VRAVEN Explanation Report
Trace c7780becb8dd422fa69262a2f10d5421
| Trace Id | c7780becb8dd422fa69262a2f10d5421 |
|---|---|
| Mode | causal |
| Layers Captured | 96 |
| Target | 208 |
| Framework | pytorch |
| Device | cpu |
| Model Class | MobileNetV2 |
| Parameter Count | 3504872 |
| Predicted Classes | [208] |
| Confidence | [0.526278555393219] |
| Predictive Entropy | [0.4889216125011444] |
| Causal Interventions | 3 |
Figures are standalone scientific views explicitly composed into this local report. Signal statistics and first-order gradients must not be interpreted as causal unless an intervention view is present. Use the Dark / Light control to switch page chrome and figure themes.
Inference before illustration
Scientific Evidence Audit
This section separates prediction, observational evidence and intervention evidence. It reports complementary quantities directly and deliberately does not manufacture a single explanation-confidence or trust score.
Evidence tier
Targeted Interventional
Highest evidence type present; not a quality grade
Effective evidence layers
45.37 / 96
exp(Shannon entropy) of absolute layer mass
50% evidence crossing
features.2.conv.1.0 (12.5% depth)
Earliest depth at which cumulative mass reaches 0.5
Zero-ablation coverage
3.1%
Uniquely matched gradient-bearing layers
Gradient ↔ ablation ρ
-1.000
Spearman rank concordance on matched layers
Permutation p
0.3333
exact (6 permutations)
Critical findings
- Absolute first-order evidence occupies 45.37 effective layers across 96 measured layers; 80% of its mass is reached at features.6.conv.0.1.
- 3 layer effects were uniquely paired with zero ablation: rank concordance ρ=-1.000, sign agreement 66.7%, and scale-normalised L1 discrepancy 0.450.
- Zero ablation produced 3 unique module tests; 3 of 96 gradient-bearing layers were uniquely matched (3.1%). All remaining layer-level causal roles are unresolved.
- Median normalised predictive entropy is 0.489 and median class margin is 0.456; these describe the output distribution, not explanation validity.
Hypothesis and evidence ledger
| Question | Estimate | Evidence tier | Interpretation boundary |
|---|---|---|---|
| Output separation | median class margin = 0.456 | Predictive | Measures output separation only; it does not validate the explanation. |
| Depth concentration | H/Hmax = 0.836; normalised Gini = 0.603 | Observational | Reports dispersion of absolute first-order evidence across captured depth. |
| Evidence cancellation | cancellation ratio = 0.927 | Observational | Quantifies support/opposition cancellation; high values are not uncertainty. |
| Local causal corroboration | ρ = -1.000; sign agreement = 66.7%; L1 discrepancy = 0.450 | Targeted interventional | Compares scale-normalised gradients with actual zero-output ablations. |
Layer evidence ledger
| Depth | Layer | Signed first-order | Absolute mass | Cumulative mass | Zero-ablation Δ | Cross-check |
|---|---|---|---|---|---|---|
| 1 | features.0.0 | -0.60129 | 2.57% | 2.57% | — | Not causally tested |
| 2 | features.0.1 | 3.3233 | 12.34% | 14.90% | 7.9521 | Direction agrees |
| 3 | features.0.2 | 3.3233 | 12.34% | 27.24% | 7.9521 | Direction agrees |
| 4 | features.1.conv.0.0 | 3.3233 | 1.55% | 28.79% | — | Not causally tested |
| 5 | features.1.conv.0.1 | 4.3992 | 3.50% | 32.29% | — | Not causally tested |
| 6 | features.1.conv.0.2 | 4.3992 | 3.50% | 35.79% | — | Not causally tested |
| 7 | features.1.conv.1 | 4.3992 | 2.99% | 38.78% | — | Not causally tested |
| 8 | features.1.conv.2 | -0.76382 | 0.89% | 39.67% | — | Not causally tested |
| 9 | features.2.conv.0.0 | -0.76382 | 1.22% | 40.89% | — | Not causally tested |
| 10 | features.2.conv.0.1 | -5.4387 | 4.44% | 45.34% | 8.0675 | Direction contradicts |
| 11 | features.2.conv.0.2 | -5.4387 | 4.44% | 49.78% | — | Not causally tested |
| 12 | features.2.conv.1.0 | -5.4387 | 2.18% | 51.96% | — | Not causally tested |
| 13 | features.2.conv.1.1 | -1.0923 | 2.44% | 54.40% | — | Not causally tested |
| 14 | features.2.conv.1.2 | -1.0923 | 2.44% | 56.85% | — | Not causally tested |
| 15 | features.2.conv.2 | -1.0923 | 1.93% | 58.77% | — | Not causally tested |
| 16 | features.2.conv.3 | 2.0106 | 0.76% | 59.54% | — | Not causally tested |
| 17 | features.3.conv.0.0 | 0.84731 | 0.90% | 60.44% | — | Not causally tested |
| 18 | features.3.conv.0.1 | 4.9838 | 1.78% | 62.21% | — | Not causally tested |
| 19 | features.3.conv.0.2 | 4.9838 | 1.78% | 63.99% | — | Not causally tested |
| 20 | features.3.conv.1.0 | 4.9838 | 0.71% | 64.70% | — | Not causally tested |
| 21 | features.3.conv.1.1 | -0.72435 | 1.44% | 66.14% | — | Not causally tested |
| 22 | features.3.conv.1.2 | -0.72435 | 1.44% | 67.57% | — | Not causally tested |
| 23 | features.3.conv.2 | -0.72435 | 0.80% | 68.37% | — | Not causally tested |
| 24 | features.3.conv.3 | 1.5137 | 0.34% | 68.71% | — | Not causally tested |
| 25 | features.4.conv.0.0 | 2.677 | 0.72% | 69.44% | — | Not causally tested |
| 26 | features.4.conv.0.1 | -2.8071 | 0.89% | 70.32% | — | Not causally tested |
| 27 | features.4.conv.0.2 | -2.8071 | 0.89% | 71.21% | — | Not causally tested |
| 28 | features.4.conv.1.0 | -2.8071 | 0.92% | 72.13% | — | Not causally tested |
| 29 | features.4.conv.1.1 | -2.4264 | 1.05% | 73.18% | — | Not causally tested |
| 30 | features.4.conv.1.2 | -2.4264 | 1.05% | 74.23% | — | Not causally tested |
| 31 | features.4.conv.2 | -2.4264 | 0.69% | 74.92% | — | Not causally tested |
| 32 | features.4.conv.3 | -0.60797 | 0.43% | 75.35% | — | Not causally tested |
| 33 | features.5.conv.0.0 | -0.049841 | 0.38% | 75.73% | — | Not causally tested |
| 34 | features.5.conv.0.1 | 0.91238 | 0.83% | 76.56% | — | Not causally tested |
| 35 | features.5.conv.0.2 | 0.91238 | 0.83% | 77.39% | — | Not causally tested |
| 36 | features.5.conv.1.0 | 0.91238 | 0.27% | 77.66% | — | Not causally tested |
| 37 | features.5.conv.1.1 | 1.0774 | 0.61% | 78.27% | — | Not causally tested |
| 38 | features.5.conv.1.2 | 1.0774 | 0.61% | 78.88% | — | Not causally tested |
| 39 | features.5.conv.2 | 1.0774 | 0.37% | 79.24% | — | Not causally tested |
| 40 | features.5.conv.3 | 0.71787 | 0.14% | 79.38% | — | Not causally tested |
| 41 | features.6.conv.0.0 | 0.27789 | 0.27% | 79.65% | — | Not causally tested |
| 42 | features.6.conv.0.1 | 0.24231 | 0.46% | 80.11% | — | Not causally tested |
| 43 | features.6.conv.0.2 | 0.24231 | 0.46% | 80.57% | — | Not causally tested |
| 44 | features.6.conv.1.0 | 0.24231 | 0.24% | 80.82% | — | Not causally tested |
| 45 | features.6.conv.1.1 | -0.59345 | 0.38% | 81.19% | — | Not causally tested |
| 46 | features.6.conv.1.2 | -0.59345 | 0.38% | 81.57% | — | Not causally tested |
| 47 | features.6.conv.2 | -0.59345 | 0.28% | 81.85% | — | Not causally tested |
| 48 | features.6.conv.3 | -0.14928 | 0.11% | 81.96% | — | Not causally tested |
| 49 | features.7.conv.0.0 | -0.26743 | 0.50% | 82.46% | — | Not causally tested |
| 50 | features.7.conv.0.1 | -1.5832 | 0.47% | 82.92% | — | Not causally tested |
| 51 | features.7.conv.0.2 | -1.5832 | 0.47% | 83.39% | — | Not causally tested |
| 52 | features.7.conv.1.0 | -1.5832 | 0.50% | 83.89% | — | Not causally tested |
| 53 | features.7.conv.1.1 | 0.65156 | 1.09% | 84.98% | — | Not causally tested |
| 54 | features.7.conv.1.2 | 0.65156 | 1.09% | 86.07% | — | Not causally tested |
| 55 | features.7.conv.2 | 0.65156 | 0.80% | 86.87% | — | Not causally tested |
| 56 | features.7.conv.3 | -0.50109 | 0.33% | 87.19% | — | Not causally tested |
| 57 | features.8.conv.0.0 | 0.90485 | 0.23% | 87.43% | — | Not causally tested |
| 58 | features.8.conv.0.1 | 0.3268 | 0.47% | 87.89% | — | Not causally tested |
| 59 | features.8.conv.0.2 | 0.3268 | 0.47% | 88.36% | — | Not causally tested |
| 60 | features.8.conv.1.0 | 0.3268 | 0.18% | 88.54% | — | Not causally tested |
| 61 | features.8.conv.1.1 | -0.93413 | 0.24% | 88.79% | — | Not causally tested |
| 62 | features.8.conv.1.2 | -0.93413 | 0.24% | 89.03% | — | Not causally tested |
| 63 | features.8.conv.2 | -0.93413 | 0.18% | 89.21% | — | Not causally tested |
| 64 | features.8.conv.3 | 1.4325 | 0.11% | 89.32% | — | Not causally tested |
| 65 | features.9.conv.0.0 | -0.55577 | 0.22% | 89.54% | — | Not causally tested |
| 66 | features.9.conv.0.1 | 0.014847 | 0.33% | 89.87% | — | Not causally tested |
| 67 | features.9.conv.0.2 | 0.014847 | 0.33% | 90.20% | — | Not causally tested |
| 68 | features.9.conv.1.0 | 0.014847 | 0.18% | 90.38% | — | Not causally tested |
| 69 | features.9.conv.1.1 | -1.0029 | 0.29% | 90.67% | — | Not causally tested |
| 70 | features.9.conv.1.2 | -1.0029 | 0.29% | 90.95% | — | Not causally tested |
| 71 | features.9.conv.2 | -1.0029 | 0.20% | 91.15% | — | Not causally tested |
| 72 | features.9.conv.3 | -0.31567 | 0.10% | 91.25% | — | Not causally tested |
| 73 | features.10.conv.0.0 | 0.029169 | 0.20% | 91.45% | — | Not causally tested |
| 74 | features.10.conv.0.1 | 0.46967 | 0.23% | 91.68% | — | Not causally tested |
| 75 | features.10.conv.0.2 | 0.46967 | 0.23% | 91.92% | — | Not causally tested |
| 76 | features.10.conv.1.0 | 0.46967 | 0.19% | 92.11% | — | Not causally tested |
| 77 | features.10.conv.1.1 | 0.73238 | 0.29% | 92.40% | — | Not causally tested |
| 78 | features.10.conv.1.2 | 0.73238 | 0.29% | 92.68% | — | Not causally tested |
| 79 | features.10.conv.2 | 0.73239 | 0.19% | 92.87% | — | Not causally tested |
| 80 | features.10.conv.3 | 0.050858 | 0.09% | 92.96% | — | Not causally tested |
| 81 | features.11.conv.0.0 | 0.28832 | 0.38% | 93.35% | — | Not causally tested |
| 82 | features.11.conv.0.1 | -1.4106 | 0.48% | 93.83% | — | Not causally tested |
| 83 | features.11.conv.0.2 | -1.4106 | 0.48% | 94.32% | — | Not causally tested |
| 84 | features.11.conv.1.0 | -1.4106 | 0.49% | 94.81% | — | Not causally tested |
| 85 | features.11.conv.1.1 | 2.363 | 0.84% | 95.65% | — | Not causally tested |
| 86 | features.11.conv.1.2 | 2.363 | 0.84% | 96.49% | — | Not causally tested |
| 87 | features.11.conv.2 | 2.363 | 0.65% | 97.14% | — | Not causally tested |
| 88 | features.11.conv.3 | -2.3242 | 0.32% | 97.46% | — | Not causally tested |
| 89 | features.12.conv.0.0 | -2.6649 | 0.33% | 97.79% | — | Not causally tested |
| 90 | features.12.conv.0.1 | -1.866 | 0.35% | 98.14% | — | Not causally tested |
| 91 | features.12.conv.0.2 | -1.866 | 0.35% | 98.48% | — | Not causally tested |
| 92 | features.12.conv.1.0 | -1.866 | 0.26% | 98.74% | — | Not causally tested |
| 93 | features.12.conv.1.1 | 3.6736 | 0.40% | 99.15% | — | Not causally tested |
| 94 | features.12.conv.1.2 | 3.6736 | 0.40% | 99.55% | — | Not causally tested |
| 95 | features.12.conv.2 | 3.6736 | 0.30% | 99.85% | — | Not causally tested |
| 96 | features.12.conv.3 | -0.18099 | 0.15% | 100.00% | — | Not causally tested |
Mathematical definitions, assumptions and machine-readable audit
p_l = ||a_l ⊙ ∂s_t/∂a_l||₁ / Σ_j ||a_j ⊙ ∂s_t/∂a_j||₁
N_eff = exp(−Σ_l p_l log p_l); participation ratio = 1/Σ_l p_l².
Concordance compares Σ(a_l ⊙ ∂s_t/∂a_l) with measured
s_t(x) − s_t(do(a_l=0)) after separate signed L1 normalisation.
Known limitations
- Activation × gradient is a local first-order approximation and is not causal evidence.
- Layer summaries are batch-reduced; they do not provide per-sample confidence intervals.
- Zero-output ablation may produce an internal state outside the training distribution.
- Targeted top-k interventions leave untested layers and multi-layer interactions unresolved.
- The permutation p-value tests exchangeability under the matched-layer null; it is not the probability that the mechanism claim is true.
Machine-readable values
{
"evidence_tier": "targeted_interventional",
"concentration": {
"layer_names": [
"features.0.0",
"features.0.1",
"features.0.2",
"features.1.conv.0.0",
"features.1.conv.0.1",
"features.1.conv.0.2",
"features.1.conv.1",
"features.1.conv.2",
"features.2.conv.0.0",
"features.2.conv.0.1",
"features.2.conv.0.2",
"features.2.conv.1.0",
"features.2.conv.1.1",
"features.2.conv.1.2",
"features.2.conv.2",
"features.2.conv.3",
"features.3.conv.0.0",
"features.3.conv.0.1",
"features.3.conv.0.2",
"features.3.conv.1.0",
"features.3.conv.1.1",
"features.3.conv.1.2",
"features.3.conv.2",
"features.3.conv.3",
"features.4.conv.0.0",
"features.4.conv.0.1",
"features.4.conv.0.2",
"features.4.conv.1.0",
"features.4.conv.1.1",
"features.4.conv.1.2",
"features.4.conv.2",
"features.4.conv.3",
"features.5.conv.0.0",
"features.5.conv.0.1",
"features.5.conv.0.2",
"features.5.conv.1.0",
"features.5.conv.1.1",
"features.5.conv.1.2",
"features.5.conv.2",
"features.5.conv.3",
"features.6.conv.0.0",
"features.6.conv.0.1",
"features.6.conv.0.2",
"features.6.conv.1.0",
"features.6.conv.1.1",
"features.6.conv.1.2",
"features.6.conv.2",
"features.6.conv.3",
"features.7.conv.0.0",
"features.7.conv.0.1",
"features.7.conv.0.2",
"features.7.conv.1.0",
"features.7.conv.1.1",
"features.7.conv.1.2",
"features.7.conv.2",
"features.7.conv.3",
"features.8.conv.0.0",
"features.8.conv.0.1",
"features.8.conv.0.2",
"features.8.conv.1.0",
"features.8.conv.1.1",
"features.8.conv.1.2",
"features.8.conv.2",
"features.8.conv.3",
"features.9.conv.0.0",
"features.9.conv.0.1",
"features.9.conv.0.2",
"features.9.conv.1.0",
"features.9.conv.1.1",
"features.9.conv.1.2",
"features.9.conv.2",
"features.9.conv.3",
"features.10.conv.0.0",
"features.10.conv.0.1",
"features.10.conv.0.2",
"features.10.conv.1.0",
"features.10.conv.1.1",
"features.10.conv.1.2",
"features.10.conv.2",
"features.10.conv.3",
"features.11.conv.0.0",
"features.11.conv.0.1",
"features.11.conv.0.2",
"features.11.conv.1.0",
"features.11.conv.1.1",
"features.11.conv.1.2",
"features.11.conv.2",
"features.11.conv.3",
"features.12.conv.0.0",
"features.12.conv.0.1",
"features.12.conv.0.2",
"features.12.conv.1.0",
"features.12.conv.1.1",
"features.12.conv.1.2",
"features.12.conv.2",
"features.12.conv.3"
],
"signed_influence": [
-0.6012938022613525,
3.3232581615448,
3.3232581615448,
3.3232569694519043,
4.399157524108887,
4.399157524108887,
4.3991570472717285,
-0.7638232707977295,
-0.7638235092163086,
-5.438686370849609,
-5.438686370849609,
-5.438686370849609,
-1.092299222946167,
-1.092299222946167,
-1.0923004150390625,
2.0105504989624023,
0.8473107814788818,
4.983786106109619,
4.983786106109619,
4.9837870597839355,
-0.7243454456329346,
-0.7243454456329346,
-0.7243451476097107,
1.5137182474136353,
2.6769580841064453,
-2.8071417808532715,
-2.8071417808532715,
-2.8071415424346924,
-2.426426887512207,
-2.426426887512207,
-2.4264273643493652,
-0.6079689860343933,
-0.04984140396118164,
0.912384033203125,
0.912384033203125,
0.9123841524124146,
1.0773696899414062,
1.0773696899414062,
1.0773695707321167,
0.7178689241409302,
0.27789387106895447,
0.2423134595155716,
0.2423134595155716,
0.2423136681318283,
-0.593453586101532,
-0.593453586101532,
-0.5934536457061768,
-0.1492791771888733,
-0.2674315571784973,
-1.5832000970840454,
-1.5832000970840454,
-1.5831999778747559,
0.6515634059906006,
0.6515634059906006,
0.6515636444091797,
-0.5010876655578613,
0.9048486948013306,
0.3268030881881714,
0.3268030881881714,
0.32680314779281616,
-0.934125542640686,
-0.934125542640686,
-0.934125542640686,
1.4324668645858765,
-0.555772602558136,
0.014847159385681152,
0.014847159385681152,
0.014847129583358765,
-1.0029391050338745,
-1.0029391050338745,
-1.0029391050338745,
-0.3156704008579254,
0.029168516397476196,
0.4696734547615051,
0.4696734547615051,
0.46967318654060364,
0.7323847413063049,
0.7323847413063049,
0.7323850393295288,
0.05085812509059906,
0.2883225679397583,
-1.410628080368042,
-1.410628080368042,
-1.4106279611587524,
2.362985610961914,
2.362985610961914,
2.362985610961914,
-2.324174404144287,
-2.664947271347046,
-1.8660250902175903,
-1.8660250902175903,
-1.8660250902175903,
3.673603057861328,
3.673603057861328,
3.673603057861328,
-0.18099215626716614
],
"absolute_influence": [
391.2163391113281,
1881.2706298828125,
1881.2706298828125,
237.06202697753906,
533.5443115234375,
533.5443115234375,
455.27752685546875,
136.0453338623047,
186.71917724609375,
677.2216796875,
677.2216796875,
333.13470458984375,
372.371826171875,
372.371826171875,
294.0095520019531,
116.23435974121094,
137.21011352539062,
270.845703125,
270.845703125,
108.59205627441406,
218.91580200195312,
218.91580200195312,
121.76220703125,
52.02739715576172,
110.45281982421875,
135.3365020751953,
135.3365020751953,
140.17169189453125,
160.31040954589844,
160.31040954589844,
105.37196350097656,
64.93258666992188,
57.53619384765625,
126.71349334716797,
126.71349334716797,
41.112205505371094,
92.8696517944336,
92.8696517944336,
56.05876922607422,
20.89837646484375,
40.682125091552734,
70.66683959960938,
70.66683959960938,
36.709129333496094,
57.26878356933594,
57.26878356933594,
42.860313415527344,
17.020038604736328,
75.81719970703125,
71.20653533935547,
71.20653533935547,
75.5960693359375,
166.19989013671875,
166.19989013671875,
122.0156478881836,
50.17181396484375,
35.12919235229492,
71.2061538696289,
71.2061538696289,
27.752405166625977,
37.27519226074219,
37.27519226074219,
26.714977264404297,
17.40493392944336,
33.68124008178711,
50.23210144042969,
50.23210144042969,
27.718338012695312,
43.68560028076172,
43.68560028076172,
29.99532699584961,
15.624992370605469,
30.068723678588867,
35.678688049316406,
35.678688049316406,
29.278034210205078,
43.560638427734375,
43.560638427734375,
28.765626907348633,
14.27054214477539,
58.248619079589844,
73.89324951171875,
73.89324951171875,
74.737548828125,
128.5771942138672,
128.5771942138672,
98.7233657836914,
49.20166778564453,
49.9727783203125,
52.988990783691406,
52.988990783691406,
39.55485534667969,
61.730709075927734,
61.730709075927734,
45.365478515625,
22.743722915649414
],
"mass_share": [
0.025655563935979876,
0.12337178716890165,
0.12337178716890165,
0.015546283173475792,
0.03498928553971419,
0.03498928553971419,
0.029856630542037233,
0.008921712648874945,
0.012244851022307433,
0.044411499124810565,
0.044411499124810565,
0.021846630261693544,
0.02441976021761074,
0.02441976021761074,
0.019280843116903684,
0.007622529403225464,
0.008998097697581019,
0.017761781803626246,
0.017761781803626246,
0.007121355025754555,
0.014356272459415769,
0.014356272459415769,
0.007985039925920102,
0.0034119030334579703,
0.007243381979767255,
0.008875228191514732,
0.008875228191514732,
0.00919231495183348,
0.010512991280094579,
0.010512991280094579,
0.006910184663554516,
0.004258212048665678,
0.0037731642375790895,
0.008309740174718317,
0.008309740174718317,
0.0026960960252532824,
0.006090295959357306,
0.006090295959357306,
0.003676276255022711,
0.0013704939695767702,
0.0026678917953906088,
0.004634258440278205,
0.004634258440278205,
0.002407346832161982,
0.003755627747390134,
0.003755627747390134,
0.0028107351386345955,
0.0011161565736455578,
0.004972013742956605,
0.004669651130134882,
0.004669651130134882,
0.0049575122413407224,
0.010899217341589686,
0.010899217341589686,
0.00800166031586552,
0.0032902158020369473,
0.0023037361948146217,
0.004669626113743742,
0.004669626113743742,
0.0018199741011506595,
0.0024444686549742376,
0.0024444686549742376,
0.001751940649544632,
0.0011413976119777295,
0.002208780978637699,
0.0032941693924926716,
0.0032941693924926716,
0.0018177400123399253,
0.002864856599881948,
0.002864856599881948,
0.001967062600889104,
0.0010246708807558052,
0.0019718758796263454,
0.0023397715557596788,
0.0023397715557596788,
0.001920023279973411,
0.002856661730472393,
0.002856661730472393,
0.001886420137656915,
0.0009358474322111437,
0.00381988434934384,
0.0048458430739155185,
0.0048458430739155185,
0.004901211352097276,
0.008431959754942795,
0.008431959754942795,
0.006474176484019153,
0.003226594616418191,
0.0032771632497958785,
0.0034749633515873735,
0.0034749633515873735,
0.002593966608425306,
0.004048236219142562,
0.004048236219142562,
0.002975021281544008,
0.0014915098861415582
],
"cumulative_mass": [
0.025655563935979876,
0.14902735110488152,
0.27239913827378315,
0.28794542144725893,
0.32293470698697313,
0.35792399252668733,
0.3877806230687246,
0.3967023357175995,
0.40894718673990693,
0.4533586858647175,
0.4977701849895281,
0.5196168152512216,
0.5440365754688323,
0.568456335686443,
0.5877371788033467,
0.5953597082065722,
0.6043578059041532,
0.6221195877077794,
0.6398813695114056,
0.6470027245371601,
0.6613589969965759,
0.6757152694559916,
0.6837003093819117,
0.6871122124153697,
0.6943555943951369,
0.7032308225866517,
0.7121060507781665,
0.72129836573,
0.7318113570100945,
0.742324348290189,
0.7492345329537435,
0.7534927450024093,
0.7572659092399884,
0.7655756494147067,
0.773885389589425,
0.7765814856146783,
0.7826717815740356,
0.7887620775333929,
0.7924383537884157,
0.7938088477579924,
0.796476739553383,
0.8011109979936613,
0.8057452564339395,
0.8081526032661014,
0.8119082310134915,
0.8156638587608817,
0.8184745938995163,
0.8195907504731619,
0.8245627642161185,
0.8292324153462534,
0.8339020664763882,
0.8388595787177289,
0.8497587960593186,
0.8606580134009083,
0.8686596737167738,
0.8719498895188107,
0.8742536257136253,
0.8789232518273691,
0.8835928779411129,
0.8854128520422635,
0.8878573206972378,
0.890301789352212,
0.8920537300017567,
0.8931951276137344,
0.8954039085923722,
0.8986980779848649,
0.9019922473773576,
0.9038099873896975,
0.9066748439895794,
0.9095397005894613,
0.9115067631903504,
0.9125314340711062,
0.9145033099507325,
0.9168430815064922,
0.9191828530622519,
0.9211028763422253,
0.9239595380726977,
0.9268161998031701,
0.928702619940827,
0.9296384673730381,
0.933458351722382,
0.9383041947962976,
0.9431500378702131,
0.9480512492223104,
0.9564832089772533,
0.9649151687321961,
0.9713893452162152,
0.9746159398326334,
0.9778931030824293,
0.9813680664340166,
0.984843029785604,
0.9874369963940293,
0.9914852326131718,
0.9955334688323143,
0.9985084901138583,
0.9999999999999999
],
"shannon_entropy": 3.8148787854882094,
"normalised_entropy": 0.8357992478793326,
"effective_layers": 45.371256394921566,
"participation_ratio": 22.696764762189606,
"herfindahl_concentration": 0.0440591428107804,
"normalised_gini": 0.6033434313644618,
"signed_cancellation": 0.9273180834157533,
"half_mass_layer": "features.2.conv.1.0",
"half_mass_depth": 0.125,
"eighty_mass_layer": "features.6.conv.0.1",
"eighty_mass_depth": 0.4375
},
"concordance": {
"layer_names": [
"features.0.1",
"features.0.2",
"features.2.conv.0.1"
],
"observational_raw": [
3.3232581615448,
3.3232581615448,
-5.438686370849609
],
"causal_raw": [
7.9521164894104,
7.9521164894104,
8.067523956298828
],
"observational_l1": [
0.27498571978535624,
0.27498571978535624,
-0.4500285604292875
],
"causal_l1": [
0.33172856336450735,
0.33172856336450735,
0.3365428732709853
],
"spearman_rho": -1.0,
"sign_agreement": 0.6666666666666666,
"cosine_similarity": 0.09023611073626861,
"normalised_l1_discrepancy": 0.4500285604292875,
"permutation_p_value": 0.3333333333333333,
"permutation_method": "exact (6 permutations)",
"excluded_nonzero_interventions": [],
"excluded_ambiguous_layers": []
},
"gradient_layer_count": 96,
"intervened_layer_count": 3,
"zero_ablation_layer_count": 3,
"causal_coverage": 0.03125,
"median_predictive_entropy": 0.4889216125011444,
"median_prediction_margin": 0.4563339054584503
}Local Input Attribution
Observational
Scientific question
Where is signed input-times-gradient influence concentrated in the input space?
Interpretation
Sign identifies the local direction of the input-times-target-gradient term and absolute magnitude ranks concentration at this input.
Strongest displayed position
row 32, column 15
Signed local attribution -0.24508.
Positive attribution mass
15.651
Sum of positive stored input-times-gradient terms.
Opposing attribution mass
16.18
Absolute sum of negative stored input-times-gradient terms.
Methods, mappings and limitations
- Measured quantities
- input_times_target_gradient
- Definitions
- attribution_i = x_i times partial target_score / partial x_i; displayed mass share = sum displayed |attribution| / sum stored |attribution|
- Visual mapping
- tabular: bar length=signed or absolute attribution; label=feature name; dense: cell position=input position; colour=signed input-times-gradient value
- Normalisation
- colour clipped symmetrically at the 99th absolute percentile; data unchanged
- Claim boundary
- Input-times-gradient is local and may be sensitive to scaling and saturation.; Large inputs are deterministically sampled and displayed in sampled order.; Feature bars summarise the stored attribution vector and are not global SHAP values.
4. Decision Flow
Observational
Scientific question
How do supporting and opposing first-order signals vary through model depth?
Interpretation
Positive net values mark layers where measured local support exceeds opposition; ribbon area reports allocation of absolute first-order mass through depth.
Dominant net layer
features.12.conv.2
Net signed share 0.776%.
Total supporting mass
49.7%
Share of global absolute first-order mass with positive sign.
Total opposing mass
50.3%
Share of global absolute first-order mass with negative sign.
Methods, mappings and limitations
- Measured quantities
- unit_contributions; activation_gradient_signed
- Definitions
- support_l = sum of positive unit contributions at layer l / global absolute mass; opposition_l = sum of absolute negative unit contributions at layer l / global absolute mass; net_l = support_l - opposition_l
- Visual mapping
- x=layer execution order; upper ribbon=supporting influence; lower ribbon=opposing influence; line=net signed influence
- Normalisation
- global L1 normalisation across captured layers
- Claim boundary
- Activation-gradient interaction is a local first-order approximation, not causal evidence.; The ribbons aggregate units and do not claim neuron-to-neuron connectivity.
18. Activation Energy
Observational
Scientific question
Where is activation magnitude concentrated across captured layers?
Interpretation
Peak and concentration values describe where squared activation magnitude is stored; they do not establish target relevance or causal necessity.
Peak-energy layer
features.2.conv.0.0
Measured activation energy 3.6467e+07.
Peak share
19.9%
Peak layer energy divided by total captured activation energy.
Captured depth
96 layers
Layers with a measured activation-energy summary.
Methods, mappings and limitations
- Measured quantities
- activation_energy
- Definitions
- activation energy_l = sum of squared captured activations in layer l; peak share = maximum layer energy / total captured layer energy
- Visual mapping
- x=execution order; y=sum of squared activations
- Normalisation
- none
- Claim boundary
- Activation energy is observational magnitude, not causal importance.
15. Evidence Trajectory
Observational
Scientific question
Where does cumulative first-order support begin to dominate opposition?
Interpretation
The crossing and final net value locate where the measured first-order evidence becomes support-dominant; they do not measure hidden-layer confidence.
Stable support transition
features.0.1
First layer after which the registered net-support rule remains satisfied.
Final net evidence
0.071%
Final cumulative supporting mass minus opposing mass.
Measured depth
96 layers
Captured layers contributing a comparable signed first-order value.
Methods, mappings and limitations
- Measured quantities
- activation_gradient_signed
- Definitions
- support_l = cumulative positive activation-gradient mass through layer l; opposition_l = cumulative absolute negative activation-gradient mass through layer l; net_l = support_l - opposition_l after global L1 normalisation
- Visual mapping
- x=layer; support line=cumulative positive influence; opposition line=cumulative negative influence; net line=support minus opposition
- Normalisation
- global L1 normalisation before cumulative summation
- Claim boundary
- This is evidence trajectory, not layer confidence.; A valid probability trajectory requires trained intermediate probes and is not inferred here.
16. Single-layer Intervention Effects
Targeted interventional
Scientific question
Which temporary layer interventions most change the selected target score?
Interpretation
A positive score drop means the ablated layer supported the selected target under the declared intervention; a negative drop means the intervention increased it.
Largest target-score drop
8.0675 score units
Measured after intervening on features.2.conv.0.1.
Smallest target-score drop
7.9521 score units
Measured after intervening on features.0.1.
Interventions tested
3 layers
Temporary single-layer interventions included in this figure.
Methods, mappings and limitations
- Measured quantities
- intervention score_delta
- Definitions
- score_delta_l = baseline target score - target score after do(layer_l = baseline)
- Visual mapping
- y=intervened layer; x=target-score drop; colour=delta direction
- Normalisation
- none
- Claim boundary
- Single-layer ablation can create out-of-distribution internal states.; Intervention sensitivity supports a causal role but does not identify a complete mechanism.
17. Mechanism Concordance
Observational ↔ interventional
Scientific question
Does local activation-gradient evidence predict the direction and relative magnitude of measured zero-ablation effects across matched layers?
Interpretation
Agreement indicates that local first-order allocation tracks measured zero-ablation effects across matched layers; disagreement identifies where the observational proxy fails.
Rank concordance
-1.000 rho
Spearman correlation across uniquely matched layers.
Sign agreement
66.7%
Share of matched layers with the same observational and ablation direction.
Permutation evidence
p=0.3333
exact (6 permutations)
Methods, mappings and limitations
- Measured quantities
- activation_gradient_signed; zero-ablation score_delta
- Definitions
- rho = Spearman rank correlation of matched signed layer effects; sign agreement = matched layers with equal non-zero effect direction / matched layers; L1 discrepancy = sum of absolute differences between separately L1-normalised effects
- Visual mapping
- x=L1-normalised first-order activation-gradient interaction; y=L1-normalised zero-ablation target-score drop; point=one unambiguously matched layer; shape and colour=sign agreement; diagonal=perfect signed magnitude agreement
- Normalisation
- each signed vector divided by its own total absolute mass
- Claim boundary
- Zero ablation can create out-of-distribution internal states.; Agreement supports local concordance, not recovery of a unique causal mechanism.; Reused modules are excluded when one intervention maps to multiple executions.