RESULTS.md
SRIC: endogenous semantic witness
RESULTS.md
1 # SRIC v2.4 Block 23 Results
2
3 ## Canonical disposition
4
5 **PASS within the released finite fixture.**
6
7 The visible compiler never receives the four protected intermediate semantic coordinates or the hidden regime label. It selects a quotient from observable future, intervention, answer, and semantic-twin evidence. A separate final witness then opens the protected ontology and determines the strongest justified certificate.
8
9 ## Audit chronology
10
11 - 24 pre-audit episodes compiled before the final witness was opened
12 - thresholds and candidate grammar frozen in the pre-audit seal
13 - 36 final episodes generated under the committed seed rule
14 - two independently gauged systems in every episode
15 - 520 construction and 320 calibration samples per system
16 - 420 audit-only samples per system in every final episode
17 - four anonymous interventions and five recurrent latent steps
18 - 4 protected semantic coordinates, 12 nuisance coordinates per step
19
20 ## Certificate matrix
21
22 | Hidden audit regime | Required certificate | Episodes | Selected method | Rank |
23 |---|---|---:|---|---:|
24 | Identified | `ONTOLOGY_IDENTIFIED` | 6 | `predictive` | 4 |
25 | Answer shortcut | `ONTOLOGY_IDENTIFIED` | 6 | `predictive` | 4 |
26 | Gauge equivalent | `ONTOLOGY_IDENTIFIED_UP_TO_GAUGE` | 6 | `intervention` | 4 |
27 | Semantic ambiguity | `ONTOLOGY_UNDERIDENTIFIED` | 6 | `intervention` | 4 |
28 | Partial ontology | `PARTIAL_ONTOLOGY` | 6 | `predictive` | 3 |
29 | Destructive loss | `SEMANTIC_INFORMATION_LOSS` | 6 | `None` | None |
30
31 All 36 final certificates match the precommitted audit truth.
32
33 ## Minimum-quotient behavior
34
35 - identified and answer-shortcut regimes select a rank-4 predictive quotient
36 - gauge-equivalent and semantically ambiguous regimes select a rank-4 intervention quotient
37 - partial-ontology regimes select a rank-3 quotient and report 75% protected-coordinate coverage
38 - destructive-loss regimes select no quotient and issue an information-loss certificate
39
40 Selected methods across the final suite:
41
42 ```text
43 predictive: 18
44 intervention: 12
45 no-go: 6
46 ```
47
48 ## Hidden witness results
49
50 | Quantity | Frozen result |
51 |---|---:|
52 | Minimum full identified semantic R2 | 1.000000 |
53 | Minimum gauge-family semantic R2 | 0.998872 |
54 | Minimum semantic-subspace overlap | 0.991563 |
55 | Minimum gauge-invariant R2 | 0.995522 |
56 | Minimum ambiguity disagreement | 0.781682 |
57 | Partial-ontology coverage | 0.750000 |
58 | Minimum destructive collision count | 64 |
59 | Minimum destructive protected-consequence gap | 1.736704 |
60
61 ## Answer-only shortcut control
62
63 The answer-only quotient predicts the binary answer with at least **0.881250** accuracy, but reaches at most:
64
65 - future R2: **0.313061**
66 - intervention-effect R2: **0.357401**
67
68 No answer-only candidate passes the semantic-construction gate. This separates label prediction from an adequate causal-predictive ontology.
69
70 ## Counterfactual quotient patching
71
72 On the identified and answer-shortcut regimes, the audit-only runtime replaces only the selected latent quotient with a donor quotient and independently recomputes future and intervention consequences.
73
74 - minimum selected-packet future R2: **0.983162**
75 - maximum answer-only control future R2: **-0.161100**
76 - maximum PCA control future R2: **-0.706230**
77 - maximum random control future R2: **-0.703696**
78
79 The selected quotient supports the intended counterfactual consequences; equal-rank correlational and random controls do not.
80
81 ## Gauge versus underidentification
82
83 The same visible repeated intervention signature appears in two regimes. The compiler therefore emits the same pre-certificate, `CANDIDATE_GAUGE_AMBIGUOUS`, in both.
84
85 - In the declared gauge regime, the protected invariant is recoverable with R2 at least **0.997034**, so the final certificate is `ONTOLOGY_IDENTIFIED_UP_TO_GAUGE`.
86 - In the semantic-ambiguity regime, two observationally equivalent orientations disagree on a protected named coordinate by at least **0.781682**, so the final certificate is `ONTOLOGY_UNDERIDENTIFIED`.
87
88 This prevents coordinate nonuniqueness from being confused with semantic nonidentifiability.
89
90 ## Information-loss no-go
91
92 Every destructive episode exposes 64 exact observation collisions whose protected future consequences disagree. The compiler receives those protected consequences but never the hidden intermediate coordinates. Because one observed state would require two different protected outcomes, deterministic target-only semantic recovery is impossible within the declared interface.
93
94 ## Receipts
95
96 ```text
97 pre-audit seal: 04785a6f21a8d9f64100e4609329333498bceaa6b5afae90472313f6ebbd4ef9
98 final result: 213bcf9fbf62b797f9903d9e05e84e5a4036bb02c927723a4570fa44e2c674ce
99 fixture smoke: e447e082f3befe5b7b6bb0254464c4cb1402684a0c93e3022579439be6f62664
100 ```