
+ )
+ second = build_contrast(
+ trace_a, trace_b, {}, n_permutations=200, n_boot=200, seed=9
+ )
+
+ assert first == second
+ assert first["calibration"]["observed"] == 1.0
+ assert first["calibration"]["null"]["n"] == 200
+ assert first["calibration"]["sharedness"]["definition"] == PAIR_SHAREDNESS_DEFINITION
+ assert first["concept_trajectories"][0]["rank_a"][0] in (1, None)
+ assert first["provenance"]["trace_a"]["lens_sha1"] == "lens-hash"
+ assert len(first["pair_rigor"]) == 4
+
+
+def test_answer_identity_check_flags_and_clears_confound():
+ base = _trace("a", "en", ["alpha"], correct=True, answer_identity="A")
+ changed_answer = _trace("b", "en", ["alpha"], correct=False, answer_identity="B")
+ same_answer = deepcopy(changed_answer)
+ same_answer["manifest"]["answer_identity"] = "A"
+
+ confounded = detect_answer_identity_confound(base["manifest"], changed_answer["manifest"])
+ controlled = detect_answer_identity_confound(base["manifest"], same_answer["manifest"])
+
+ assert confounded["status"] == "confounded"
+ assert confounded["controlled"] is False
+ assert controlled["status"] == "controlled"
+ assert controlled["controlled"] is True
+
+
+def test_answer_identity_check_never_guesses_from_free_text():
+ result = detect_answer_identity_confound({"prompt": "correct"}, {"prompt": "wrong"})
+ assert result["status"] == "not_assessable"
+ assert result["controlled"] is None
diff --git a/tests/test_recorder.py b/tests/test_recorder.py
index 465cd0c..91807fb 100644
--- a/tests/test_recorder.py
+++ b/tests/test_recorder.py
@@ -115,3 +115,38 @@ def test_hf_backend_rejects_precomputed_jacobian_without_metadata(tmp_path):
with pytest.raises(FileNotFoundError, match="metadata"):
backend._load_precomputed_jacobian()
+
+
+def test_hf_backend_exposes_activation_patch_with_provenance(monkeypatch):
+ import j7scope.patching
+
+ backend = HFBackend(model_name="test/model", layer=12)
+ backend._model = object()
+ backend._jlens = object()
+ backend.model_revision_resolved = "revision-1"
+ backend.jacobian_sha1 = "lens-sha"
+
+ def fake_patch(jlens, source, target, patch_layer, src_pos, tgt_pos, k):
+ assert jlens is backend._jlens
+ assert (source, target, patch_layer, src_pos, tgt_pos, k) == (
+ "source", "target", 6, 2, 3, 5
+ )
+ return {
+ "readout": [("concept", 4.2)],
+ "next_token": [("answer", 3.1)],
+ }
+
+ monkeypatch.setattr(j7scope.patching, "patch_and_readout", fake_patch)
+ result = backend.intervene(
+ source_prompt="source",
+ target_prompt="target",
+ patch_layer=6,
+ source_position=2,
+ target_position=3,
+ topk=5,
+ )
+
+ assert backend.supports_intervention is True
+ assert result["readout"] == [("concept", 4.2)]
+ assert result["jacobian_sha1"] == "lens-sha"
+ assert result["model_revision"] == "revision-1"
diff --git a/tests/test_trace_gallery.py b/tests/test_trace_gallery.py
index 97d1e28..c3ea87e 100644
--- a/tests/test_trace_gallery.py
+++ b/tests/test_trace_gallery.py
@@ -24,6 +24,22 @@ def test_demo_builder_produces_valid_gallery(tmp_path):
)
assert validate_trace_gallery(tmp_path) == []
+ align = json.loads(
+ (tmp_path / "demo-parallel-en" / "align.json").read_text(encoding="utf-8")
+ )
+ assert align["contrast"]["calibration"]["null"]["n"] == 1000
+ assert align["contrast"]["concept_trajectories"]
+ assert align["contrast"]["answer_identity_check"]["status"] == "not_assessable"
+
+
+def test_contrast_lab_page_keeps_research_statistics_precomputed():
+ page = (ROOT / "apps" / "site" / "contrast.html").read_text(encoding="utf-8")
+
+ assert "Contrast Lab" in page
+ assert "C.calibration" in page
+ assert "jspace/interventions" in page
+ assert "j7scope.claim_card" in page
+ assert "cross_lang_overlap -" not in page
def test_gallery_validator_detects_index_drift(tmp_path):
