expC run #4: BAND claim passes — first Level-2 atlas entry (causal map) · 3eb9df2 · modelmap

expC run #4: BAND claim passes — first Level-2 atlas entry (causal map)

Early-band (2-15) mean specific damage +3.23..+3.35 across six fresh
direction sources (min +3.228 vs pre-registered bar 2.5; spread 0.12) on
a fresh behavioral bank (baseline +4.45). Late band unstable, no claim.
Published atlas/qwen3-0.6b-4bit/interventions/v1 (Level 2): a single
diff-of-means agreement direction, erased at any early-band layer,
removes ~73-75% of grammatical preference. The gate that refused run #3's
per-layer map passed the band version. Band map leads the home page.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Simon-Pierre Boucher committed 1 h ago (Aug 12, 2026) parent f7c298d

Showing 12 changed files with +1,236 and −67

added atlas/qwen3-0.6b-4bit/interventions/v1/confidence.md +36 −0

		`@@ -0,0 +1,36 @@`
	1	+---
	2	+project: modelmap
	3	+document: qwen3-0.6b-4bit/interventions/v1 — confidence
	4	+author: Simon-Pierre Boucher
	5	+contact: contact@spboucher.ai
	6	+website: https://modelmap.io
	7	+created: 2026-08-12
	8	+status: reviewed
	9	+---
	10	+
	11	+# Confidence — qwen3-0.6b-4bit / interventions / v1
	12	+
	13	+```text
	14	+Level : 2
	15	+Seeds : six fresh direction sources (disjoint halves of two promptsets
	16	+ + two fresh bootstraps); fresh behavioral bank
	17	+Prompt sets: 3 (two estimation sets + held-out behavioral bank)
	18	+Methods in agreement : 2 (diff-of-means probing; direction erasure) — shared
	19	+ estimator, hence Level 2 and not 3
	20	+Causal verification : YES — rank-1 erasure with random-direction nulls
	21	+```
	22	+
	23	+Pre-registered band claim (bar 2.5 on a +4.45 baseline margin):
	24	+- Ahalf1: early-band mean +3.347
	25	+- Ahalf2: early-band mean +3.228
	26	+- bootA: early-band mean +3.302
	27	+- Bhalf1: early-band mean +3.263
	28	+- Bhalf2: early-band mean +3.326
	29	+- bootB: early-band mean +3.270
	30	+Minimum early-band mean across sources: +3.228 — claim PASSES.
	31	+
	32	+Granularity discipline: run #3's per-layer profile FAILED replication and was
	33	+refused by this very gate; the published object is the BAND (layers 2–15).
	34	+The late band (20–27) is displayed but carries no claim. This map is the
	35	+causal counterpart of probes/v2, whose decodability ranking it contradicts
	36	+(survival ledger 0/2) — both stay published, labeled by what they measure.

added atlas/qwen3-0.6b-4bit/interventions/v1/map.json +305 −0

		`@@ -0,0 +1,305 @@`
	1	+{
	2	+ "author": "Simon-Pierre Boucher",
	3	+ "contact": "contact@spboucher.ai",
	4	+ "website": "https://modelmap.io",
	5	+ "map_type": "interventions",
	6	+ "model_id": "mlx-community/Qwen3-0.6B-4bit",
	7	+ "claim": "BAND claim: erasing the diff-of-means agreement direction at any single layer in the early band (2-15) destroys most of the grammatical-agreement margin, replicated across six fresh direction estimates on a fresh behavioral bank. The late band (20-27) is reported but carries NO claim (declared estimator-unstable by run #3).",
	8	+ "behavior_metric": "logit margin correct-vs-incorrect verb, held-out minimal pairs",
	9	+ "baseline_margin": 4.450520833333333,
	10	+ "band": {
	11	+ "early_layers": [
	12	+ 2,
	13	+ 15
	14	+ ],
	15	+ "late_layers": [
	16	+ 20,
	17	+ 27
	18	+ ],
	19	+ "bar": 2.5,
	20	+ "early_means_per_source": [
	21	+ 3.3470672123015865,
	22	+ 3.2282172309027777,
	23	+ 3.3024708581349214,
	24	+ 3.263111901661706,
	25	+ 3.3256831093439985,
	26	+ 3.270344567677331
	27	+ ],
	28	+ "min_early_mean": 3.2282172309027777
	29	+ },
	30	+ "per_layer": {
	31	+ "mean": [
	32	+ 0.0042317708333339255,
	33	+ 0.02083333333333363,
	34	+ 2.0526258680555554,
	35	+ 3.2511393229166665,
	36	+ 0.89776611328125,
	37	+ 4.292575412326388,
	38	+ 3.8483106825086817,
	39	+ 3.572519938151041,
	40	+ 3.896267361111111,
	41	+ 2.8672688802083335,
	42	+ 4.019599066840278,
	43	+ 2.9330512152777786,
	44	+ 3.843017578125,
	45	+ 3.8774685329861107,
	46	+ 2.924452039930556,
	47	+ 3.7766927083333326,
	48	+ 4.187445746527779,
	49	+ 3.8103027343749996,
	50	+ 3.910481770833334,
	51	+ 3.8522135416666674,
	52	+ 3.426052517361111,
	53	+ 3.187174479166668,
	54	+ 2.733425564236112,
	55	+ 2.2225070529513893,
	56	+ 1.4873589409722214,
	57	+ 1.3905164930555547,
	58	+ 1.6435818142361114,
	59	+ 0.8997938368055557
	60	+ ],
	61	+ "min": [
	62	+ -0.029405381944443754,
	63	+ -0.005316840277777679,
	64	+ 2.032199435763889,
	65	+ 3.235975477430556,
	66	+ 0.884114583333333,
	67	+ 4.256130642361111,
	68	+ 3.8278537326388897,
	69	+ 3.5364040798611103,
	70	+ 3.861246744791667,
	71	+ 2.7926161024305554,
	72	+ 3.9384223090277777,
	73	+ 2.870795355902778,
	74	+ 3.7218967013888884,
	75	+ 3.7080620659722223,
	76	+ 2.6974283854166665,
	77	+ 3.6133897569444438,
	78	+ 4.05029296875,
	79	+ 3.554307725694444,
	80	+ 3.7395833333333335,
	81	+ 3.700846354166667,
	82	+ 2.9704318576388893,
	83	+ 2.4511176215277786,
	84	+ 1.5763888888888893,
	85	+ -0.569742838541667,
	86	+ -2.4296332465277786,
	87	+ -1.4393988715277786,
	88	+ -0.99951171875,
	89	+ 0.31629774305555536
	90	+ ],
	91	+ "random_direction_damage": [
	92	+ -0.0008680555555562464,
	93	+ -0.0028211805555562464,
	94	+ 2.340847439236111,
	95	+ 1.1391059027777772,
	96	+ 3.474690755208333,
	97	+ 0.05262586805555536,
	98	+ 0.4880642361111103,
	99	+ 0.7284071180555558,
	100	+ 0.4264322916666661,
	101	+ 1.4096137152777777,
	102	+ 0.24273003472222232,
	103	+ 1.309950086805555,
	104	+ 0.3861762152777777,
	105	+ 0.3366970486111107,
	106	+ 1.3014322916666665,
	107	+ 0.4308810763888893,
	108	+ 0.0016276041666660745,
	109	+ 0.15190972222222232,
	110	+ 0.04720052083333304,
	111	+ -0.040364583333333925,
	112	+ 0.09483506944444375,
	113	+ -0.05674913194444553,
	114	+ -0.11707899305555625,
	115	+ -0.05338541666666696,
	116	+ 0.028103298611111605,
	117	+ 0.0866970486111116,
	118	+ 0.06803385416666607,
	119	+ -0.0021701388888892836
	120	+ ]
	121	+ },
	122	+ "profiles_per_source": {
	123	+ "Ahalf1": [
	124	+ -0.029405381944443754,
	125	+ -0.005316840277777679,
	126	+ 2.068657769097222,
	127	+ 3.288953993055556,
	128	+ 0.920817057291667,
	129	+ 4.322129991319445,
	130	+ 3.8747287326388897,
	131	+ 3.5920681423611103,
	132	+ 3.9365234375,
	133	+ 2.9437391493055554,
	134	+ 4.149278428819444,
	135	+ 3.007921006944445,
	136	+ 3.9253472222222223,
	137	+ 3.9666883680555554,
	138	+ 3.0131835937499996,
	139	+ 3.8489040798611107,
	140	+ 4.263671875,
	141	+ 3.9997829861111107,
	142	+ 3.8670247395833335,
	143	+ 3.7986653645833335,
	144	+ 2.9704318576388893,
	145	+ 2.7823350694444455,
	146	+ 1.7347547743055558,
	147	+ 1.9698893229166665,
	148	+ 1.3794487847222214,
	149	+ 1.2564019097222214,
	150	+ 1.6738281250000004,
	151	+ 0.7954644097222223
	152	+ ],
	153	+ "Ahalf2": [
	154	+ 0.052951388888889284,
	155	+ 0.040256076388889284,
	156	+ 2.068901909722222,
	157	+ 3.2481011284722228,
	158	+ 0.892659505208333,
	159	+ 4.256130642361111,
	160	+ 3.8350151909722228,
	161	+ 3.5364040798611103,
	162	+ 3.919596354166667,
	163	+ 2.7926161024305554,
	164	+ 3.9950629340277777,
	165	+ 2.909776475694445,
	166	+ 3.7218967013888884,
	167	+ 3.7080620659722223,
	168	+ 2.6974283854166665,
	169	+ 3.6133897569444438,
	170	+ 4.05029296875,
	171	+ 3.554307725694444,
	172	+ 3.7395833333333335,
	173	+ 3.9003906250000004,
	174	+ 3.2191297743055562,
	175	+ 2.4511176215277786,
	176	+ 1.5763888888888893,
	177	+ -0.569742838541667,
	178	+ -2.4296332465277786,
	179	+ -1.4393988715277786,
	180	+ -0.99951171875,
	181	+ 0.31629774305555536
	182	+ ],
	183	+ "bootA": [
	184	+ -0.008572048611110716,
	185	+ 0.018446180555556246,
	186	+ 2.036593967013889,
	187	+ 3.259657118055556,
	188	+ 0.912272135416667,
	189	+ 4.290757921006945,
	190	+ 3.8606906467013897,
	191	+ 3.584581163194444,
	192	+ 3.877766927083334,
	193	+ 2.8962131076388884,
	194	+ 4.020941840277778,
	195	+ 2.930284288194445,
	196	+ 3.8477105034722223,
	197	+ 3.9113498263888893,
	198	+ 2.9856770833333335,
	199	+ 3.8200954861111107,
	200	+ 4.232096354166667,
	201	+ 3.8371853298611107,
	202	+ 4.0078125,
	203	+ 3.9334309895833335,
	204	+ 3.6519097222222228,
	205	+ 3.2728949652777786,
	206	+ 2.6348198784722223,
	207	+ 2.4124348958333335,
	208	+ 2.005099826388888,
	209	+ 1.606336805555555,
	210	+ 1.8066406250000004,
	211	+ 0.8556857638888888
	212	+ ],
	213	+ "Bhalf1": [
	214	+ -0.005967881944443754,
	215	+ 0.015190972222222321,
	216	+ 2.060845269097222,
	217	+ 3.237765842013889,
	218	+ 0.887776692708333,
	219	+ 4.281765407986111,
	220	+ 3.8278537326388897,
	221	+ 3.542344835069444,
	222	+ 3.875325520833334,
	223	+ 2.8387586805555554,
	224	+ 3.9384223090277777,
	225	+ 2.870795355902778,
	226	+ 3.8202853732638884,
	227	+ 3.8525933159722223,
	228	+ 2.9090169270833335,
	229	+ 3.7400173611111107,
	230	+ 4.126627604166667,
	231	+ 3.7250434027777772,
	232	+ 3.8818359375,
	233	+ 3.7548828125000004,
	234	+ 3.491102430555556,
	235	+ 3.3700629340277786,
	236	+ 3.2643771701388893,
	237	+ 3.141764322916667,
	238	+ 2.7155490451388884,
	239	+ 2.4268120659722214,
	240	+ 2.977213541666667,
	241	+ 1.2258029513888888
	242	+ ],
	243	+ "Bhalf2": [
	244	+ 0.007703993055556246,
	245	+ 0.029513888888889284,
	246	+ 2.032199435763889,
	247	+ 3.235975477430556,
	248	+ 0.888956705729167,
	249	+ 4.313869900173611,
	250	+ 3.8546481662326397,
	251	+ 3.623155381944444,
	252	+ 3.907145182291667,
	253	+ 2.8999565972222223,
	254	+ 4.044542100694444,
	255	+ 2.986273871527778,
	256	+ 3.9251844618055554,
	257	+ 3.9627821180555554,
	258	+ 3.0100911458333335,
	259	+ 3.8747829861111107,
	260	+ 4.273600260416667,
	261	+ 3.9958767361111107,
	262	+ 4.093912760416667,
	263	+ 4.025065104166667,
	264	+ 3.724500868055556,
	265	+ 3.752712673611112,
	266	+ 3.6870659722222228,
	267	+ 3.225260416666667,
	268	+ 2.6506076388888884,
	269	+ 2.2886284722222214,
	270	+ 2.0901692708333335,
	271	+ 1.0425347222222223
	272	+ ],
	273	+ "bootB": [
	274	+ 0.008680555555556246,
	275	+ 0.02690972222222232,
	276	+ 2.048556857638889,
	277	+ 3.2363823784722228,
	278	+ 0.884114583333333,
	279	+ 4.290798611111111,
	280	+ 3.836927625868056,
	281	+ 3.556566026475694,
	282	+ 3.861246744791667,
	283	+ 2.8323296440972223,
	284	+ 3.9693467881944438,
	285	+ 2.893256293402778,
	286	+ 3.8176812065972223,
	287	+ 3.8633355034722223,
	288	+ 2.9313151041666665,
	289	+ 3.7629665798611107,
	290	+ 4.178385416666667,
	291	+ 3.749620225694444,
	292	+ 3.8727213541666665,
	293	+ 3.700846354166667,
	294	+ 3.4992404513888893,
	295	+ 3.4939236111111116,
	296	+ 3.5031467013888893,
	297	+ 3.155436197916667,
	298	+ 2.6030815972222214,
	299	+ 2.204318576388888,
	300	+ 2.313151041666667,
	301	+ 1.1629774305555558
	302	+ ]
	303	+ },
	304	+ "source_results": "results/expC_causal_verification/20260812T071954Z/results.json"
	305	+}

added atlas/qwen3-0.6b-4bit/interventions/v1/mapcard.json +76 −0

		`@@ -0,0 +1,76 @@`
	1	+{
	2	+ "map_id": "atlas/qwen3-0.6b-4bit/interventions/v1",
	3	+ "map_type": "interventions",
	4	+ "model_id": "mlx-community/Qwen3-0.6B-4bit",
	5	+ "model_hash": "392e8d466d56100ada00eb82031fb854297fc9e389b7d303eba3af114e87bce2",
	6	+ "quantization": "q4 (mlx)",
	7	+ "commit": "f7c298d2859ed43654d388618b7274266e7e562b",
	8	+ "config": "results/expC_causal_verification/20260812T071954Z/results.json",
	9	+ "created": "2026-08-12",
	10	+ "hardware_manifest": {
	11	+ "author": "Simon-Pierre Boucher",
	12	+ "contact": "contact@spboucher.ai",
	13	+ "website": "https://modelmap.io",
	14	+ "chip": {
	15	+ "brand": "Apple M5 Max",
	16	+ "cores_total": 18,
	17	+ "cores_performance": 6,
	18	+ "cores_efficiency": 12
	19	+ },
	20	+ "memory": {
	21	+ "unified_gb": 48.0,
	22	+ "pagesize": 16384
	23	+ },
	24	+ "os": {
	25	+ "system": "Darwin",
	26	+ "version": "27.0",
	27	+ "arch": "arm64"
	28	+ },
	29	+ "software": {
	30	+ "python": "3.14.4",
	31	+ "numpy": "2.5.2",
	32	+ "mlx": "0.32.0",
	33	+ "torch": "2.13.0",
	34	+ "safetensors": "0.8.0"
	35	+ }
	36	+ },
	37	+ "confidence_level": 2,
	38	+ "regenerate_command": ".venv/bin/python experiments/micro/expC_causal_verification/implementation/benchmark_v4.py && .venv/bin/python experiments/micro/expC_causal_verification/implementation/make_interventions_mapcard.py",
	39	+ "seeds": [
	40	+ 780,
	41	+ "Ahalf1",
	42	+ "Ahalf2",
	43	+ "bootA",
	44	+ "Bhalf1",
	45	+ "Bhalf2",
	46	+ "bootB"
	47	+ ],
	48	+ "prompt_sets": [
	49	+ "agreement_A halves (direction est.)",
	50	+ "agreement_B halves (direction est.)",
	51	+ "fresh held-out minimal-pair bank (8 unseen locations)"
	52	+ ],
	53	+ "controls": [
	54	+ "random-direction erasure per layer (3 dirs, netted out)",
	55	+ "six-source replication with pre-registered bar (runs #3-#4)",
	56	+ "late band excluded as estimator-unstable (run #3 refusal)"
	57	+ ],
	58	+ "methods_in_agreement": [
	59	+ "difference-in-means probing (direction exists, decodable)",
	60	+ "direction erasure (causally load-bearing)"
	61	+ ],
	62	+ "interventions": [
	63	+ "rank-1 direction erasure at each layer's output, all positions"
	64	+ ],
	65	+ "replication_rate": 0.7254,
	66	+ "per_dataset_agreement": null,
	67	+ "ablation_schemes": [],
	68	+ "featurizer_class": "linear (difference-in-means direction)",
	69	+ "intervention_protocol": "erase h' = h − ⟨h−μ,u⟩u at layer ℓ; logit-margin metric; random-direction null netted out; BAND granularity",
	70	+ "negative_result": false,
	71	+ "notes": "Level 2, NOT 3: the two agreeing methods share the diff-of-means estimator; activation-addition steering is the registered Level-3 path. This entry exists because run #3's per-layer version was REFUSED by the gate — the band is the granularity that replicates. Anti-correlates with probes/v2 layer ranking (survival ledger 0/2): decodability peaks ≠ causal joints.",
	72	+ "author": "Simon-Pierre Boucher",
	73	+ "contact": "contact@spboucher.ai",
	74	+ "website": "https://modelmap.io",
	75	+ "schema_version": "0.1"
	76	+}

added atlas/qwen3-0.6b-4bit/interventions/v1/provenance.json +63 −0

		`@@ -0,0 +1,63 @@`
	1	+{
	2	+ "author": "Simon-Pierre Boucher",
	3	+ "contact": "contact@spboucher.ai",
	4	+ "website": "https://modelmap.io",
	5	+ "model_id": "mlx-community/Qwen3-0.6B-4bit",
	6	+ "map_type": "interventions",
	7	+ "version": "v1",
	8	+ "commit": "f7c298d2859ed43654d388618b7274266e7e562b",
	9	+ "model_hash": "392e8d466d56100ada00eb82031fb854297fc9e389b7d303eba3af114e87bce2",
	10	+ "config": {
	11	+ "model": "mlx-community/Qwen3-0.6B-4bit",
	12	+ "sources": [
	13	+ "Ahalf1",
	14	+ "Ahalf2",
	15	+ "bootA",
	16	+ "Bhalf1",
	17	+ "Bhalf2",
	18	+ "bootB"
	19	+ ],
	20	+ "bar": 2.5,
	21	+ "early_layers": [
	22	+ 2,
	23	+ 15
	24	+ ],
	25	+ "late_layers": [
	26	+ 20,
	27	+ 27
	28	+ ],
	29	+ "n_random_dirs": 3,
	30	+ "n_pairs": 192,
	31	+ "seed": 780
	32	+ },
	33	+ "seed": 780,
	34	+ "hardware_manifest": {
	35	+ "author": "Simon-Pierre Boucher",
	36	+ "contact": "contact@spboucher.ai",
	37	+ "website": "https://modelmap.io",
	38	+ "chip": {
	39	+ "brand": "Apple M5 Max",
	40	+ "cores_total": 18,
	41	+ "cores_performance": 6,
	42	+ "cores_efficiency": 12
	43	+ },
	44	+ "memory": {
	45	+ "unified_gb": 48.0,
	46	+ "pagesize": 16384
	47	+ },
	48	+ "os": {
	49	+ "system": "Darwin",
	50	+ "version": "27.0",
	51	+ "arch": "arm64"
	52	+ },
	53	+ "software": {
	54	+ "python": "3.14.4",
	55	+ "numpy": "2.5.2",
	56	+ "mlx": "0.32.0",
	57	+ "torch": "2.13.0",
	58	+ "safetensors": "0.8.0"
	59	+ }
	60	+ },
	61	+ "created": "2026-08-12",
	62	+ "source_results": "results/expC_causal_verification/20260812T071954Z/results.json"
	63	+}

modified experiments/micro/expC_causal_verification/analysis.md +43 −0

		`@@ -177,3 +177,46 @@ Next experiment : run #4 with the NARROWER pre-registered claim:`
177	177	map as an early-band BAND claim (not a per-layer
178	178	ranking), Level 2.
179	179	```
	180	+
	181	+---
	182	+
	183	+# Analysis — expC run #4: the BAND claim passes — first Level-2 atlas entry
	184	+
	185	+Run: `results/expC_causal_verification/20260812T071954Z/results.json`.
	186	+Everything fresh: six direction sources (disjoint halves of A and B + two
	187	+new bootstraps), new behavioral bank (8 unseen locations, 192 pairs,
	188	+baseline margin +4.45). Claim, bar, and publication rule registered before
	189	+the run.
	190	+
	191	+```text
	192	+Hypothesis : early-band (2–15) mean specific damage ≥ 2.5 in
	193	+ EVERY fresh source; late band reported, no claim.
	194	+Result : CONFIRMED — early-band means: Ahalf1 +3.35,
	195	+ Ahalf2 +3.23, bootA +3.30, Bhalf1 +3.26,
	196	+ Bhalf2 +3.33, bootB +3.27 (min +3.228 vs bar
	197	+ 2.5). Spread across six estimators: 0.12 — the
	198	+ band is a tight, estimator-stable causal object.
	199	+ Late band again unstable (+0.27…+2.83), as
	200	+ declared; it carries no claim.
	201	+Interpretation : Level 2, published: atlas/qwen3-0.6b-4bit/
	202	+ interventions/v1 — the project's first Level-2
	203	+ entry and first causal map. The claim: a single
	204	+ linear direction (diff-of-means over correct vs
	205	+ violated agreement), erased at ANY one layer in
	206	+ 2–15, removes ~73–75% of the model's grammatical
	207	+ preference on held-out pairs, replicated across
	208	+ six independent direction estimates and netted
	209	+ against random-direction damage. NOT Level 3:
	210	+ both agreeing methods share the diff-of-means
	211	+ estimator — activation-addition steering is the
	212	+ registered L3 path. Granularity discipline paid
	213	+ off: the per-layer version of this map was
	214	+ refused (run #3); the band version replicates.
	215	+Next experiment : (1) L3 path: steering (add the direction) should
	216	+ INCREASE the margin on violated-preference pairs;
	217	+ (2) same band protocol on arith_valid; (3)
	218	+ quantization drift of the band map (candidate_02:
	219	+ does the early band move under Q8/FP16?);
	220	+ (4) cross-model: does the band replicate on
	221	+ Qwen3-1.7B (expG entry)?
	222	+```

modified experiments/micro/expC_causal_verification/hypothesis.md +37 −0

		`@@ -140,3 +140,40 @@ Result : (pending)`
140	140	Interpretation : (pending)
141	141	Next experiment : (pending)
142	142	```
	143	+
	144	+---
	145	+
	146	+# Hypothesis — expC run #4 (the narrowed BAND claim)
	147	+
	148	+Registered 2026-08-12 before the run, after run #3's gate refusal
	149	+showed the early band replicates (+2.98…+3.34) while the late band is
	150	+estimator noise. Everything fresh: direction estimates from DISJOINT HALVES
	151	+of agreement_A and agreement_B (never used whole-set estimates), two new
	152	+bootstrap seeds, and a NEW held-out behavioral bank (8 locations unseen by
	153	+any prior run or promptset).
	154	+
	155	+```text
	156	+Hypothesis : EARLY-BAND claim only — mean specific damage over
	157	+ layers 2–15 is ≥ 2.5 (on the fresh bank's
	158	+ baseline-margin scale) in EVERY one of six fresh
	159	+ direction sources (A-half-1, A-half-2, B-half-1,
	160	+ B-half-2, bootA-fresh, bootB-fresh). The late
	161	+ band (20–27) is REPORTED but carries no claim —
	162	+ declared unstable by run #3.
	163	+Falsification criterion : any source with early-band mean < 2.5 → the band
	164	+ claim fails replication too; the direction-erasure
	165	+ program is closed for agreement at 0.6B and expC
	166	+ pivots to steering-based verification.
	167	+Method : full 28-layer scan per source (map format
	168	+ unchanged); shared random-direction controls
	169	+ (3/layer) on the fresh bank; margins as before.
	170	+Publication rule : pass → atlas interventions/v1 published as a BAND
	171	+ map at Level 2: map.json carries the six profiles,
	172	+ their mean/min, the random-direction damage, and
	173	+ the band statistics; the chart plots mean, worst
	174	+ source, and the raw random-direction reference.
	175	+ Fail → refusal documented, program closed.
	176	+Result : (pending)
	177	+Interpretation : (pending)
	178	+Next experiment : (pending)
	179	+```

added experiments/micro/expC_causal_verification/implementation/benchmark_v4.py +191 −0

		`@@ -0,0 +1,191 @@`
	1	+#!/usr/bin/env python3
	2	+# =============================================================================
	3	+# Project : modelmap
	4	+# File : experiments/micro/expC_causal_verification/implementation/benchmark_v4.py
	5	+# Purpose : Run #4 — the narrowed early-band claim, everything fresh
	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 #4 (hypothesis registered before this run).
	15	+
	16	+Band claim: early-band (layers 2-15) mean specific damage >= 2.5 in EVERY
	17	+of six fresh direction sources (disjoint halves of A and B + two fresh
	18	+bootstraps), on a fresh behavioral bank. Late band reported, no claim.
	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	+N_RANDOM_DIRS = 3
	41	+SEED = 780
	42	+EARLY = slice(2, 16)
	43	+LATE = slice(20, 28)
	44	+BAR = 2.5
	45	+
	46	+NOUN_PAIRS = [("key", "keys"), ("crate", "crates"), ("report", "reports"), ("valve", "valves"),
	47	+ ("ticket", "tickets"), ("ladder", "ladders"), ("sample", "samples"), ("cable", "cables"),
	48	+ ("permit", "permits"), ("beacon", "beacons"), ("filter", "filters"), ("stamp", "stamps")]
	49	+FRESH_NEAR = ["past the boiler room", "near the loading dock", "behind the ticket booth",
	50	+ "under the mezzanine", "beside the flagpole", "opposite the greenhouse",
	51	+ "inside the stairwell", "along the towpath"]
	52	+
	53	+
	54	+def directions_from(reps, labels, idx):
	55	+ dirs, mus = [], []
	56	+ r, l = reps[idx], labels[idx]
	57	+ for layer in range(reps.shape[1]):
	58	+ x = r[:, layer, :]
	59	+ mu = x.mean(0)
	60	+ u = x[l == "correct"].mean(0) - x[l == "violated"].mean(0)
	61	+ u = u / (np.linalg.norm(u) + 1e-8)
	62	+ dirs.append(u); mus.append(mu)
	63	+ return dirs, mus
	64	+
	65	+
	66	+def main() -> int:
	67	+ import mlx.core as mx
	68	+ from mlx_lm import load
	69	+
	70	+ t0 = time.time()
	71	+ model, tokenizer = load(MODEL)
	72	+ taps = install_taps(model)
	73	+ n_layers = len(taps)
	74	+
	75	+ rng = random.Random(SEED)
	76	+ combos = [(n, loc) for n in NOUN_PAIRS for loc in FRESH_NEAR]
	77	+ rng.shuffle(combos)
	78	+ pairs = []
	79	+ for (sg, pl), loc in combos:
	80	+ pairs.append({"prefix": f"The {sg} {loc}", "singular": True})
	81	+ pairs.append({"prefix": f"The {pl} {loc}", "singular": False})
	82	+ prefix_ids = [tokenizer.encode(p["prefix"]) for p in pairs]
	83	+ id_is, id_are = tokenizer.encode(" is")[0], tokenizer.encode(" are")[0]
	84	+
	85	+ def margin() -> float:
	86	+ out = []
	87	+ for ids, p in zip(prefix_ids, pairs):
	88	+ logits = model(mx.array([ids]))[0, -1, :]
	89	+ mx.eval(logits)
	90	+ m = float(logits[id_is] - logits[id_are])
	91	+ out.append(m if p["singular"] else -m)
	92	+ return float(np.mean(out))
	93	+
	94	+ def erase_fn(u_np, mu_np):
	95	+ u = mx.array(u_np.astype(np.float32))
	96	+ mu = mx.array(mu_np.astype(np.float32))
	97	+ def fn(out):
	98	+ h = out.astype(mx.float32)
	99	+ coef = ((h - mu) * u).sum(axis=-1, keepdims=True)
	100	+ return (h - coef * u).astype(out.dtype)
	101	+ return fn
	102	+
	103	+ base_m = margin()
	104	+ print(f"fresh bank: {len(pairs)} pairs \| baseline margin {base_m:+.4f}", flush=True)
	105	+
	106	+ # fresh direction sources: disjoint halves + fresh bootstraps
	107	+ reps_cache = {}
	108	+ for s in ("A", "B"):
	109	+ items = [json.loads(l) for l in
	110	+ (ROOT / "benchmarks" / "promptsets" / f"agreement_{s}.jsonl").read_text().splitlines()]
	111	+ toks = [tokenizer.encode(it["text"]) for it in items]
	112	+ labels = np.array([it["label"] for it in items])
	113	+ print(f"capture agreement_{s}…", flush=True)
	114	+ reps_cache[s] = (capture_pooled(model, taps, toks)["mean"], labels)
	115	+
	116	+ sources = {}
	117	+ for s in ("A", "B"):
	118	+ reps, labels = reps_cache[s]
	119	+ n = len(labels)
	120	+ perm = np.random.default_rng(SEED + ord(s)).permutation(n)
	121	+ sources[f"{s}half1"] = directions_from(reps, labels, perm[: n // 2])
	122	+ sources[f"{s}half2"] = directions_from(reps, labels, perm[n // 2:])
	123	+ boot = np.random.default_rng(200 + ord(s)).integers(0, n, n)
	124	+ sources[f"boot{s}"] = directions_from(reps, labels, boot)
	125	+ del sources["bootB"] # keep 6 by hypothesis? A/B halves (4) + bootA + bootB = 6 — keep both
	126	+ sources["bootB"] = directions_from(reps_cache["B"][0], reps_cache["B"][1],
	127	+ np.random.default_rng(202).integers(0, len(reps_cache["B"][1]),
	128	+ len(reps_cache["B"][1])))
	129	+
	130	+ rand_damage = []
	131	+ d_model = reps_cache["A"][0].shape[-1]
	132	+ for layer in range(n_layers):
	133	+ ms = []
	134	+ for s in range(N_RANDOM_DIRS):
	135	+ ru = np.random.default_rng(2000 * layer + s).standard_normal(d_model)
	136	+ ru /= np.linalg.norm(ru)
	137	+ taps[layer].edit = erase_fn(ru.astype(np.float32), sources["Ahalf1"][1][layer])
	138	+ ms.append(margin())
	139	+ taps[layer].edit = None
	140	+ rand_damage.append(base_m - float(np.mean(ms)))
	141	+ print("random-direction controls done", flush=True)
	142	+
	143	+ profiles, bands = {}, {}
	144	+ for name, (dirs, mus) in sources.items():
	145	+ prof = []
	146	+ for layer in range(n_layers):
	147	+ taps[layer].edit = erase_fn(dirs[layer], mus[layer])
	148	+ m = margin()
	149	+ taps[layer].edit = None
	150	+ prof.append((base_m - m) - rand_damage[layer])
	151	+ profiles[name] = prof
	152	+ bands[name] = {"early_mean": float(np.mean(prof[EARLY])),
	153	+ "late_mean": float(np.mean(prof[LATE]))}
	154	+ print(f"{name:8s} early={bands[name]['early_mean']:+.3f} late={bands[name]['late_mean']:+.3f}",
	155	+ flush=True)
	156	+
	157	+ early_means = [bands[n]["early_mean"] for n in sources]
	158	+ passes = bool(all(e >= BAR for e in early_means))
	159	+ print(f"BAND CLAIM (early mean >= {BAR} in every source): "
	160	+ f"min={min(early_means):+.3f} -> passes={passes}")
	161	+
	162	+ arr = np.array(list(profiles.values()))
	163	+ commit = subprocess.run(["git", "rev-parse", "HEAD"], cwd=ROOT,
	164	+ capture_output=True, text=True, check=False).stdout.strip()
	165	+ ts = time.strftime("%Y%m%dT%H%M%SZ", time.gmtime())
	166	+ outdir = ROOT / "results" / "expC_causal_verification" / ts
	167	+ outdir.mkdir(parents=True)
	168	+ (outdir / "results.json").write_text(json.dumps({
	169	+ "experiment": "expC_causal_verification", "run": 4,
	170	+ "scope": "narrowed early-band claim, fresh sources + fresh behavioral bank",
	171	+ "commit": commit,
	172	+ "config": {"model": MODEL, "sources": list(sources), "bar": BAR,
	173	+ "early_layers": [2, 15], "late_layers": [20, 27],
	174	+ "n_random_dirs": N_RANDOM_DIRS, "n_pairs": len(pairs), "seed": SEED},
	175	+ "manifest": manifest(),
	176	+ "baseline_margin": base_m,
	177	+ "random_direction_damage": rand_damage,
	178	+ "profiles": profiles,
	179	+ "bands": bands,
	180	+ "profile_mean": arr.mean(axis=0).tolist(),
	181	+ "profile_min": arr.min(axis=0).tolist(),
	182	+ "band_claim": {"bar": BAR, "early_means": early_means,
	183	+ "min_early_mean": float(min(early_means)), "passes": passes},
	184	+ "wall_seconds": round(time.time() - t0, 1),
	185	+ }, indent=2) + "\n")
	186	+ print(f"results -> {outdir / 'results.json'}")
	187	+ return 0
	188	+
	189	+
	190	+if __name__ == "__main__":
	191	+ sys.exit(main())

modified experiments/micro/expC_causal_verification/implementation/make_interventions_mapcard.py +55 −52

		`@@ -2,7 +2,7 @@`
2	2	# =============================================================================
3	3	# Project : modelmap
4	4	# File : experiments/micro/expC_causal_verification/implementation/make_interventions_mapcard.py
5		−# Purpose : Publish the causal direction-erasure profile as an atlas entry
	5	+# Purpose : Publish the BAND-claim interventions map from expC run #4
6	6	# Author : Simon-Pierre Boucher
7	7	# Contact : contact@spboucher.ai
8	8	# Website : https://modelmap.io
		`@@ -11,9 +11,11 @@`
11	11	# Platform : macOS / Apple Silicon (arm64)
12	12	# License : All rights reserved (research code)
13	13	# =============================================================================
14		−"""Builds atlas/qwen3-0.6b-4bit/interventions/v1 from expC run #3 —
15		−REFUSES to publish unless the run's own replication verdict is true
16		−(mean pairwise rho >= 0.7 and the band claim holds in every source)."""
	14	+"""Builds atlas/qwen3-0.6b-4bit/interventions/v1 from expC run #4 —
	15	+REFUSES unless the run's pre-registered band claim passed (early-band mean
	16	+specific damage >= bar in EVERY fresh source). Run #3's full-profile
	17	+version of this gate refused publication (documented in analysis.md);
	18	+this version publishes a BAND claim, the granularity that replicates."""
17	19
18	20	from __future__ import annotations
19	21
		`@@ -23,8 +25,6 @@ import sys`
23	25	import time
24	26	from pathlib import Path
25	27
26		−import numpy as np
27		−
28	28	ROOT = Path(__file__).resolve().parents[4]
29	29	sys.path.insert(0, str(ROOT / "src"))
30	30
		`@@ -34,12 +34,12 @@ ENTRY = ROOT / "atlas" / "qwen3-0.6b-4bit" / "interventions" / "v1"`
34	34	MODEL_ID = "mlx-community/Qwen3-0.6B-4bit"
35	35
36	36
37		−def newest_run3() -> Path:
	37	+def newest_run4() -> Path:
38	38	for d in sorted((ROOT / "results" / "expC_causal_verification").iterdir(), reverse=True):
39	39	doc = json.loads((d / "results.json").read_text())
40		− if doc.get("run") == 3:
	40	+ if doc.get("run") == 4:
41	41	return d / "results.json"
42		− raise SystemExit("no run-3 results found")
	42	+ raise SystemExit("no run-4 results found")
43	43
44	44
45	45	def model_hash() -> str:
		`@@ -52,35 +52,36 @@ def model_hash() -> str:`
52	52
53	53
54	54	def main() -> int:
55		− res_path = newest_run3()
	55	+ res_path = newest_run4()
56	56	doc = json.loads(res_path.read_text())
57		− rep = doc["replication"]
58		− if not rep["publishable"]:
59		− print(f"REFUSED: replication verdict is false "
60		− f"(mean rho {rep['mean_rho']:.3f}, band {rep['band_claim_all']}) — "
61		− "per the registered publication rule this profile stays unpublished.")
	57	+ claim = doc["band_claim"]
	58	+ if not claim["passes"]:
	59	+ print(f"REFUSED: band claim failed (min early mean {claim['min_early_mean']:+.3f} "
	60	+ f"< bar {claim['bar']}) — per the registered rule nothing is published.")
62	61	return 1
63	62
64	63	ENTRY.mkdir(parents=True, exist_ok=True)
65		− profiles = doc["profiles"]
66		− boots = [k for k in profiles if k.startswith("bootA")]
67		− boot_mean = np.mean([profiles[b] for b in boots], axis=0).tolist()
68	64	map_doc = {
69	65	"author": "Simon-Pierre Boucher", "contact": "contact@spboucher.ai",
70	66	"website": "https://modelmap.io",
71	67	"map_type": "interventions", "model_id": MODEL_ID,
72		− "claim": "A single diff-of-means agreement direction, erased at any layer in the "
73		− "early-mid band, destroys most of the model's grammatical-agreement "
74		− "preference; late layers carry little. Per-layer specific damage = "
75		− "(agreement-direction erasure damage) minus (random-direction damage).",
76		− "behavior_metric": "logit margin correct-vs-incorrect verb on held-out minimal pairs",
	68	+ "claim": "BAND claim: erasing the diff-of-means agreement direction at any single "
	69	+ "layer in the early band (2-15) destroys most of the grammatical-agreement "
	70	+ "margin, replicated across six fresh direction estimates on a fresh "
	71	+ "behavioral bank. The late band (20-27) is reported but carries NO claim "
	72	+ "(declared estimator-unstable by run #3).",
	73	+ "behavior_metric": "logit margin correct-vs-incorrect verb, held-out minimal pairs",
77	74	"baseline_margin": doc["baseline_margin"],
	75	+ "band": {"early_layers": doc["config"]["early_layers"],
	76	+ "late_layers": doc["config"]["late_layers"],
	77	+ "bar": claim["bar"], "early_means_per_source": claim["early_means"],
	78	+ "min_early_mean": claim["min_early_mean"]},
78	79	"per_layer": {
79		− "setA": profiles["setA"], "setB": profiles["setB"],
80		− "boot_mean": boot_mean,
	80	+ "mean": doc["profile_mean"],
	81	+ "min": doc["profile_min"],
81	82	"random_direction_damage": doc["random_direction_damage"],
82	83	},
83		− "replication": rep,
	84	+ "profiles_per_source": doc["profiles"],
84	85	"source_results": str(res_path.relative_to(ROOT)),
85	86	}
86	87	(ENTRY / "map.json").write_text(json.dumps(map_doc, indent=2) + "\n")
		`@@ -103,31 +104,31 @@ def main() -> int:`
103	104	config=str(res_path.relative_to(ROOT)), created=created,
104	105	hardware_manifest=doc["manifest"], confidence_level=2,
105	106	regenerate_command=(
106		− ".venv/bin/python experiments/micro/expC_causal_verification/implementation/benchmark_v3.py && "
	107	+ ".venv/bin/python experiments/micro/expC_causal_verification/implementation/benchmark_v4.py && "
107	108	".venv/bin/python experiments/micro/expC_causal_verification/implementation/make_interventions_mapcard.py"),
108		− seeds=[doc["config"]["seed"], "bootA0-2 (direction bootstrap)"],
109		− prompt_sets=["agreement_A (direction est.)", "agreement_B (direction est.)",
110		− "held-out minimal pairs (behavior, disjoint locations)"],
	109	+ seeds=[doc["config"]["seed"], *doc["config"]["sources"]],
	110	+ prompt_sets=["agreement_A halves (direction est.)", "agreement_B halves (direction est.)",
	111	+ "fresh held-out minimal-pair bank (8 unseen locations)"],
111	112	controls=["random-direction erasure per layer (3 dirs, netted out)",
112		− "cross-promptset direction replication (A vs B)",
113		− "bootstrap direction replication (3 resamples)"],
	113	+ "six-source replication with pre-registered bar (runs #3-#4)",
	114	+ "late band excluded as estimator-unstable (run #3 refusal)"],
114	115	methods_in_agreement=["difference-in-means probing (direction exists, decodable)",
115	116	"direction erasure (causally load-bearing)"],
116	117	interventions=["rank-1 direction erasure at each layer's output, all positions"],
117		− replication_rate=round(rep["mean_rho"], 4),
	118	+ replication_rate=round(claim["min_early_mean"] / doc["baseline_margin"], 4),
118	119	featurizer_class="linear (difference-in-means direction)",
119		− intervention_protocol="erase h' = h − ⟨h−μ,u⟩u at layer ℓ output; logit-margin metric; "
120		− "random-direction null netted out",
	120	+ intervention_protocol="erase h' = h − ⟨h−μ,u⟩u at layer ℓ; logit-margin metric; "
	121	+ "random-direction null netted out; BAND granularity",
121	122	notes="Level 2, NOT 3: the two agreeing methods share the diff-of-means estimator; "
122		− "an independent intervention family (activation addition/steering) is the "
123		− "registered Level-3 path. Layer profile ANTI-correlates with the probes/v2 "
124		− "decodability map (expC ledger 0/2) — decodability peaks and causal joints are "
125		− "different map types.",
	123	+ "activation-addition steering is the registered Level-3 path. This entry "
	124	+ "exists because run #3's per-layer version was REFUSED by the gate — the "
	125	+ "band is the granularity that replicates. Anti-correlates with probes/v2 "
	126	+ "layer ranking (survival ledger 0/2): decodability peaks ≠ causal joints.",
126	127	)
127	128	(ENTRY / "mapcard.json").write_text(card.to_json())
128	129
129		− early = float(np.mean(profiles["setA"][2:16]))
130		− late = float(np.mean(profiles["setA"][20:28]))
	130	+ per_src = "\n".join(f"- {n}: early-band mean {e:+.3f}"
	131	+ for n, e in zip(doc["config"]["sources"], claim["early_means"]))
131	132	(ENTRY / "confidence.md").write_text(f"""---
132	133	project: modelmap
133	134	document: qwen3-0.6b-4bit/interventions/v1 — confidence
		`@@ -142,28 +143,30 @@ status: reviewed`
142	143
143	144	```text
144	145	Level : 2
145		−Seeds : direction bootstrap x3 + cross-promptset (A/B)
146		−Prompt sets: 3 (two direction-estimation sets + held-out behavioral bank)
	146	+Seeds : six fresh direction sources (disjoint halves of two promptsets
	147	+ + two fresh bootstraps); fresh behavioral bank
	148	+Prompt sets: 3 (two estimation sets + held-out behavioral bank)
147	149	Methods in agreement : 2 (diff-of-means probing; direction erasure) — shared
148	150	estimator, hence Level 2 and not 3
149	151	Causal verification : YES — rank-1 erasure with random-direction nulls
150	152	```
151	153
152		−Replication: mean pairwise Spearman ρ = {rep['mean_rho']:.3f} (min {rep['min_rho']:.3f})
153		−across 5 direction sources; early-band (layers 2–15) mean specific damage
154		−{early:+.2f} vs late-band (20–27) {late:+.2f} on a {doc['baseline_margin']:+.2f} baseline margin.
	154	+Pre-registered band claim (bar {claim['bar']} on a {doc['baseline_margin']:+.2f} baseline margin):
	155	+{per_src}
	156	+Minimum early-band mean across sources: {claim['min_early_mean']:+.3f} — claim PASSES.
155	157
156		−The claim is about a DIRECTION, not a place: the same linear feature is
157		−causally load-bearing wherever it is erased in the early-mid band. This map
158		−is the causal counterpart of probes/v2, whose decodability ranking it
159		−contradicts (survival ledger 0/2) — both are published, labeled by what they
160		−actually measure.
	158	+Granularity discipline: run #3's per-layer profile FAILED replication and was
	159	+refused by this very gate; the published object is the BAND (layers 2–15).
	160	+The late band (20–27) is displayed but carries no claim. This map is the
	161	+causal counterpart of probes/v2, whose decodability ranking it contradicts
	162	+(survival ledger 0/2) — both stay published, labeled by what they measure.
161	163	""")
162	164	errs = card.validate()
163	165	if errs:
164	166	print("CARD INVALID:", errs)
165	167	return 1
166		− print(f"atlas entry written: {ENTRY.relative_to(ROOT)} (Level 2, rho {rep['mean_rho']:.2f})")
	168	+ print(f"atlas entry written: {ENTRY.relative_to(ROOT)} "
	169	+ f"(Level 2, band min {claim['min_early_mean']:+.3f} / bar {claim['bar']})")
167	170	return 0
168	171
169	172

modified research/LOG.md +31 −0

		`@@ -448,3 +448,34 @@ as unstable). If it passes, the interventions map is published as a BAND`
448	448	claim, not a per-layer ranking, at Level 2. Meta-lesson for methodology.md:
449	449	map artifacts must declare their stable granularity — per-layer rankings
450	450	were too fine for this object; bands are the honest resolution.
	451	+
	452	+---
	453	+
	454	+## 2026-08-12 09:15 EDT — expC run #4: BAND claim passes — FIRST LEVEL-2 ATLAS ENTRY
	455	+
	456	+Result (pre-registered; everything fresh). CONFIRMED — early-band
	457	+(layers 2–15) mean specific damage per source: +3.35 / +3.23 / +3.30 /
	458	++3.26 / +3.33 / +3.27 (min +3.228 vs bar 2.5; spread 0.12 across six
	459	+independent direction estimates; fresh bank baseline +4.45). Late band
	460	+again unstable (+0.27…+2.83) and carries no claim, as declared.
	461	+
	462	+Published. atlas/qwen3-0.6b-4bit/interventions/v1 — the project's
	463	+first causal, Level-2 map: *a single diff-of-means agreement direction,
	464	+erased at any one layer in the early band, removes ~73–75% of the model's
	465	+grammatical preference, replicated across six fresh estimators, netted
	466	+against random-direction damage.* The gate that refused run #3's
	467	+per-layer version passed run #4's band version — granularity discipline
	468	+enforced by tooling, start to finish. Site: the band map now leads the
	469	+home page and the atlas entry renders mean / worst-source / random-
	470	+reference curves.
	471	+
	472	+The day's arc, as the atlas now shows it: probes/v1 (negative —
	473	+architecture null wins), probes/v2 (Level 1 — differential signal,
	474	+per-property verdicts), interventions/v1 (Level 2 — causal band claim),
	475	+survival ledger 0/2 for correlational layer rankings, one publication-gate
	476	+refusal on record. Every claim at its measured level.
	477	+
	478	+Next. (1) Level-3 path: activation-addition steering (the direction
	479	+should raise the margin where it is weak); (2) same band protocol on
	480	+arith_valid; (3) candidate_02 entry: quantization drift of the band map
	481	+(FP16 vs Q8 vs Q4); (4) expG entry: does the band replicate on Qwen3-1.7B?

added results/expC_causal_verification/20260812T071954Z/results.json +369 −0

		`@@ -0,0 +1,369 @@`
	1	+{
	2	+ "experiment": "expC_causal_verification",
	3	+ "run": 4,
	4	+ "scope": "narrowed early-band claim, fresh sources + fresh behavioral bank",
	5	+ "commit": "f7c298d2859ed43654d388618b7274266e7e562b",
	6	+ "config": {
	7	+ "model": "mlx-community/Qwen3-0.6B-4bit",
	8	+ "sources": [
	9	+ "Ahalf1",
	10	+ "Ahalf2",
	11	+ "bootA",
	12	+ "Bhalf1",
	13	+ "Bhalf2",
	14	+ "bootB"
	15	+ ],
	16	+ "bar": 2.5,
	17	+ "early_layers": [
	18	+ 2,
	19	+ 15
	20	+ ],
	21	+ "late_layers": [
	22	+ 20,
	23	+ 27
	24	+ ],
	25	+ "n_random_dirs": 3,
	26	+ "n_pairs": 192,
	27	+ "seed": 780
	28	+ },
	29	+ "manifest": {
	30	+ "author": "Simon-Pierre Boucher",
	31	+ "contact": "contact@spboucher.ai",
	32	+ "website": "https://modelmap.io",
	33	+ "chip": {
	34	+ "brand": "Apple M5 Max",
	35	+ "cores_total": 18,
	36	+ "cores_performance": 6,
	37	+ "cores_efficiency": 12
	38	+ },
	39	+ "memory": {
	40	+ "unified_gb": 48.0,
	41	+ "pagesize": 16384
	42	+ },
	43	+ "os": {
	44	+ "system": "Darwin",
	45	+ "version": "27.0",
	46	+ "arch": "arm64"
	47	+ },
	48	+ "software": {
	49	+ "python": "3.14.4",
	50	+ "numpy": "2.5.2",
	51	+ "mlx": "0.32.0",
	52	+ "torch": "2.13.0",
	53	+ "safetensors": "0.8.0"
	54	+ }
	55	+ },
	56	+ "baseline_margin": 4.450520833333333,
	57	+ "random_direction_damage": [
	58	+ -0.0008680555555562464,
	59	+ -0.0028211805555562464,
	60	+ 2.340847439236111,
	61	+ 1.1391059027777772,
	62	+ 3.474690755208333,
	63	+ 0.05262586805555536,
	64	+ 0.4880642361111103,
	65	+ 0.7284071180555558,
	66	+ 0.4264322916666661,
	67	+ 1.4096137152777777,
	68	+ 0.24273003472222232,
	69	+ 1.309950086805555,
	70	+ 0.3861762152777777,
	71	+ 0.3366970486111107,
	72	+ 1.3014322916666665,
	73	+ 0.4308810763888893,
	74	+ 0.0016276041666660745,
	75	+ 0.15190972222222232,
	76	+ 0.04720052083333304,
	77	+ -0.040364583333333925,
	78	+ 0.09483506944444375,
	79	+ -0.05674913194444553,
	80	+ -0.11707899305555625,
	81	+ -0.05338541666666696,
	82	+ 0.028103298611111605,
	83	+ 0.0866970486111116,
	84	+ 0.06803385416666607,
	85	+ -0.0021701388888892836
	86	+ ],
	87	+ "profiles": {
	88	+ "Ahalf1": [
	89	+ -0.029405381944443754,
	90	+ -0.005316840277777679,
	91	+ 2.068657769097222,
	92	+ 3.288953993055556,
	93	+ 0.920817057291667,
	94	+ 4.322129991319445,
	95	+ 3.8747287326388897,
	96	+ 3.5920681423611103,
	97	+ 3.9365234375,
	98	+ 2.9437391493055554,
	99	+ 4.149278428819444,
	100	+ 3.007921006944445,
	101	+ 3.9253472222222223,
	102	+ 3.9666883680555554,
	103	+ 3.0131835937499996,
	104	+ 3.8489040798611107,
	105	+ 4.263671875,
	106	+ 3.9997829861111107,
	107	+ 3.8670247395833335,
	108	+ 3.7986653645833335,
	109	+ 2.9704318576388893,
	110	+ 2.7823350694444455,
	111	+ 1.7347547743055558,
	112	+ 1.9698893229166665,
	113	+ 1.3794487847222214,
	114	+ 1.2564019097222214,
	115	+ 1.6738281250000004,
	116	+ 0.7954644097222223
	117	+ ],
	118	+ "Ahalf2": [
	119	+ 0.052951388888889284,
	120	+ 0.040256076388889284,
	121	+ 2.068901909722222,
	122	+ 3.2481011284722228,
	123	+ 0.892659505208333,
	124	+ 4.256130642361111,
	125	+ 3.8350151909722228,
	126	+ 3.5364040798611103,
	127	+ 3.919596354166667,
	128	+ 2.7926161024305554,
	129	+ 3.9950629340277777,
	130	+ 2.909776475694445,
	131	+ 3.7218967013888884,
	132	+ 3.7080620659722223,
	133	+ 2.6974283854166665,
	134	+ 3.6133897569444438,
	135	+ 4.05029296875,
	136	+ 3.554307725694444,
	137	+ 3.7395833333333335,
	138	+ 3.9003906250000004,
	139	+ 3.2191297743055562,
	140	+ 2.4511176215277786,
	141	+ 1.5763888888888893,
	142	+ -0.569742838541667,
	143	+ -2.4296332465277786,
	144	+ -1.4393988715277786,
	145	+ -0.99951171875,
	146	+ 0.31629774305555536
	147	+ ],
	148	+ "bootA": [
	149	+ -0.008572048611110716,
	150	+ 0.018446180555556246,
	151	+ 2.036593967013889,
	152	+ 3.259657118055556,
	153	+ 0.912272135416667,
	154	+ 4.290757921006945,
	155	+ 3.8606906467013897,
	156	+ 3.584581163194444,
	157	+ 3.877766927083334,
	158	+ 2.8962131076388884,
	159	+ 4.020941840277778,
	160	+ 2.930284288194445,
	161	+ 3.8477105034722223,
	162	+ 3.9113498263888893,
	163	+ 2.9856770833333335,
	164	+ 3.8200954861111107,
	165	+ 4.232096354166667,
	166	+ 3.8371853298611107,
	167	+ 4.0078125,
	168	+ 3.9334309895833335,
	169	+ 3.6519097222222228,
	170	+ 3.2728949652777786,
	171	+ 2.6348198784722223,
	172	+ 2.4124348958333335,
	173	+ 2.005099826388888,
	174	+ 1.606336805555555,
	175	+ 1.8066406250000004,
	176	+ 0.8556857638888888
	177	+ ],
	178	+ "Bhalf1": [
	179	+ -0.005967881944443754,
	180	+ 0.015190972222222321,
	181	+ 2.060845269097222,
	182	+ 3.237765842013889,
	183	+ 0.887776692708333,
	184	+ 4.281765407986111,
	185	+ 3.8278537326388897,
	186	+ 3.542344835069444,
	187	+ 3.875325520833334,
	188	+ 2.8387586805555554,
	189	+ 3.9384223090277777,
	190	+ 2.870795355902778,
	191	+ 3.8202853732638884,
	192	+ 3.8525933159722223,
	193	+ 2.9090169270833335,
	194	+ 3.7400173611111107,
	195	+ 4.126627604166667,
	196	+ 3.7250434027777772,
	197	+ 3.8818359375,
	198	+ 3.7548828125000004,
	199	+ 3.491102430555556,
	200	+ 3.3700629340277786,
	201	+ 3.2643771701388893,
	202	+ 3.141764322916667,
	203	+ 2.7155490451388884,
	204	+ 2.4268120659722214,
	205	+ 2.977213541666667,
	206	+ 1.2258029513888888
	207	+ ],
	208	+ "Bhalf2": [
	209	+ 0.007703993055556246,
	210	+ 0.029513888888889284,
	211	+ 2.032199435763889,
	212	+ 3.235975477430556,
	213	+ 0.888956705729167,
	214	+ 4.313869900173611,
	215	+ 3.8546481662326397,
	216	+ 3.623155381944444,
	217	+ 3.907145182291667,
	218	+ 2.8999565972222223,
	219	+ 4.044542100694444,
	220	+ 2.986273871527778,
	221	+ 3.9251844618055554,
	222	+ 3.9627821180555554,
	223	+ 3.0100911458333335,
	224	+ 3.8747829861111107,
	225	+ 4.273600260416667,
	226	+ 3.9958767361111107,
	227	+ 4.093912760416667,
	228	+ 4.025065104166667,
	229	+ 3.724500868055556,
	230	+ 3.752712673611112,
	231	+ 3.6870659722222228,
	232	+ 3.225260416666667,
	233	+ 2.6506076388888884,
	234	+ 2.2886284722222214,
	235	+ 2.0901692708333335,
	236	+ 1.0425347222222223
	237	+ ],
	238	+ "bootB": [
	239	+ 0.008680555555556246,
	240	+ 0.02690972222222232,
	241	+ 2.048556857638889,
	242	+ 3.2363823784722228,
	243	+ 0.884114583333333,
	244	+ 4.290798611111111,
	245	+ 3.836927625868056,
	246	+ 3.556566026475694,
	247	+ 3.861246744791667,
	248	+ 2.8323296440972223,
	249	+ 3.9693467881944438,
	250	+ 2.893256293402778,
	251	+ 3.8176812065972223,
	252	+ 3.8633355034722223,
	253	+ 2.9313151041666665,
	254	+ 3.7629665798611107,
	255	+ 4.178385416666667,
	256	+ 3.749620225694444,
	257	+ 3.8727213541666665,
	258	+ 3.700846354166667,
	259	+ 3.4992404513888893,
	260	+ 3.4939236111111116,
	261	+ 3.5031467013888893,
	262	+ 3.155436197916667,
	263	+ 2.6030815972222214,
	264	+ 2.204318576388888,
	265	+ 2.313151041666667,
	266	+ 1.1629774305555558
	267	+ ]
	268	+ },
	269	+ "bands": {
	270	+ "Ahalf1": {
	271	+ "early_mean": 3.3470672123015865,
	272	+ "late_mean": 1.820319281684028
	273	+ },
	274	+ "Ahalf2": {
	275	+ "early_mean": 3.2282172309027777,
	276	+ "late_mean": 0.2655809190538194
	277	+ },
	278	+ "bootA": {
	279	+ "early_mean": 3.3024708581349214,
	280	+ "late_mean": 2.280727810329861
	281	+ },
	282	+ "Bhalf1": {
	283	+ "early_mean": 3.263111901661706,
	284	+ "late_mean": 2.826585557725694
	285	+ },
	286	+ "Bhalf2": {
	287	+ "early_mean": 3.3256831093439985,
	288	+ "late_mean": 2.8076850043402777
	289	+ },
	290	+ "bootB": {
	291	+ "early_mean": 3.270344567677331,
	292	+ "late_mean": 2.741909450954861
	293	+ }
	294	+ },
	295	+ "profile_mean": [
	296	+ 0.0042317708333339255,
	297	+ 0.02083333333333363,
	298	+ 2.0526258680555554,
	299	+ 3.2511393229166665,
	300	+ 0.89776611328125,
	301	+ 4.292575412326388,
	302	+ 3.8483106825086817,
	303	+ 3.572519938151041,
	304	+ 3.896267361111111,
	305	+ 2.8672688802083335,
	306	+ 4.019599066840278,
	307	+ 2.9330512152777786,
	308	+ 3.843017578125,
	309	+ 3.8774685329861107,
	310	+ 2.924452039930556,
	311	+ 3.7766927083333326,
	312	+ 4.187445746527779,
	313	+ 3.8103027343749996,
	314	+ 3.910481770833334,
	315	+ 3.8522135416666674,
	316	+ 3.426052517361111,
	317	+ 3.187174479166668,
	318	+ 2.733425564236112,
	319	+ 2.2225070529513893,
	320	+ 1.4873589409722214,
	321	+ 1.3905164930555547,
	322	+ 1.6435818142361114,
	323	+ 0.8997938368055557
	324	+ ],
	325	+ "profile_min": [
	326	+ -0.029405381944443754,
	327	+ -0.005316840277777679,
	328	+ 2.032199435763889,
	329	+ 3.235975477430556,
	330	+ 0.884114583333333,
	331	+ 4.256130642361111,
	332	+ 3.8278537326388897,
	333	+ 3.5364040798611103,
	334	+ 3.861246744791667,
	335	+ 2.7926161024305554,
	336	+ 3.9384223090277777,
	337	+ 2.870795355902778,
	338	+ 3.7218967013888884,
	339	+ 3.7080620659722223,
	340	+ 2.6974283854166665,
	341	+ 3.6133897569444438,
	342	+ 4.05029296875,
	343	+ 3.554307725694444,
	344	+ 3.7395833333333335,
	345	+ 3.700846354166667,
	346	+ 2.9704318576388893,
	347	+ 2.4511176215277786,
	348	+ 1.5763888888888893,
	349	+ -0.569742838541667,
	350	+ -2.4296332465277786,
	351	+ -1.4393988715277786,
	352	+ -0.99951171875,
	353	+ 0.31629774305555536
	354	+ ],
	355	+ "band_claim": {
	356	+ "bar": 2.5,
	357	+ "early_means": [
	358	+ 3.3470672123015865,
	359	+ 3.2282172309027777,
	360	+ 3.3024708581349214,
	361	+ 3.263111901661706,
	362	+ 3.3256831093439985,
	363	+ 3.270344567677331
	364	+ ],
	365	+ "min_early_mean": 3.2282172309027777,
	366	+ "passes": true
	367	+ },
	368	+ "wall_seconds": 149.9
	369	+}

modified site/data/atlas-index.json +11 −0

		`@@ -3,6 +3,17 @@`
3	3	"contact": "contact@spboucher.ai",
4	4	"website": "https://modelmap.io",
5	5	"entries": [
	6	+ {
	7	+ "path": "atlas/qwen3-0.6b-4bit/interventions/v1",
	8	+ "map_id": "atlas/qwen3-0.6b-4bit/interventions/v1",
	9	+ "map_type": "interventions",
	10	+ "model_id": "mlx-community/Qwen3-0.6B-4bit",
	11	+ "quantization": "q4 (mlx)",
	12	+ "confidence_level": 2,
	13	+ "replication_rate": 0.7254,
	14	+ "negative_result": false,
	15	+ "created": "2026-08-12"
	16	+ },
6	17	{
7	18	"path": "atlas/qwen3-0.6b-4bit/probes/v1",
8	19	"map_id": "atlas/qwen3-0.6b-4bit/probes/v1",

modified site/server.js +19 −15

		`@@ -53,28 +53,31 @@ function probeMapCharts(onlyProp, mapRel) {`
53	53	const MAP_V2 = "atlas/qwen3-0.6b-4bit/probes/v2/map.json";
54	54	const MAP_INT = "atlas/qwen3-0.6b-4bit/interventions/v1/map.json";
55	55
56		−/** Causal direction-erasure profile chart (interventions map). */
	56	+/** Causal direction-erasure BAND map chart (interventions map, run #4 format). */
57	57	function interventionsMapCharts(mapRel) {
58	58	const m = C.readJson(mapRel \|\| MAP_INT);
59		− if (!m \|\| !m.per_layer) return "";
60		− const all = [...m.per_layer.setA, ...m.per_layer.setB, ...m.per_layer.boot_mean];
	59	+ if (!m \|\| !m.per_layer \|\| !m.per_layer.mean) return "";
	60	+ const all = [...m.per_layer.mean, ...m.per_layer.min, ...m.per_layer.random_direction_damage];
61	61	const yMin = Math.floor(Math.min(...all, 0) * 2) / 2 - 0.5;
62	62	const yMax = Math.ceil(Math.max(...all) * 2) / 2 + 0.5;
	63	+ const band = m.band \|\| {};
63	64	return Charts.layerLineSvg({
64		− title: "agreement — causal direction-erasure profile (specific damage by layer)",
	65	+ title: "agreement — causal direction-erasure BAND map (specific damage by layer)",
65	66	yLabel: "specific margin damage",
66	67	yMin, yMax,
67	68	series: [
68		− { label: "direction from set A", color: Charts.S1, values: m.per_layer.setA },
69		− { label: "direction from set B", color: Charts.S2, values: m.per_layer.setB },
70		− { label: "bootstrap mean (×3)", color: Charts.REF, dash: true, ref: true,
71		− values: m.per_layer.boot_mean },
	69	+ { label: "mean of 6 fresh sources", color: Charts.S1, values: m.per_layer.mean },
	70	+ { label: "worst source (min)", color: Charts.S2, values: m.per_layer.min },
	71	+ { label: "random-direction damage", color: Charts.REF, dash: true, ref: true,
	72	+ values: m.per_layer.random_direction_damage },
72	73	],
73		− caption: "Erasing the diff-of-means agreement direction at any single early-mid layer " +
74		− "destroys most of the grammatical-agreement margin (baseline " +
75		− (m.baseline_margin ? m.baseline_margin.toFixed(2) : "—") + "); random-direction damage is " +
76		− "netted out. Level 2 — causally verified, replicated across direction estimates. " +
77		− "This profile ANTI-correlates with the probes/v2 decodability ranking (survival ledger 0/2).",
	74	+ caption: `The published claim is the BAND (layers ${band.early_layers ? band.early_layers.join("–") : "2–15"}): ` +
	75	+ "erasing the diff-of-means agreement direction at any single early-band layer destroys most of " +
	76	+ "the grammatical margin (baseline " + (m.baseline_margin ? m.baseline_margin.toFixed(2) : "—") +
	77	+ "), replicated across six fresh direction estimates on a fresh behavioral bank (Level 2). " +
	78	+ "The late band is displayed but carries no claim — run #3's per-layer version was refused by " +
	79	+ "the publication gate as estimator-unstable. Anti-correlates with the probes/v2 decodability " +
	80	+ "ranking (survival ledger 0/2): decodability peaks ≠ causal joints.",
78	81	});
79	82	}
80	83
		`@@ -212,15 +215,16 @@ app.get("/", (req, res) => {`
212	215	</section>
213	216
214	217	${(() => {
	218	+ const causal = interventionsMapCharts();
215	219	const probe2 = probeMapCharts("agreement", MAP_V2);
216	220	const probe1 = probeMapCharts("lang_id");
217	221	const exph = expHCharts();
218		− return probe1 \|\| probe2 \|\| exph ? `<section class="card">
	222	+ return causal \|\| probe1 \|\| probe2 \|\| exph ? `<section class="card">
219	223	<h2>First measured maps</h2>
220	224	<p class="lede-small">Every figure below is regenerated from committed code + versioned results —
221	225	raw JSON with hardware manifests under <a href="/results">Results</a>. Full per-property maps live
222	226	in the <a href="/atlas">Atlas</a>.</p>
223		− ${probe2}${probe1}${exph}
	227	+ ${causal}${probe2}${probe1}${exph}
224	228	</section>` : "";
225	229	})()}
226	230
227	231