Datasets:
File size: 12,798 Bytes
e1ced61 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 | """Core types for question-preserving visual interventions (``docs/02`` §10).
An *intervention operator* transforms a structured ``World`` into a new
``World`` (plus bookkeeping) **without touching the human-authored question or
choices** and **without assigning a gold state**. The operator's only job is to
produce a well-formed transformed world and record what it did; the resulting
state (``FULL`` / ``A_SAME`` / ``A_CHANGED`` / ``U_MISSING`` / ``U_INVALID``) is
decided later by the P7 state-proof builder, which executes the program before
and after (``docs/02`` §11.1). This separation is a hard P5 gate
("operator가 state를 직접 assign하지 않음").
Design rules every operator obeys (``docs/07`` §7 P5 completion gate):
* **Source world immutable** — :func:`apply` returns a brand-new world; the
passed dict is never mutated.
* **Deterministic** — the same ``(world, target_node, seed, operation)`` always
reproduces the same ``transformed_world_sha256``; a different seed or
operation changes it.
* **Reject invalid mutations** — an operator that would break world
well-formedness (a dangling geometry constraint, a label collision, an
out-of-range substitute, a multiple-choice answer leaving the unchanged
choice set) raises :class:`InterventionError` rather than emitting a broken
world. Never a silent accept.
* **No state assignment** — :class:`TransformedWorld` carries no ``state``
field by construction.
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any, Literal, Protocol
from ..dsl.ast import canonical_number_str, parse_number
from ..hashing import canonical_json_hash
from ..ingest.base import IngestError
from ..sources.base import World
from ..sources.world_ops import (
NODE_VISIBILITY_KEY,
clone_world,
find_point,
find_row,
find_series,
find_stating_constraint,
hidden_nodes,
looks_numeric,
node_kind,
world_sha256,
)
# The four deterministic intervention operators (``docs/02`` §10). REDACT_BLUR_V1
# is deliberately absent: it is a shortcut stress test only and never appears in
# a primary U certificate (§10.2).
OperatorName = Literal[
"REDACT_SOLID_V1",
"CLEAN_DELETE_V1",
"SUBSTITUTE_V1",
"CONTROL_MATCHED_V1",
]
class InterventionError(IngestError):
"""A requested intervention is infeasible / would break the world.
Raised for an invalid geometry or plot mutation, a label collision, an
out-of-range substitute, or a multiple-choice answer that leaves the
unchanged choice set (``docs/02`` §10.4). An infeasible intervention is
rejected — never silently coerced into a broken world or a faked state.
"""
@dataclass(frozen=True)
class TransformedWorld:
"""One operator's deterministic output over a source world.
``world`` is the transformed (or, for REDACT_SOLID_V1, visibility-marked)
world. ``hidden_nodes`` lists semantic node ids a renderer must patch
(REDACT_SOLID_V1 keeps the node in the world but hides its pixels, §10.1).
``substitute_value`` records the new value a SUBSTITUTE_V1 / completion
placed on the target (canonical string, or ``None`` for non-substitute
operators). ``transformed_world_sha256`` is the content-addressed identity of
the *transformation* — deterministic over the full record, so the same
``(seed, operation, target, source)`` reproduces it byte-for-byte.
There is deliberately **no** ``state`` field: the operator never assigns a
gold state (P5 gate). P7's state-proof builder derives state by execution.
"""
world: World
operator: OperatorName
target_node: str
seed: int
transformed_world_sha256: str
hidden_nodes: frozenset[str] = field(default_factory=frozenset)
substitute_value: str | None = None
description: str = ""
@property
def is_hidden(self) -> bool:
return bool(self.hidden_nodes)
class InterventionOperator(Protocol):
"""One deterministic intervention operator (``docs/02`` §10)."""
name: OperatorName
def apply(
self,
world: World,
target_node: str,
*,
seed: int,
**kwargs: Any,
) -> TransformedWorld:
"""Return a transformed world; never mutate ``world``, never assign state."""
...
# --- transformation hashing ------------------------------------------------
def transformed_hash(
*,
operator: OperatorName,
target_node: str,
seed: int,
source_world_sha256: str,
transformed_world_sha256: str,
hidden_nodes: frozenset[str],
substitute_value: str | None,
) -> str:
"""Content-addressed identity of one transformation (deterministic).
Hashed over the full operator record — not just the world — so two
operators that happen to leave the world identical (e.g. REDACT_SOLID over a
node vs. a no-op) still differ, and the same transformation reproduces the
same hash across runs (P5 gate: "seed/operation → same transformed world
hash").
"""
return canonical_json_hash(
{
"operator": operator,
"target_node": target_node,
"seed": seed,
"source_world_sha256": source_world_sha256,
"transformed_world_sha256": transformed_world_sha256,
"hidden_nodes": sorted(hidden_nodes),
"substitute_value": substitute_value,
}
)
# --- value setters (read-only in, new world out) ---------------------------
def with_value(world: World, node_id: str, value: Any) -> World:
"""Return a new world with ``node_id``'s value set to ``value``.
The value is written to the same field the executor reads (a plot point's
``y``, a row cell, the stating constraint's numeric arg), so a completion
built this way is executable. Raises :class:`InterventionError` when the
node has no mutable value site. The source world is unchanged.
"""
new = clone_world(world)
if not _set_value(new, node_id, value):
raise InterventionError(f"node {node_id!r} has no mutable value site")
return new
def _set_value(world: World, node_id: str, value: Any) -> bool:
"""Write ``value`` onto ``node_id`` in place; return whether a site was found.
The write site follows :func:`~explicit_learning.sources.world_ops.node_kind`:
a plot point's ``y``, a series' ``label``, a row's first cell, a geometry
entity's ``label``, or a stating constraint's numeric arg. Dispatching on the
*node* kind (not the world kind) keeps a category substitute (a series/entity
label) from being mis-written onto a numeric site.
"""
kind = node_kind(world, node_id)
if kind == "point":
point = find_point(world, node_id)
if point is None:
return False
sidx, pidx, p = point
p["y"] = _canonical(value)
world["series"][sidx]["points"][pidx] = p
return True
if kind == "series":
series = find_series(world, node_id)
if series is None:
return False
sidx, s = series
s["label"] = _canonical(value)
world["series"][sidx] = s
return True
if kind == "row":
row = find_row(world, node_id)
if row is None:
return False
ridx, r = row
cells = r.get("cells", {}) or {}
if not cells:
return False
key = next(iter(cells))
cells[key] = _canonical(value)
world["rows"][ridx]["cells"] = cells
return True
if kind == "entity":
for i, e in enumerate(world.get("entities", []) or []):
if e.get("id") == node_id:
e["label"] = _canonical(value)
world["entities"][i] = e
return True
return False
if kind == "constraint":
found = find_stating_constraint(world, node_id)
if found is None:
return False
cidx, c = found
args = list(c.get("args") or [])
pred = c.get("predicate")
if pred == "MeasureOf" and len(args) >= 2 and args[0] == node_id:
args[1] = _canonical(value)
elif pred == "Equals" and len(args) == 2:
if args[0] == node_id and looks_numeric(args[1]):
args[1] = _canonical(value)
elif args[1] == node_id and looks_numeric(args[0]):
args[0] = _canonical(value)
else:
return False
else:
return False
c["args"] = args
world["constraints"][cidx] = c
return True
return False
def _canonical(value: Any) -> str:
"""Canonical string form of a substitute value (numbers via §3.2/§3.3)."""
if isinstance(value, str):
return value
try:
return canonical_number_str(parse_number(value))
except (ValueError, ZeroDivisionError):
return str(value)
def set_hidden(world: World, node_ids: frozenset[str]) -> World:
"""Return a new world carrying a visibility marker hiding ``node_ids``.
The nodes stay in the world (REDACT_SOLID_V1, §10.1: "semantic node는 world에
남기고 node_visibility[node]=hidden"); only the reserved
:data:`~explicit_learning.sources.world_ops.NODE_VISIBILITY_KEY` map is set,
which the renderer (P6) reads and executors ignore.
"""
new = clone_world(world)
visibility = dict(new.get(NODE_VISIBILITY_KEY, {}))
for nid in node_ids:
visibility[nid] = "hidden"
new[NODE_VISIBILITY_KEY] = visibility
return new
def remove_observed_value(world: World, node_id: str) -> World:
"""Return a copy whose target node remains identifiable but has no value.
REDACT is an epistemic operation: the renderer hides the value and the
executor evaluating the *observed* transformed world must not retain a
privileged copy of that value. Identity/position fields remain so the
question referent and a content-independent redaction anchor still exist.
Completion witnesses can restore alternative values with :func:`with_value`
while retaining the same visibility marker.
"""
new = clone_world(world)
kind = node_kind(new, node_id)
if kind == "point":
found_point = find_point(new, node_id)
if found_point is None:
raise InterventionError(f"point {node_id!r} not found")
sidx, pidx, point = found_point
point["y"] = None
new["series"][sidx]["points"][pidx] = point
return new
if kind == "row":
found_row = find_row(new, node_id)
if found_row is None:
raise InterventionError(f"row {node_id!r} not found")
ridx, row = found_row
cells = dict(row.get("cells", {}) or {})
if not cells:
raise InterventionError(f"row {node_id!r} has no observable cells")
cells[next(iter(cells))] = None
new["rows"][ridx]["cells"] = cells
return new
if kind == "constraint":
found_constraint = find_stating_constraint(new, node_id)
if found_constraint is None:
raise InterventionError(f"no stating constraint for {node_id!r}")
cidx, constraint = found_constraint
args = list(constraint.get("args") or [])
if constraint.get("predicate") == "MeasureOf" and len(args) >= 2:
args[1] = None
elif constraint.get("predicate") == "Equals" and len(args) == 2:
if str(args[0]) == node_id:
args[1] = None
elif str(args[1]) == node_id:
args[0] = None
else:
raise InterventionError(f"constraint for {node_id!r} has no value site")
else:
raise InterventionError(f"constraint for {node_id!r} has no value site")
constraint["args"] = args
new["constraints"][cidx] = constraint
return new
if kind == "series":
found_series = find_series(new, node_id)
if found_series is None:
raise InterventionError(f"series {node_id!r} not found")
sidx, series = found_series
series["label"] = None
new["series"][sidx] = series
return new
if kind == "entity":
for idx, entity in enumerate(new.get("entities", []) or []):
if entity.get("id") == node_id:
entity["label"] = None
new["entities"][idx] = entity
return new
raise InterventionError(f"node {node_id!r} has no redactable observed value")
__all__ = [
"InterventionError",
"InterventionOperator",
"OperatorName",
"TransformedWorld",
"hidden_nodes",
"set_hidden",
"remove_observed_value",
"transformed_hash",
"with_value",
"world_sha256",
]
|