Контракт интегратора
00 / module apiCONSUMES
Compo0.Scene
Compo0.Constraint[]
Compo0.StylePreset
Compo0.Seed
Compo1.00 role descriptors
Compo1.01 weight model
Compo1.02 placement provider
Compo1.03 pair geometry/metrics
Compo1.04 cluster geometry/metrics
Compo1.05 rhythm candidates/metrics
Compo1.06 grid candidates/metrics
Compo1.07 topology candidates/metrics
Compo1.08 noise field policy/metrics
Compo1.09 contact/conflict validation
Compo1.10 color/tone candidates/metrics
Compo1.12 scoring provider? ← DEFERRED
PROVIDES
PointSceneCandidate[]
PointSceneState
DependencyState
DirtySet
PassReport[]
CorrectionRequest[]
RejectedCandidate[]
AuditEvent[]
seedUsed
iterationCount
convergenceState
unresolvedInputs[]
DOES NOT OWN
point.visual_weight
point.local_contrast
point.placement_score
pair.* metrics
cluster.* metrics
rhythm.* metrics
grid.* metrics
conflict.* metrics
noise.* metrics
pointSystem.composition_score
global.* metrics
final winner
1.11 оркестрирует владельцев, но не переопределяет их метрики.
Главная архитектурная поправка
01 / no circular dependencyPASS A — intrinsic
До появления окружения точка получает только свойства, не требующие контекста.
1.00 assign roles
1.01 PASS_A_BASE
radiusShortN
visualWeightBase
NO contrast modifier
NO isolation modifier
NO crowding modifier
NO edge modifier
PASS B — contextual
После placement, соседей, grid, conflicts и local background вес пересчитывается владельцем 1.01.
after 1.02..1.10:
1.01 PASS_B_CONTEXTUAL
consume owner-provided modifiers
compute visualWeightContextual
if deltaWeight > epsilonWeight:
mark dependent modules DIRTY
Запрещённый цикл: нельзя требовать contextual visual weight до размещения точки, если этот contextual weight зависит от размещения. 1.11 обязан сначала построить сцену по
visualWeightBase, затем выполнить contextual recompute.Полный последовательный pipeline
02 / deterministic execution order0 INPUTnormalize scene, seed, canvas, constraints
1 ROLESCompo 1.00 functional + visual role candidates
2 BASE SIZECompo 1.01 PASS A: radius + base weight
3 PLACECompo 1.02 placement candidates
4 RELATE1.03 pairs + 1.04 clusters + 1.05 rhythm
5 STRUCTURE1.06 grid + 1.07 topology + 1.08 noise
6 VALIDATE1.09 contacts / tangencies / protected geometry
7 COLOR1.10 sample background + color/tone candidates
8 CONTEXT WEIGHT1.01 PASS B using fresh owned context
9 DIRTY RECOMPUTEre-run only modules invalidated by changed state
10 SCORECompo 1.12 aggregates owned metrics; 1.11 does not
11 CORRECT / HANDOFFapply repair requests, iterate, then pass candidates global
repeat passes 6..11 while:
hardViolations exist
OR scoringProvider requests repair
OR dirtySet is not empty
stop when:
dirtySet empty
AND hardViolations == 0
AND (scoreDelta < epsilonScore OR scoringProvider unavailable)
OR iterationCount >= maxIterations
IMPORTANT:
if Compo 1.12 is unavailable:
output LOCALLY_VALID_UNRANKED candidates
DO NOT invent score
DO NOT choose "best" by hidden heuristic
Contextual weight loop
03 / explicit feedback cycleКонтекст не переписывает формулу веса сам. Он поставляет владельцу 1.01 свежие входы, после чего 1.01 пересчитывает contextual weight.
Freshness contract
ContextInput {
metricId,
ownerModule,
ownerVersion,
value,
sourceStateHash,
computedAtPass,
status: FRESH | STALE | MISSING
}
1.01 PASS B may consume only FRESH values.
STALE value → omit + mark unresolved.
MISSING value → omit + never invent neutral fake value.
Это не даёт модулю использовать старый contrast после перемещения точки или старый crowding после изменения соседей.
Dirty graph и точечный recompute
04 / dependency invalidation| Что изменилось | Нужно инвалидировать | Почему |
|---|---|---|
point.positionN | 1.03, 1.04, 1.05*, 1.06, 1.07, 1.08*, 1.09, 1.10, 1.01B | меняются расстояния, окружение, конфликт, фон и contextual weight |
point.radiusShortN | 1.03, 1.04, 1.06, 1.07*, 1.08*, 1.09, 1.10*, 1.01B | меняются surface gap, occupancy, contact clearance и sample footprint |
point.color | 1.10, 1.01B | меняется local contrast → contextual weight |
visualWeightContextual | 1.04, 1.06, 1.07, 1.08 | эти модули могут использовать effective fresh weight |
grid | 1.02*, 1.05*, 1.06, 1.07* | grid может быть candidate provider или structural constraint |
protectedZones | 1.02, 1.08, 1.09 | меняется feasible geometry |
stylePreset | все модули, использующие STYLE_PRESET | политика кандидатов и диапазоны могли измениться |
* — только если конкретный объект/кандидат использует соответствующую зависимость. Dirty propagation должна быть по dependency flags, а не слепо запускать всё.
markDirty(change):
affected = dependencyGraph.query(change.field)
for module in affected:
if module.dependencyPredicate(change.sceneState):
dirtySet.add(module)
recomputeDirty():
modules = topologicalSort(dirtySet)
for module in modules:
module.recompute()
updateStateHash(module.outputs)
dirtySet.clearResolved()
Генерация кандидатов
05 / bounded searchDeterministic base
Сначала генерируются структурные кандидаты из ролей, grid, anchors и разрешённых областей.
baseCandidates =
roleProviders
+ gridProviders
+ anchorProviders
+ relationProviders
Seeded variants
Затем разрешён bounded jitter, изменение размера и локальных вариантов.
rng = PRNG(hash(seed, candidateId, variant))
variant = base + boundedDelta(rng)
Hard filter first
Невалидные кандидаты не получают эстетического score.
if hardViolation:
reject(candidate)
log(reason)
skip scoring
INITIAL_HEURISTIC Количество кандидатов — инженерный бюджет, а не композиционный закон. Стартово допустимо
candidateBudget = clamp(12 + 3*N_semantic, 24, 128); для noise-field используется отдельный генератор 1.08.Corrective pass
06 / repair instead of guessДопустимые операции
MOVE(point, delta)
RESIZE(point, radius)
RECOLOR(point, colorCandidate)
ROLE_SWITCH(point, allowedRoleCandidate)
DROP(optionalPoint)
REGENERATE_SUBSET(groupId)
SNAP_TO_GRID(point, gridCandidate)
UNSNAP(point)
CHANGE_RELATION(pair/profile)
Любая операция обязана перечислить, какие модули становятся DIRTY.
Источник correction request
HARD owner → mandatory repair/reject
Compo1.12 → score-directed repair request
global solver → contextual override
user lock → immutable / constraint
style preset → low-priority preference
1.11 не изобретает причину коррекции. Он исполняет запрос владельца правила и ведёт audit log.
applyCorrection(request):
assert request.owner != null
assert request.reasonCode != null
assert request.priorityClass != null
if conflictsWithHigherPriority(request):
rejectCorrection(request)
else:
mutateCandidate(request.operation)
markDirty(request.changedFields)
recomputeDirty()
if scoringProvider exists:
rescoreViaCompo1_12()
Random — только последний слой
07 / reproducible variationSEED
reproducible
Одинаковый вход + seed + module versions → одинаковые candidates.
δ
bounded
Random не выходит за диапазоны owner-модулей.
LOW
priority
Hard, semantic и global rules всегда выше random.
LOG
audit
Каждое stochastic решение сохраняет seed fragment и выбранное значение.
Worked example: почему нужен второй проход
08 / visual proofДо контекста точка имеет только intrinsic/base weight.
Соседи меняют isolation/crowding. Эти данные принадлежат контекстным владельцам.
Correction меняет position, поэтому 1.03/1.04/1.06/1.09/1.10 и 1.01B должны пересчитаться.
STATE A
visualWeightBase = 0.41
visualWeightContextual = MISSING
STATE B
localContrast = 0.82 [owner: 1.10]
crowdingModifier = 0.68 [owner: 1.04/1.08 context]
edgeModifier = 1.00 [owner: 1.02]
visualWeightContextual = 0.31 [owner: 1.01]
deltaWeight = 0.10 > epsilonWeight
ACTION
mark dependent modules DIRTY
generate MOVE candidate
recompute owners
ask Compo 1.12 to rescore
NO hidden "designer intuition" is inserted.
Deterministic tests
09 / orchestration tests| Test | Input event | Expected orchestration result |
|---|---|---|
| T01 | fresh input, no context | run 1.01 PASS A; never request PASS B before placement/context exists |
| T02 | point position changed | invalidate relation/context modules; 1.10 background sample becomes stale; then 1.01B stale |
| T03 | color changed only | rerun 1.10 → 1.01B; do not rerun geometry unless contextual weight change invalidates weight-dependent modules |
| T04 | Compo 1.12 missing | return valid candidates with status LOCALLY_VALID_UNRANKED; no invented best candidate |
| T05 | hard conflict from 1.09 | reject or mandatory repair before score aggregation |
| T06 | same input + same seed + same versions | candidate ids, stochastic deltas and audit trail deterministic |
| T07 | stale localContrast after move | 1.01B must omit it; must not reuse previous value |
| T08 | maxIterations reached with dirty set | status NON_CONVERGED + diagnostics; never silently call result final |
DATA FOR LAYOUT ENGINE
10 / machine contract{
"module": "Compo 1.11",
"version": "2.0",
"name": "point_system_generator_orchestrator",
"scope": "generation, dependency scheduling, corrective execution, handoff",
"consumes": [
"Compo0.Scene",
"Compo0.Constraint[]",
"Compo0.StylePreset",
"Compo0.Seed",
"Compo1.00.*",
"Compo1.01.*",
"Compo1.02.*",
"Compo1.03.*",
"Compo1.04.*",
"Compo1.05.*",
"Compo1.06.*",
"Compo1.07.*",
"Compo1.08.*",
"Compo1.09.*",
"Compo1.10.*",
"Compo1.12.ScoreProvider?"
],
"provides": [
"PointSceneCandidate[]",
"PointSceneState",
"DependencyState",
"DirtySet",
"PassReport[]",
"CorrectionRequest[]",
"RejectedCandidate[]",
"AuditEvent[]",
"ConvergenceState"
],
"owns_metrics": [],
"owns_orchestration_counters": [
"iterationCount",
"hardRejectCount",
"dirtyModuleCount",
"unresolvedInputCount",
"candidateCount",
"convergenceDelta"
],
"deferred_inputs": {
"pointSystem.composition_score": "Compo 1.12",
"global.balance_score": "future global owner",
"global.hierarchy.score": "future global owner",
"global.negative_space.score": "future global owner",
"global.eye_flow.score": "future global owner"
},
"pass_order": [
"NORMALIZE_INPUT",
"ASSIGN_ROLES_1_00",
"BASE_WEIGHT_1_01A",
"PLACEMENT_1_02",
"PAIR_1_03",
"CLUSTER_1_04",
"RHYTHM_1_05",
"GRID_1_06",
"TOPOLOGY_1_07",
"NOISE_1_08",
"CONFLICT_1_09",
"COLOR_1_10",
"CONTEXT_WEIGHT_1_01B",
"DIRTY_RECOMPUTE",
"SCORE_1_12_IF_AVAILABLE",
"CORRECTIVE_PASS",
"HANDOFF"
],
"freshness_policy": {
"stale_external_metric": "OMIT_AND_MARK_UNRESOLVED",
"missing_external_metric": "OMIT_NEVER_INVENT",
"contextual_weight": "ONLY_FROM_FRESH_OWNER_INPUTS"
},
"candidate_budget": {
"formula": "clamp(12 + 3*N_semantic, 24, 128)",
"status": "INITIAL_HEURISTIC",
"noise_points": "managed separately by Compo 1.08"
},
"convergence": {
"epsilonWeight": "consume Compo1.01 epsilonWeight",
"epsilonScore": {"value":0.005,"status":"INITIAL_HEURISTIC"},
"maxIterations": {"value":8,"status":"INITIAL_HEURISTIC"},
"withoutScoringProvider": "LOCALLY_VALID_UNRANKED"
},
"rules": [
"PASS_A_BASE_WEIGHT_BEFORE_CONTEXT",
"PASS_B_CONTEXT_WEIGHT_AFTER_1_02_TO_1_10",
"NO_STALE_METRIC_REUSE",
"NO_HIDDEN_SCORE_AGGREGATION",
"HARD_CONFLICTS_BEFORE_SCORING",
"EVERY_MUTATION_MUST_PROPAGATE_DIRTY_FLAGS",
"SAME_INPUT_SEED_AND_VERSIONS_MUST_BE_REPRODUCIBLE",
"NON_CONVERGED_RESULT_MUST_BE_EXPLICIT",
"GLOBAL_SOLVER_HAS_FINAL_AUTHORITY"
],
"finality": "POINT_SYSTEM_CANDIDATE_GENERATOR_ONLY"
}