File size: 13,319 Bytes
767a051 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 | """Deterministic observation-to-hypothesis pipeline for TOM Learner 0.1."""
from __future__ import annotations
from dataclasses import dataclass
import hashlib
from typing import Any, Iterable, Mapping
from tom_world03.canonical import attach_hash, canonical_bytes, require_hash
from tom_world03.rational import Q
from .affine import (
candidate_coefficients,
counterexamples_from_evidence,
derive_candidates,
evaluate_candidate,
find_contradictions,
select_from_training,
)
from .model import ObservationSet, PROFILE
from .split import deterministic_split
DECISION_SCHEMA = "TOM-LEARNER-ACCEPTANCE-DECISION-0.1"
CERTIFICATE_SCHEMA = "TOM-LEARNING-CERTIFICATE-0.1"
LEARNED_DEFINITION_SCHEMA = "TOM-LEARNED-AFFINE-DEFINITION-0.1"
REJECTION_LINEAGE_SCHEMA = "TOM-REJECTED-CANDIDATE-LINEAGE-0.1"
def _hash_values(value: Any) -> str:
return "sha256:" + hashlib.sha256(canonical_bytes(value)).hexdigest()
def _selected_candidate(
selection: Mapping[str, Any],
candidates: Iterable[Mapping[str, Any]],
) -> Mapping[str, Any] | None:
selected_hash = selection.get("selected_candidate_hash")
if selected_hash is None:
return None
for candidate in candidates:
if candidate.get("content_hash") == selected_hash:
return candidate
raise ValueError("selection references a candidate absent from enumeration")
def _evidence_for(
candidate: Mapping[str, Any] | None,
candidates: list[Mapping[str, Any]],
evidence: list[Mapping[str, Any]],
) -> Mapping[str, Any] | None:
if candidate is None:
return evidence[0] if evidence else None
wanted = candidate["content_hash"]
for item in evidence:
if item.get("candidate_hash") == wanted:
return item
raise ValueError("selected candidate has no residual evidence")
def _decision(
dataset: ObservationSet,
split: Mapping[str, Any],
selection: Mapping[str, Any],
candidate: Mapping[str, Any] | None,
evidence: Mapping[str, Any] | None,
contradictions: list[Mapping[str, Any]],
) -> dict[str, Any]:
reasons: list[str] = []
checks: list[dict[str, Any]] = []
def check(name: str, passed: bool, detail: str) -> None:
checks.append({"name": name, "passed": bool(passed), "detail": detail})
if not passed:
reasons.append(detail)
counts = split["counts"]
check("minimum training observations", int(counts["train"]) >= 2,
"training split has fewer than two observations")
check("minimum validation observations", int(counts["validation"]) >= 1,
"validation split is empty")
check("minimum holdout observations", int(counts["holdout"]) >= 1,
"holdout split is empty")
check("selected exact training candidate", candidate is not None,
str(selection.get("reason", "no exact training candidate")))
if candidate is not None:
check("model complexity",
int(candidate["complexity"]) <= dataset.acceptance_policy.max_model_complexity,
"selected model exceeds the declared complexity budget")
else:
check("model complexity", False, "no selected model exists")
if evidence is None:
check("exact training residuals", False, "no residual evidence exists")
check("exact validation residuals", False, "no validation residual evidence exists")
check("exact holdout residuals", False, "no holdout residual evidence exists")
else:
train_exact = bool(evidence["splits"]["train"]["metrics"]["exact"])
validation_exact = bool(evidence["splits"]["validation"]["metrics"]["exact"])
holdout_exact = bool(evidence["splits"]["holdout"]["metrics"]["exact"])
check("exact training residuals", train_exact,
"selected/best candidate has a nonzero training residual")
check("exact validation residuals",
validation_exact or not dataset.acceptance_policy.require_exact_validation,
"selected candidate has a nonzero validation residual")
check("exact holdout residuals",
holdout_exact or not dataset.acceptance_policy.require_exact_holdout,
"selected candidate has a nonzero holdout residual")
check("no exact-input contradictions",
not contradictions or not dataset.acceptance_policy.reject_contradictions,
"observation set contains contradictory exact outputs for one input")
accepted = all(item["passed"] for item in checks)
return attach_hash({
"schema": DECISION_SCHEMA,
"profile": PROFILE,
"observation_set_id": dataset.id,
"observation_set_hash": dataset.content_hash,
"split_certificate_hash": split["content_hash"],
"acceptance_policy_hash": dataset.acceptance_policy.content_hash,
"selection_certificate_hash": selection["content_hash"],
"selected_candidate_hash": None if candidate is None else candidate["content_hash"],
"evidence_hash": None if evidence is None else evidence["content_hash"],
"contradiction_hashes": [item["content_hash"] for item in contradictions],
"checks": checks,
"accepted": accepted,
"reasons": reasons,
"policy_boundary": {
"fit_and_selection_splits": ["train"],
"validation_role": "acceptance gate only",
"holdout_role": "final acceptance audit only",
},
})
def _learned_definition(
dataset: ObservationSet,
candidate: Mapping[str, Any],
certificate_hash: str,
) -> dict[str, Any]:
a, b = candidate_coefficients(candidate)
suffix = hashlib.sha256(canonical_bytes({
"dataset": dataset.id,
"candidate": candidate["content_hash"],
})).hexdigest()[:24]
return attach_hash({
"schema": LEARNED_DEFINITION_SCHEMA,
"profile": PROFILE,
"id": f"definition:learned-affine:{suffix}",
"kind": "learned-exact-affine-relation",
"domain": {
"input": dataset.input_name,
"output": dataset.output_name,
"numeric": "exact-rational",
},
"codomain": "exact-rational-residual",
"relation_interface": "SDF0@Def",
"model": "y=a*t+b",
"coefficients": {"a": a.to_record(), "b": b.to_record()},
"expression": {
"op": "sub",
"args": [
{"op": "field", "name": dataset.output_name},
{
"op": "add",
"args": [
{
"op": "mul",
"args": [
{"op": "const", "value": a.to_record()},
{"op": "field", "name": dataset.input_name},
],
},
{"op": "const", "value": b.to_record()},
],
},
],
},
"zero_locus": f"{dataset.output_name} = a*{dataset.input_name} + b",
"source_observation_set_hash": dataset.content_hash,
"source_candidate_hash": candidate["content_hash"],
"source_learning_certificate_hash": certificate_hash,
"base_world_hash": dataset.base_world_hash,
"base_handoff_hash": dataset.base_handoff_hash,
"provenance": {
"class": "deterministically-induced-and-held-out-verified",
"learner": PROFILE,
},
})
@dataclass(frozen=True, slots=True)
class LearningRun:
dataset: ObservationSet
split_certificate: Mapping[str, Any]
candidates: tuple[Mapping[str, Any], ...]
enumeration: Mapping[str, Any]
evidence: tuple[Mapping[str, Any], ...]
selection: Mapping[str, Any]
contradictions: tuple[Mapping[str, Any], ...]
counterexamples: tuple[Mapping[str, Any], ...]
decision: Mapping[str, Any]
certificate: Mapping[str, Any]
learned_definition: Mapping[str, Any] | None
rejection_lineage: Mapping[str, Any] | None
@property
def accepted(self) -> bool:
return bool(self.decision["accepted"])
def all_records(self) -> list[Mapping[str, Any]]:
values: list[Mapping[str, Any]] = [
*self.dataset.source_records(),
self.split_certificate,
*self.candidates,
self.enumeration,
*self.evidence,
self.selection,
*self.contradictions,
*self.counterexamples,
self.decision,
self.certificate,
]
if self.learned_definition is not None:
values.append(self.learned_definition)
if self.rejection_lineage is not None:
values.append(self.rejection_lineage)
return values
def summary(self) -> dict[str, Any]:
candidate = self.learned_definition
return {
"observation_set_id": self.dataset.id,
"observation_set_hash": self.dataset.content_hash,
"accepted": self.accepted,
"selected_candidate_hash": self.selection.get("selected_candidate_hash"),
"learned_definition_hash": None if candidate is None else candidate["content_hash"],
"candidate_count": len(self.candidates),
"counterexample_count": len(self.counterexamples),
"contradiction_count": len(self.contradictions),
"decision_hash": self.decision["content_hash"],
"certificate_hash": self.certificate["content_hash"],
}
def learn_observation_set(dataset: ObservationSet) -> LearningRun:
split = deterministic_split(dataset)
candidates, enumeration = derive_candidates(dataset, split)
evidence = [evaluate_candidate(dataset, split, candidate) for candidate in candidates]
selection = select_from_training(dataset, split, candidates, evidence)
selected = _selected_candidate(selection, candidates)
selected_evidence = _evidence_for(selected, candidates, evidence)
contradictions = find_contradictions(dataset)
counterexamples = [] if selected_evidence is None else counterexamples_from_evidence(dataset, selected_evidence)
decision = _decision(dataset, split, selection, selected, selected_evidence, contradictions)
certificate = attach_hash({
"schema": CERTIFICATE_SCHEMA,
"profile": PROFILE,
"observation_set_id": dataset.id,
"observation_set_hash": dataset.content_hash,
"base_world_hash": dataset.base_world_hash,
"base_handoff_hash": dataset.base_handoff_hash,
"split_certificate_hash": split["content_hash"],
"assignment_basis_hash": split["assignment_basis_hash"],
"candidate_enumeration_hash": enumeration["content_hash"],
"fit_input_hash": enumeration["fit_input_hash"],
"selection_certificate_hash": selection["content_hash"],
"selected_candidate_hash": selection.get("selected_candidate_hash"),
"selected_evidence_hash": None if selected_evidence is None else selected_evidence["content_hash"],
"contradiction_hashes": [item["content_hash"] for item in contradictions],
"counterexample_hashes": [item["content_hash"] for item in counterexamples],
"acceptance_decision_hash": decision["content_hash"],
"accepted": decision["accepted"],
"phase_trace": [
"parse-observations",
"deterministic-id-only-split",
"fit-train-only",
"select-train-only",
"validation-gate",
"holdout-audit",
"contradiction-gate",
"candidate-certificate",
],
"causal_boundary": "certificate is a proposal record; it is not authoritative world state",
})
learned_definition = None
rejection_lineage = None
if decision["accepted"]:
if selected is None:
raise AssertionError("accepted decision has no selected candidate")
learned_definition = _learned_definition(dataset, selected, certificate["content_hash"])
else:
rejection_lineage = attach_hash({
"schema": REJECTION_LINEAGE_SCHEMA,
"profile": PROFILE,
"observation_set_id": dataset.id,
"observation_set_hash": dataset.content_hash,
"learning_certificate_hash": certificate["content_hash"],
"acceptance_decision_hash": decision["content_hash"],
"candidate_hashes": [candidate["content_hash"] for candidate in candidates],
"counterexample_hashes": [item["content_hash"] for item in counterexamples],
"contradiction_hashes": [item["content_hash"] for item in contradictions],
"reasons": list(decision["reasons"]),
"retention": "rejected candidates and counterexamples remain in append-only lineage",
})
return LearningRun(
dataset=dataset,
split_certificate=split,
candidates=tuple(candidates),
enumeration=enumeration,
evidence=tuple(evidence),
selection=selection,
contradictions=tuple(contradictions),
counterexamples=tuple(counterexamples),
decision=decision,
certificate=certificate,
learned_definition=learned_definition,
rejection_lineage=rejection_lineage,
)
|