SPB Git

spb/modelmap Public License

Internal cartography of local LLMs on Apple Silicon — registered, gated, negative-first. Public atlas at modelmap.io.

🔗 https://modelmap.io
interpretabilityllmapple-siliconmlxreproducibility
Python 66.3% JavaScript 24.5% CSS 8.1% Shell 0.7%

expC run #6: L12 handle fails replication — Level 3 abandoned (as registered)

Only 1/4 fresh sources monotone (positive dose arm unstable); specificity
failed to replicate (random |delta| 3.36 vs bound 1.14 — run #5's 0.45 was
3-draw sampling luck). Halving held 4/4: necessary but not a reliable
additive handle. interventions/v1 (Level 2) stands as the final causal
claim. Gate record: refuse -> pass(L2) -> refuse -> refuse. Methodology
rule adopted: specificity controls need >=10 random draws + percentile
bound. Charts: x-axis customization for dose-response figures.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Simon-Pierre Boucher committed 1 h ago (Aug 12, 2026) parent f2f47f9

Showing 7 changed files with +431 and −27

modified experiments/micro/expC_causal_verification/analysis.md +44 −0
@@ -273,3 +273,47 @@ Three refusals/passes to date, all pre-registered: run #3 per-layer profile
273 273 REFUSED → run #4 band claim PASSED (Level 2 published) → run #5 L3
274 274 conjunction REFUSED (v1 unchanged). The atlas never received a claim its
275 275 evidence didn't carry.
276 +
277 +---
278 +
279 +# Analysis — expC run #6: the L12 handle claim fails replication — Level 3 abandoned
280 +
281 +Run: `results/expC_causal_verification/20260812T072913Z/results.json`.
282 +Four fresh direction sources at L12, third fresh bank (192 pairs, baseline
283 ++4.57), fresh random-direction seeds. Registered rule: any source failing
284 +monotonicity/halving, or specificity failing, kills the L3 claim and closes
285 +the steering program for this object.
286 +
287 +```text
288 +Hypothesis : L12 is an estimator-stable dose-controlled handle.
289 +Result : FALSIFIED. Monotone: 1/4 sources only (Ahalf2);
290 + the POSITIVE dose arm is unstable (Ahalf1 +2σ dips
291 + 5.59<6.29; Bhalf1 +1σ dips below baseline; Bhalf2
292 + +2σ < +1σ). Halving at −2σ: 4/4 — the negative
293 + (erasure-like) arm is robust, again. SPECIFICITY
294 + FAILED TO REPLICATE: random-direction mean |Δ| =
295 + 3.36 vs bound 1.14 on the fresh bank and fresh
296 + random seeds (run #5's L12 value was 0.45 — with
297 + only 3 random draws, that pass now reads as
298 + sampling luck).
299 +Interpretation : Level 1 for the negative. The direction is
300 + NECESSARY (v1, Level 2, band-replicated) but NOT
301 + a reliable additive handle: pushing along it does
302 + not control the behavior in a dose-stable,
303 + direction-specific way. Per the pre-registered
304 + rule, Level 3 is abandoned for this object and
305 + the steering program is closed. Had run #5's L12
306 + observation been published without fresh
307 + re-registration, the atlas would now contain a
308 + false Level-3 claim — the gate earned its keep a
309 + third time.
310 +Methodology lesson : 3 random-direction draws are too few for a
311 + specificity bound; methodology.md will require
312 + ≥10 draws with a percentile bound for any
313 + specificity control from now on.
314 +Next experiment : program closed here. Proceeding tracks: arith_valid
315 + band protocol; candidate_02 (quantization drift of
316 + the Level-2 band map); expG (band replication on
317 + Qwen3-1.7B). The final gate record for this arc:
318 + refuse → pass(L2) → refuse → refuse.
319 +```
modified experiments/micro/expC_causal_verification/hypothesis.md +39 −0
@@ -216,3 +216,42 @@ Result : (pending)
216 216 Interpretation : (pending)
217 217 Next experiment : (pending)
218 218 ```
219 +
220 +---
221 +
222 +# Hypothesis — expC run #6 (minimal L3 claim: layer 12 as a single-layer handle)
223 +
224 +Registered 2026-08-12 **before** the run, after run #5's conjunction failed
225 +while L12 passed every gate. The claim is narrowed to the object that
226 +showed textbook behavior — scope discipline again, now applied to layers.
227 +
228 +```text
229 +Hypothesis : At LAYER 12 ONLY, the diff-of-means agreement
230 + direction is a dose-controlled causal handle, and
231 + this is estimator-stable: for EVERY one of four
232 + fresh direction sources (disjoint halves of
233 + agreement_A and agreement_B, new permutation
234 + seed), on a THIRD fresh behavioral bank (8
235 + never-used locations):
236 + (a) margins strictly monotone in
237 + α ∈ {−2,−1,0,+1,+2}·σ;
238 + (b) margin(−2σ) ≤ 0.5 × baseline;
239 + (c) shared specificity control at L12: 3 random
240 + directions at ±2σ move the margin by < 25%
241 + of baseline on average.
242 +Falsification criterion : any source failing (a) or (b), or (c) failing —
243 + the single-layer L3 claim dies and Level 3 is
244 + abandoned for this object (steering program
245 + closed; candidate_02/expG proceed from the
246 + Level-2 band entry).
247 +Method : same margin metric; direction + σ re-estimated
248 + per source at L12 from half-set mean-pooled reps;
249 + doses applied at L12 output, all positions.
250 +Publication rule : pass → interventions/v2 at LEVEL 3 with scope
251 + declared as {band necessity (v1) + single-layer
252 + L12 handle}; fail → documented refusal, v1
253 + unchanged.
254 +Result : (pending)
255 +Interpretation : (pending)
256 +Next experiment : (pending)
257 +```
added experiments/micro/expC_causal_verification/implementation/benchmark_v6.py +169 −0
@@ -0,0 +1,169 @@
1 +#!/usr/bin/env python3
2 +# =============================================================================
3 +# Project : modelmap
4 +# File : experiments/micro/expC_causal_verification/implementation/benchmark_v6.py
5 +# Purpose : Run #6 — minimal L3 claim: layer-12 handle, four fresh sources
6 +# Author : Simon-Pierre Boucher
7 +# Contact : contact@spboucher.ai
8 +# Website : https://modelmap.io
9 +# Created : 2026-08-12
10 +# Modified : 2026-08-12
11 +# Platform : macOS / Apple Silicon (arm64) — MLX / Metal
12 +# License : All rights reserved (research code)
13 +# =============================================================================
14 +"""expC run #6 (hypothesis registered before this run).
15 +
16 +Single-layer (L12) dose-response handle claim, four fresh direction sources
17 +(new-permutation disjoint halves of A and B), third fresh behavioral bank,
18 +shared random-direction specificity control.
19 +"""
20 +
21 +from __future__ import annotations
22 +
23 +import json
24 +import random
25 +import subprocess
26 +import sys
27 +import time
28 +from pathlib import Path
29 +
30 +import numpy as np
31 +
32 +ROOT = Path(__file__).resolve().parents[4]
33 +sys.path.insert(0, str(ROOT / "src"))
34 +sys.path.insert(0, str(ROOT / "benchmarks"))
35 +from hardware_manifest import manifest
36 +
37 +from modelmap.capture.mlx_capture import capture_pooled, install_taps
38 +
39 +MODEL = "mlx-community/Qwen3-0.6B-4bit"
40 +LAYER = 12
41 +DOSES = (-2.0, -1.0, 0.0, 1.0, 2.0)
42 +N_RANDOM_DIRS = 3
43 +SEED = 782
44 +
45 +NOUN_PAIRS = [("key", "keys"), ("crate", "crates"), ("report", "reports"), ("valve", "valves"),
46 + ("ticket", "tickets"), ("ladder", "ladders"), ("sample", "samples"), ("cable", "cables"),
47 + ("permit", "permits"), ("beacon", "beacons"), ("filter", "filters"), ("stamp", "stamps")]
48 +THIRD_NEAR = ["beside the turbine hall", "near the cold frame", "behind the switchboard",
49 + "under the gantry", "next to the signal box", "opposite the pump house",
50 + "inside the drying shed", "along the breakwater"]
51 +
52 +
53 +def main() -> int:
54 + import mlx.core as mx
55 + from mlx_lm import load
56 +
57 + t0 = time.time()
58 + model, tokenizer = load(MODEL)
59 + taps = install_taps(model)
60 +
61 + rng = random.Random(SEED)
62 + combos = [(n, loc) for n in NOUN_PAIRS for loc in THIRD_NEAR]
63 + rng.shuffle(combos)
64 + pairs = []
65 + for (sg, pl), loc in combos:
66 + pairs.append({"prefix": f"The {sg} {loc}", "singular": True})
67 + pairs.append({"prefix": f"The {pl} {loc}", "singular": False})
68 + prefix_ids = [tokenizer.encode(p["prefix"]) for p in pairs]
69 + id_is, id_are = tokenizer.encode(" is")[0], tokenizer.encode(" are")[0]
70 +
71 + def margin() -> float:
72 + out = []
73 + for ids, p in zip(prefix_ids, pairs):
74 + logits = model(mx.array([ids]))[0, -1, :]
75 + mx.eval(logits)
76 + m = float(logits[id_is] - logits[id_are])
77 + out.append(m if p["singular"] else -m)
78 + return float(np.mean(out))
79 +
80 + def add_fn(u_np, delta):
81 + d = mx.array((u_np * delta).astype(np.float32))
82 + def fn(out):
83 + return (out.astype(mx.float32) + d).astype(out.dtype)
84 + return fn
85 +
86 + # four fresh direction sources at L12 (new permutation seed)
87 + sources = {}
88 + d_model = None
89 + for s in ("A", "B"):
90 + items = [json.loads(l) for l in
91 + (ROOT / "benchmarks" / "promptsets" / f"agreement_{s}.jsonl").read_text().splitlines()]
92 + toks = [tokenizer.encode(it["text"]) for it in items]
93 + labels = np.array([it["label"] for it in items])
94 + print(f"capture agreement_{s}…", flush=True)
95 + reps = capture_pooled(model, taps, toks)["mean"][:, LAYER, :]
96 + d_model = reps.shape[-1]
97 + n = len(labels)
98 + perm = np.random.default_rng(SEED + ord(s)).permutation(n)
99 + for hi, idx in enumerate((perm[: n // 2], perm[n // 2:])):
100 + x, l = reps[idx], labels[idx]
101 + u = x[l == "correct"].mean(0) - x[l == "violated"].mean(0)
102 + u = u / (np.linalg.norm(u) + 1e-8)
103 + sources[f"{s}half{hi + 1}"] = {"u": u, "sigma": float((x @ u).std())}
104 +
105 + base_m = margin()
106 + print(f"baseline margin {base_m:+.4f} | pairs {len(pairs)}", flush=True)
107 +
108 + per_source = {}
109 + all_ab = True
110 + for name, sv in sources.items():
111 + dm = {}
112 + for a in DOSES:
113 + if a == 0.0:
114 + dm[a] = base_m
115 + continue
116 + taps[LAYER].edit = add_fn(sv["u"], a * sv["sigma"])
117 + dm[a] = margin()
118 + taps[LAYER].edit = None
119 + seq = [dm[a] for a in DOSES]
120 + mono = all(seq[i] < seq[i + 1] for i in range(len(seq) - 1))
121 + halve = dm[-2.0] <= 0.5 * base_m
122 + all_ab &= (mono and halve)
123 + per_source[name] = {"sigma": sv["sigma"],
124 + "dose_margins": {str(a): dm[a] for a in DOSES},
125 + "monotone": mono, "halved": halve}
126 + print(f"{name:8s} " + " ".join(f"{a:+.0f}σ:{dm[a]:+.2f}" for a in DOSES) +
127 + f" | mono={mono} halve={halve}", flush=True)
128 +
129 + sigma_ref = float(np.mean([sv["sigma"] for sv in sources.values()]))
130 + rand_changes = []
131 + for s in range(N_RANDOM_DIRS):
132 + ru = np.random.default_rng(4000 + s).standard_normal(d_model)
133 + ru /= np.linalg.norm(ru)
134 + for a in (-2.0, 2.0):
135 + taps[LAYER].edit = add_fn(ru.astype(np.float32), a * sigma_ref)
136 + rand_changes.append(abs(margin() - base_m))
137 + taps[LAYER].edit = None
138 + spec = float(np.mean(rand_changes)) < 0.25 * base_m
139 + passes = bool(all_ab and spec)
140 + print(f"L12 HANDLE CLAIM: all-sources mono+halve={all_ab} "
141 + f"specific={spec} (randΔ {np.mean(rand_changes):.2f} vs bound {0.25 * base_m:.2f}) "
142 + f"-> passes={passes}")
143 +
144 + commit = subprocess.run(["git", "rev-parse", "HEAD"], cwd=ROOT,
145 + capture_output=True, text=True, check=False).stdout.strip()
146 + ts = time.strftime("%Y%m%dT%H%M%SZ", time.gmtime())
147 + outdir = ROOT / "results" / "expC_causal_verification" / ts
148 + outdir.mkdir(parents=True)
149 + (outdir / "results.json").write_text(json.dumps({
150 + "experiment": "expC_causal_verification", "run": 6,
151 + "scope": "minimal L3 claim: layer-12 handle, four fresh sources, third bank",
152 + "commit": commit,
153 + "config": {"model": MODEL, "layer": LAYER, "doses": list(DOSES),
154 + "sources": list(sources), "n_random_dirs": N_RANDOM_DIRS,
155 + "n_pairs": len(pairs), "seed": SEED},
156 + "manifest": manifest(),
157 + "baseline_margin": base_m,
158 + "per_source": per_source,
159 + "specificity": {"random_dir_mean_abs_change": float(np.mean(rand_changes)),
160 + "bound": 0.25 * base_m, "specific": spec},
161 + "criterion": {"all_sources_mono_halve": all_ab, "specific": spec, "passes": passes},
162 + "wall_seconds": round(time.time() - t0, 1),
163 + }, indent=2) + "\n")
164 + print(f"results -> {outdir / 'results.json'}")
165 + return 0
166 +
167 +
168 +if __name__ == "__main__":
169 + sys.exit(main())
modified experiments/micro/expC_causal_verification/implementation/make_l3_mapcard.py +31 −23
@@ -54,10 +54,10 @@ def main() -> int:
54 54 if not (V1 / "mapcard.json").exists():
55 55 print("REFUSED: the Level-2 band entry (v1) must exist first.")
56 56 return 1
57 − res5 = newest_run(5)
57 + res5 = newest_run(6)
58 58 doc5 = json.loads(res5.read_text())
59 59 if not doc5["criterion"]["passes"]:
60 − print(f"REFUSED: run #5 criterion failed ({doc5['criterion']}) — v1 stays Level 2.")
60 + print(f"REFUSED: run #6 criterion failed ({doc5['criterion']}) — v1 stays Level 2.")
61 61 return 1
62 62 res4 = newest_run(4)
63 63 doc4 = json.loads(res4.read_text())
@@ -68,16 +68,21 @@ def main() -> int:
68 68 "author": "Simon-Pierre Boucher", "contact": "contact@spboucher.ai",
69 69 "website": "https://modelmap.io",
70 70 "map_type": "interventions", "model_id": MODEL_ID,
71 − "claim": "LEVEL-3 claim: the diff-of-means agreement direction in the early band "
72 − "(layers 2-15) is a causal HANDLE on grammatical-agreement behavior — "
73 − "necessary (erasure removes ~73-75% of the margin, six fresh estimators, "
74 − "run #4) AND controllable (activation-addition steering moves the margin "
75 − "strictly monotonically with dose at layers 4/8/12, halves it at -2sigma, "
76 − "while random directions move it by less than the registered 25% bound, "
77 − "run #5).",
71 + "claim": "LEVEL-3 claim, scope-limited: (i) band necessity — erasing the "
72 + "diff-of-means agreement direction at any early-band layer (2-15) removes "
73 + "~73-75% of the grammatical margin (six fresh estimators, run #4, Level-2 "
74 + "record in v1); (ii) SINGLE-LAYER HANDLE at layer 12 — activation-addition "
75 + "steering moves the margin strictly monotonically with dose and halves it "
76 + "at -2 sigma for EVERY of four fresh direction sources on a third fresh "
77 + "behavioral bank, while random directions at the same doses move it by "
78 + "less than the registered 25% bound (run #6). Runs #3 and #5 record the "
79 + "gate refusals that scoped this claim.",
78 80 "band": v1_map["band"],
79 81 "per_layer": v1_map["per_layer"],
80 − "steering": doc5["results"],
82 + "steering": {"layer": doc5["config"]["layer"],
83 + "per_source": doc5["per_source"],
84 + "specificity": doc5["specificity"],
85 + "baseline_margin": doc5["baseline_margin"]},
81 86 "source_results": {"erasure_band": str(res4.relative_to(ROOT)),
82 87 "steering": str(res5.relative_to(ROOT))},
83 88 }
@@ -113,13 +118,13 @@ def main() -> int:
113 118 controls=["random-direction erasure nulls (netted)",
114 119 "random-direction steering specificity (< 25% bound)",
115 120 "six-source direction replication (run #4)",
116 − "dose-response monotonicity across 5 doses x 3 layers",
117 − "late band excluded as estimator-unstable (run #3 refusal)"],
121 + "four fresh sources x 5 doses at L12, third fresh bank (run #6)",
122 + "late band excluded (run #3 refusal); L3 scope narrowed to L12 (run #5 refusal)"],
118 123 methods_in_agreement=["difference-in-means probing (direction exists)",
119 124 "rank-1 erasure (necessity, band-replicated)",
120 125 "activation-addition steering (dose-controlled sufficiency)"],
121 126 interventions=["rank-1 direction erasure per layer (run #4)",
122 − "activation-addition steering, 5 doses x 3 layers (run #5)"],
127 + "activation-addition steering at L12, 5 doses x 4 sources (run #6)"],
123 128 replication_rate=round(json.loads((V1 / "mapcard.json").read_text())["replication_rate"], 4),
124 129 featurizer_class="linear (difference-in-means direction)",
125 130 intervention_protocol="erasure h-⟨h−μ,u⟩u; steering h+ασℓu, α∈{−2..+2}; "
@@ -133,10 +138,12 @@ def main() -> int:
133 138 (ENTRY / "mapcard.json").write_text(card.to_json())
134 139
135 140 steer_lines = "\n".join(
136 − f"- L{l}: doses " + " ".join(f"{a}σ→{doc5['results']['per_layer'][l]['dose_margins'][a]:+.2f}"
137 − for a in ("-2.0", "-1.0", "0.0", "1.0", "2.0")) +
138 − f" (random-dir mean |Δ| {doc5['results']['per_layer'][l]['random_dir_mean_abs_change']:.2f})"
139 − for l in map(str, doc5["config"]["test_layers"]))
141 + f"- {name}: " + " ".join(f"{a}σ→{sv['dose_margins'][a]:+.2f}"
142 + for a in ("-2.0", "-1.0", "0.0", "1.0", "2.0"))
143 + for name, sv in doc5["per_source"].items()) + (
144 + f"\n- specificity: random-direction mean |Δ| "
145 + f"{doc5['specificity']['random_dir_mean_abs_change']:.2f} vs bound "
146 + f"{doc5['specificity']['bound']:.2f}")
140 147 (ENTRY / "confidence.md").write_text(f"""---
141 148 project: modelmap
142 149 document: qwen3-0.6b-4bit/interventions/v2 — confidence
@@ -159,14 +166,15 @@ Causal verification : YES, both directions — necessity (erasure) and
159 166 dose-controlled sufficiency (steering)
160 167 ```
161 168
162 −Steering dose-response (baseline margin {doc5['results']['baseline_margin']:+.2f}):
169 +Steering dose-response at layer {doc5['config']['layer']} (baseline margin {doc5['baseline_margin']:+.2f}):
163 170 {steer_lines}
164 171
165 −All three pre-registered gates passed: strict monotonicity at every test
166 −layer, halving at −2σ, random-direction specificity under the 25% bound.
167 −Scope honesty: one model, one behavior, band granularity; generality across
168 −sizes (expG) and quantization levels (candidate_02) is untested and is the
169 −registered next question.
172 +All pre-registered gates passed at L12 for every fresh source: strict
173 +monotonicity, halving at −2σ, random-direction specificity under the 25%
174 +bound. Scope honesty: the HANDLE claim is single-layer (L12) and the
175 +NECESSITY claim is band-level (v1); one model, one behavior. Runs #3 and #5
176 +are the refusals that scoped this entry. Generality across sizes (expG) and
177 +quantization (candidate_02) is untested and registered next.
170 178 """)
171 179 errs = card.validate()
172 180 if errs:
modified research/LOG.md +27 −0
@@ -507,3 +507,30 @@ dose calibration + handle claim on the perturbation-tolerant sub-band;
507 507 or the minimal L12 single-layer L3 claim with fresh direction estimates.
508 508 Parallel tracks unchanged: arith_valid band, candidate_02 quantization
509 509 drift, expG cross-model.
510 +
511 +---
512 +
513 +## 2026-08-12 10:15 EDT — expC run #6: L12 handle fails replication — Level 3 abandoned, program closed
514 +
515 +**Result (pre-registered kill rule).** FALSIFIED on both prongs: only 1/4
516 +fresh sources monotone (the POSITIVE dose arm is unstable); specificity
517 +did not replicate (random-direction |Δ| 3.36 vs bound 1.14 on fresh bank +
518 +fresh seeds — run #5's 0.45 at L12 was sampling luck of 3 draws). Halving
519 +at −2σ held 4/4: the negative (erasure-like) arm is robust, consistent
520 +with the Level-2 entry. **Verdict: the agreement direction is NECESSARY
521 +but NOT a reliable additive handle. Level 3 abandoned for this object;
522 +steering program closed; interventions/v1 (Level 2) stands as the final
523 +causal claim for agreement at 0.6B.**
524 +
525 +**The gate earned its keep a third time:** had run #5's L12 observation
526 +been published without fresh re-registration, the atlas would now hold a
527 +false Level-3 claim. Final arc record: refuse (r3) → pass L2 (r4) →
528 +refuse (r5) → refuse (r6).
529 +
530 +**Methodology rule adopted:** specificity controls need ≥10 random-
531 +direction draws with a percentile bound (3 was too few) — goes into
532 +methodology.md and the map-card doctrine.
533 +
534 +**Proceeding tracks (from the Level-2 base):** arith_valid band protocol;
535 +candidate_02 quantization drift of the band map; expG cross-model
536 +replication on Qwen3-1.7B.
added results/expC_causal_verification/20260812T072913Z/results.json +115 −0
@@ -0,0 +1,115 @@
1 +{
2 + "experiment": "expC_causal_verification",
3 + "run": 6,
4 + "scope": "minimal L3 claim: layer-12 handle, four fresh sources, third bank",
5 + "commit": "f2f47f94aa921f3bc7637a0f83a22739c36c5620",
6 + "config": {
7 + "model": "mlx-community/Qwen3-0.6B-4bit",
8 + "layer": 12,
9 + "doses": [
10 + -2.0,
11 + -1.0,
12 + 0.0,
13 + 1.0,
14 + 2.0
15 + ],
16 + "sources": [
17 + "Ahalf1",
18 + "Ahalf2",
19 + "Bhalf1",
20 + "Bhalf2"
21 + ],
22 + "n_random_dirs": 3,
23 + "n_pairs": 192,
24 + "seed": 782
25 + },
26 + "manifest": {
27 + "author": "Simon-Pierre Boucher",
28 + "contact": "contact@spboucher.ai",
29 + "website": "https://modelmap.io",
30 + "chip": {
31 + "brand": "Apple M5 Max",
32 + "cores_total": 18,
33 + "cores_performance": 6,
34 + "cores_efficiency": 12
35 + },
36 + "memory": {
37 + "unified_gb": 48.0,
38 + "pagesize": 16384
39 + },
40 + "os": {
41 + "system": "Darwin",
42 + "version": "27.0",
43 + "arch": "arm64"
44 + },
45 + "software": {
46 + "python": "3.14.4",
47 + "numpy": "2.5.2",
48 + "mlx": "0.32.0",
49 + "torch": "2.13.0",
50 + "safetensors": "0.8.0"
51 + }
52 + },
53 + "baseline_margin": 4.567708333333333,
54 + "per_source": {
55 + "Ahalf1": {
56 + "sigma": 42.242393493652344,
57 + "dose_margins": {
58 + "-2.0": 1.73828125,
59 + "-1.0": 2.4296875,
60 + "0.0": 4.567708333333333,
61 + "1.0": 6.28857421875,
62 + "2.0": 5.590657552083333
63 + },
64 + "monotone": false,
65 + "halved": true
66 + },
67 + "Ahalf2": {
68 + "sigma": 49.152095794677734,
69 + "dose_margins": {
70 + "-2.0": 1.2859700520833333,
71 + "-1.0": 2.7734375,
72 + "0.0": 4.567708333333333,
73 + "1.0": 4.690755208333333,
74 + "2.0": 5.4912109375
75 + },
76 + "monotone": true,
77 + "halved": true
78 + },
79 + "Bhalf1": {
80 + "sigma": 45.88282012939453,
81 + "dose_margins": {
82 + "-2.0": 1.9191080729166667,
83 + "-1.0": 3.6044921875,
84 + "0.0": 4.567708333333333,
85 + "1.0": 3.935546875,
86 + "2.0": 4.4736328125
87 + },
88 + "monotone": false,
89 + "halved": true
90 + },
91 + "Bhalf2": {
92 + "sigma": 38.40946578979492,
93 + "dose_margins": {
94 + "-2.0": 0.7444661458333334,
95 + "-1.0": 2.0065104166666665,
96 + "0.0": 4.567708333333333,
97 + "1.0": 5.039388020833333,
98 + "2.0": 4.46142578125
99 + },
100 + "monotone": false,
101 + "halved": true
102 + }
103 + },
104 + "specificity": {
105 + "random_dir_mean_abs_change": 3.3635207282172312,
106 + "bound": 1.1419270833333333,
107 + "specific": false
108 + },
109 + "criterion": {
110 + "all_sources_mono_halve": false,
111 + "specific": false,
112 + "passes": false
113 + },
114 + "wall_seconds": 15.3
115 +}
modified site/lib/charts.js +6 −4
@@ -32,7 +32,7 @@ const fmt = (v) => (Math.abs(v) >= 100 ? v.toFixed(0) : Math.abs(v) >= 10 ? v.to
32 32 * series: [{label, color, dash?, ref?, values:[y per layer]}]
33 33 * Direct labels at line ends + legend; per-point hover targets.
34 34 */
35 −function layerLineSvg({ title, series, yLabel, yMin = 0, yMax = 1, caption }) {
35 +function layerLineSvg({ title, series, yLabel, yMin = 0, yMax = 1, caption, xTitle = "layer", xTickLabels = null }) {
36 36 const W = 720, H = 300, ML = 52, MR = 118, MT = 16, MB = 40;
37 37 const iw = W - ML - MR, ih = H - MT - MB;
38 38 const nx = Math.max(...series.map((s) => s.values.length));
@@ -52,10 +52,12 @@ function layerLineSvg({ title, series, yLabel, yMin = 0, yMax = 1, caption }) {
52 52 grid += `<line x1="${ML}" y1="${Y(0)}" x2="${ML + iw}" y2="${Y(0)}" class="axis"/>`;
53 53 }
54 54 let xt = "";
55 + const xlab = (i) => (xTickLabels ? (xTickLabels[i] ?? "") : String(i));
55 56 for (let i = 0; i < nx; i += Math.ceil(nx / 8)) {
56 − xt += `<text x="${X(i)}" y="${MT + ih + 18}" class="tick" text-anchor="middle">${i}</text>`;
57 + xt += `<text x="${X(i)}" y="${MT + ih + 18}" class="tick" text-anchor="middle">${xlab(i)}</text>`;
57 58 }
58 − xt += `<text x="${X(nx - 1)}" y="${MT + ih + 18}" class="tick" text-anchor="middle">${nx - 1}</text>`;
59 + if (!xTickLabels) xt += `<text x="${X(nx - 1)}" y="${MT + ih + 18}" class="tick" text-anchor="middle">${xlab(nx - 1)}</text>`;
60 + else for (let i = 1; i < nx; i += 1) if (i % Math.ceil(nx / 8) !== 0) xt += `<text x="${X(i)}" y="${MT + ih + 18}" class="tick" text-anchor="middle">${xlab(i)}</text>`;
59 61
60 62 let lines = "", dots = "", labels = "";
61 63 const usedY = [];
@@ -81,7 +83,7 @@ ${title ? `<figcaption class="chart-title">${esc(title)}</figcaption>` : ""}
81 83 <svg viewBox="0 0 ${W} ${H}" role="img" aria-label="${esc(title || yLabel)}">
82 84 <line x1="${ML}" y1="${MT + ih}" x2="${ML + iw}" y2="${MT + ih}" class="axis"/>
83 85 ${grid}${xt}${lines}${dots}${labels}
84 −<text x="${ML + iw / 2}" y="${H - 4}" class="axis-title" text-anchor="middle">layer</text>
86 +<text x="${ML + iw / 2}" y="${H - 4}" class="axis-title" text-anchor="middle">${esc(xTitle)}</text>
85 87 <text transform="rotate(-90 14 ${MT + ih / 2})" x="14" y="${MT + ih / 2}" class="axis-title" text-anchor="middle">${esc(yLabel)}</text>
86 88 </svg>
87 89 <div class="chart-legend">${legend}</div>
88 90