File size: 7,101 Bytes
feaa032
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
import torch
import torch.nn as nn
import torch.nn.functional as F

from config import DROPOUT


class HypothesisValidator(nn.Module):
    """
    Scores hypothesis-vs-option compatibility and emits a guidance
    potential for the reasoner.

    On the "Born rule" here
    -----------------------
    The distribution computed below weights the guidance potential. With
    real, non-negative amplitudes it is a temperature-scaled softmax and
    nothing more -- that is stated rather than dressed up. The place
    where amplitudes are genuinely complex, and where the Born rule
    therefore does real work, is models/interference.py; this module
    feeds it (through `validator_energy`) but does not itself perform a
    measurement.

    Fixed in the v44 audit, kept
    ----------------------------
    * Question conditioning. The question embedding was never computed
      anywhere in the pipeline, so the validator judged options using
      hypotheses alone -- a third of which were template fallbacks.
      Measured with a plain MLP on cached ARC embeddings: options only
      0.3585, +question 0.4849.
    * L2 (not L1) normalization of the amplitude. L1 is not the Born
      convention and gave this module half the collapse controller's
      effective temperature at the same nominal T.
    * Padded options are excluded from every mean over N.

    New in v45
    ----------
    * `align` / `polarity` features, so the validator can distinguish
      "the support for option n" from "the objection to option n"
      instead of averaging them together.
    """

    def __init__(self, dim, use_question=True):

        super().__init__()

        self.use_question = use_question

        # [H, O, H-O, H*O] plus, when the question is available,
        # [Q*O, Q-O]; then three scalars: align, polarity, align*polarity.
        n_vec = 6 if use_question else 4

        n_feat = dim * n_vec + 3

        self.encoder = nn.Sequential(
            nn.Linear(n_feat, dim * 2),
            nn.GELU(),
            nn.LayerNorm(dim * 2),
            nn.Dropout(DROPOUT),
            nn.Linear(dim * 2, dim),
            nn.GELU(),
        )

        self.causal = nn.Linear(dim, 1)
        self.diversity = nn.Linear(dim, 1)
        self.specificity = nn.Linear(dim, 1)
        self.relevance = nn.Linear(dim, 1)

        self.observable_gate = nn.Sequential(
            nn.Linear(dim, dim),
            nn.GELU(),
            nn.Linear(dim, 4),
        )

        self.reliability = nn.Sequential(
            nn.Linear(dim, dim),
            nn.GELU(),
            nn.Linear(dim, 1),
            nn.Sigmoid(),
        )

        self.potential = nn.Sequential(
            nn.Linear(dim + 1, dim),
            nn.GELU(),
            nn.Dropout(DROPOUT),
            nn.Linear(dim, dim),
        )

        self.temperature = nn.Parameter(torch.tensor(1.0))

    def forward(self, H, O, Q=None, y=None, H_mask=None, O_mask=None,
                align=None, polarity=None):

        B, K, D = H.shape
        _, N, _ = O.shape

        H = F.normalize(H, dim=-1)
        O = F.normalize(O, dim=-1)

        H_exp = H.unsqueeze(2).expand(B, K, N, D)
        O_exp = O.unsqueeze(1).expand(B, K, N, D)

        parts = [H_exp, O_exp, H_exp - O_exp, H_exp * O_exp]

        if self.use_question:

            if Q is None:
                raise ValueError(
                    "HypothesisValidator was built with use_question=True "
                    "but forward() got Q=None. Rebuild the cache so it "
                    "carries question embeddings, or construct the model "
                    "with use_question=False."
                )

            Q_exp = F.normalize(Q, dim=-1).view(B, 1, 1, D).expand(B, K, N, D)

            parts += [Q_exp * O_exp, Q_exp - O_exp]

        if align is None:
            align = torch.zeros(B, K, N, device=H.device, dtype=H.dtype)

        if polarity is None:
            polarity = torch.zeros(B, K, device=H.device, dtype=H.dtype)

        a = align.unsqueeze(-1)
        p = polarity.view(B, K, 1, 1).expand(B, K, N, 1)

        parts += [a, p, a * p]

        features = torch.cat(parts, dim=-1)

        z = self.encoder(features)

        causal = self.causal(z).squeeze(-1)
        diversity = self.diversity(z).squeeze(-1)
        specificity = self.specificity(z).squeeze(-1)
        relevance = self.relevance(z).squeeze(-1)

        target = None

        if y is not None:
            target = torch.zeros_like(relevance)
            target[torch.arange(B, device=H.device), :, y] = 1.0

        # ------------------------------------------------------------
        # Masked reductions over the option axis
        # ------------------------------------------------------------

        if O_mask is not None:
            om = O_mask.view(B, 1, N, 1).to(z.dtype)
            denom = om.sum(dim=2).clamp(min=1.0)
            z_mean = (z * om).sum(dim=2) / denom
            om2 = O_mask.view(B, 1, N).to(causal.dtype)
            d2 = om2.sum(dim=2).clamp(min=1.0)
            mean = lambda t: (t * om2).sum(dim=2) / d2
        else:
            z_mean = z.mean(dim=2)
            mean = lambda t: t.mean(dim=2)

        gate = self.observable_gate(z_mean)

        weights = F.softmax(gate, dim=-1)

        energy = (
            weights[..., 0] * mean(causal)
            + weights[..., 1] * mean(diversity)
            + weights[..., 2] * mean(specificity)
            + weights[..., 3] * mean(relevance)
        )

        energy = -energy

        # ------------------------------------------------------------
        # Weighting distribution over hypotheses (see class docstring:
        # this is a softmax, and is not claimed to be more)
        # ------------------------------------------------------------

        temperature = 0.5 + F.softplus(self.temperature)

        log_amp = -energy / (2.0 * temperature)

        if H_mask is not None:
            log_amp = log_amp.masked_fill(~H_mask, float("-inf"))

        log_amp = log_amp - log_amp.max(dim=1, keepdim=True).values

        amplitude = torch.exp(log_amp)

        amplitude = amplitude / torch.sqrt(
            (amplitude ** 2).sum(dim=1, keepdim=True) + 1e-8
        )

        probabilities = amplitude.pow(2)

        probabilities = probabilities / (probabilities.sum(dim=1, keepdim=True) + 1e-8)

        reliability = self.reliability(z_mean).squeeze(-1)

        potential_input = torch.cat([z_mean, reliability.unsqueeze(-1)], dim=-1)

        potential = self.potential(potential_input)

        potential = potential * probabilities.unsqueeze(-1)

        return {
            "potential": potential,
            "validator_energy": energy,
            "validator_probabilities": probabilities,
            "reliability": reliability,
            "observable_weights": weights,
            "causal": mean(causal),
            "diversity": mean(diversity),
            "specificity": mean(specificity),
            "relevance": mean(relevance),
            "relevance_logits": relevance,
            "relevance_target": target,
        }