Solution: DeepSeek core instrument factorial

Подход#

Исторические Flash/strong prompts, scorers и artifacts не изменяются. Новый sibling harness:

загружает frozen baseline prompts через hash-checked boundary;
детерминированно вставляет один semantic-invariant block только для treatment;
строит четыре configs из prompt × reasoning;
materializes один LLM call и один raw SaveText sink на request;
оценивает raw bytes прежним scorer;
объединяет scorecard с metadata-hidden semantic review;
применяет frozen selector;
materializes confirmation только для eligible config.

Структура#

prototype/
├── artifact-publish.ts   # atomic no-overwrite bundle publication
├── manifest.ts           # 24 + conditional 18 matrix, gates and ranking
├── prompt-instrument.ts  # exact baseline/treatment composition
├── transport-capture.ts  # terminal Comfy batch capture contract
├── transport.ts          # request/workflow/job/raw/telemetry binding
├── source-evidence.ts    # fresh recomputation from captured raw bytes
├── semantic-packet.ts    # source-bound blinded review packet
├── factorial-evidence.ts # eligibility, selection and factorial effects
├── finalize.ts           # full-chain screening/confirmation finalization
└── run.ts                # preparation and CLI orchestration

Команды#

Из корня репозитория:

npm test
npm run typecheck

После live catalog capture:

node \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/prototype/run.ts \
  prepare-screening \
  --catalog-preflight \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-preflight-v2/catalog-preflight.json \
  --output-root \
  <fresh-output-root>

Canonical paid screening использует уже проверенный root artifacts/derived/screening-preflight-v2; повторная команда выше нужна только для byte-for-byte audit и не должна писать поверх него.

Provider execution не имеет отдельной локальной CLI-команды. Operator:

передаёт 24 prepared workflows в один submit_batch({client_os: "darwin", confirm: true, items});
сохраняет returned batch ID и request-to-job map до ожидания;
вызывает wait_for_batch для того же batch до 24 ready / 0 failed;
вызывает get_batch_output ровно один раз после terminal state;
скачивает bytes по job ID, не нормализуя JSON или newline;
записывает capture-v1 без signed URLs; недоступную telemetry оставляет null.

Transport failure после definite submission не разрешает второй submit. Screening run следовал этому контракту: одна successful submission, 24/24 terminal jobs и 0 provider failures.

После сохранения terminal batch capture и exact raw responses:

node \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/prototype/run.ts \
  materialize-transport \
  --prepared-manifest \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-preflight-v2/manifest.json \
  --batch-capture \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-batch-capture.json \
  --raw-root \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/raw/screening/responses \
  --output-root \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-transport

node \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/prototype/run.ts \
  score-screening \
  --raw-root \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/raw/screening/responses \
  --prepared-manifest \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-preflight-v2/manifest.json \
  --transport \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-transport/transport.json \
  --output \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-scorecard.json

node \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/prototype/run.ts \
  prepare-semantic-review \
  --phase screening \
  --prepared-manifest \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-preflight-v2/manifest.json \
  --transport \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-transport/transport.json \
  --scorecard \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-scorecard.json \
  --raw-root \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/raw/screening/responses \
  --output-root \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-semantic

node \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/prototype/run.ts \
  finalize-screening \
  --prepared-manifest \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-preflight-v2/manifest.json \
  --scorecard \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-scorecard.json \
  --transport \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-transport/transport.json \
  --raw-root \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/raw/screening/responses \
  --private-map \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-semantic/semantic-private-map.json \
  --semantic-review \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-semantic-review.json \
  --output-root \
  RnD/technical/2026-07-26-deepseek-core-instrument-factorial/artifacts/derived/screening-finalization

prepare-semantic-review создаёт blind packet без request/config/model/language, role, seed, prompt и reasoning metadata. Reviewer записывает отдельный review; private map объединяется с ним только внутри finalize-screening.

Conditional confirmation sequence#

Эта ветка допустима только для selection.status=selected, ненулевого selectedConfigId и соответствующей записи rankings[] с eligible=true. Текущий live selector вернул no_eligible_config, поэтому команды ниже не выполнялись и paid confirmation calls равны 0.

study=RnD/technical/2026-07-26-deepseek-core-instrument-factorial
runner="$study/prototype/run.ts"

node "$runner" prepare-confirmation \
  --catalog-preflight "$study/artifacts/derived/screening-preflight-v2/catalog-preflight.json" \
  --selection "$study/artifacts/derived/screening-finalization/selection.json" \
  --screening-prepared-manifest "$study/artifacts/derived/screening-preflight-v2/manifest.json" \
  --screening-scorecard "$study/artifacts/derived/screening-scorecard.json" \
  --screening-transport "$study/artifacts/derived/screening-transport/transport.json" \
  --screening-raw-root "$study/artifacts/raw/screening/responses" \
  --screening-private-map "$study/artifacts/derived/screening-semantic/semantic-private-map.json" \
  --screening-semantic-review "$study/artifacts/derived/screening-semantic-review.json" \
  --output-root "$study/artifacts/derived/confirmation-preflight-v2"

# После 18/18 sequential dry-runs, одной confirmed provider submission,
# terminal capture и exact raw download:
node "$runner" materialize-transport \
  --prepared-manifest "$study/artifacts/derived/confirmation-preflight-v2/manifest.json" \
  --batch-capture "$study/artifacts/derived/confirmation-batch-capture.json" \
  --raw-root "$study/artifacts/raw/confirmation/responses" \
  --output-root "$study/artifacts/derived/confirmation-transport"

node "$runner" score-confirmation \
  --raw-root "$study/artifacts/raw/confirmation/responses" \
  --prepared-manifest "$study/artifacts/derived/confirmation-preflight-v2/manifest.json" \
  --transport "$study/artifacts/derived/confirmation-transport/transport.json" \
  --selection "$study/artifacts/derived/screening-finalization/selection.json" \
  --output "$study/artifacts/derived/confirmation-scorecard.json"

node "$runner" prepare-semantic-review \
  --phase confirmation \
  --prepared-manifest "$study/artifacts/derived/confirmation-preflight-v2/manifest.json" \
  --transport "$study/artifacts/derived/confirmation-transport/transport.json" \
  --scorecard "$study/artifacts/derived/confirmation-scorecard.json" \
  --raw-root "$study/artifacts/raw/confirmation/responses" \
  --output-root "$study/artifacts/derived/confirmation-semantic"

node "$runner" finalize-confirmation \
  --prepared-manifest "$study/artifacts/derived/confirmation-preflight-v2/manifest.json" \
  --scorecard "$study/artifacts/derived/confirmation-scorecard.json" \
  --transport "$study/artifacts/derived/confirmation-transport/transport.json" \
  --raw-root "$study/artifacts/raw/confirmation/responses" \
  --private-map "$study/artifacts/derived/confirmation-semantic/semantic-private-map.json" \
  --semantic-review "$study/artifacts/derived/confirmation-semantic-review.json" \
  --selection "$study/artifacts/derived/screening-finalization/selection.json" \
  --screening-prepared-manifest "$study/artifacts/derived/screening-preflight-v2/manifest.json" \
  --screening-scorecard "$study/artifacts/derived/screening-scorecard.json" \
  --screening-transport "$study/artifacts/derived/screening-transport/transport.json" \
  --screening-raw-root "$study/artifacts/raw/screening/responses" \
  --screening-private-map "$study/artifacts/derived/screening-semantic/semantic-private-map.json" \
  --screening-semantic-review "$study/artifacts/derived/screening-semantic-review.json" \
  --output-root "$study/artifacts/derived/confirmation-finalization"

Path ownership#

Layer	Caller chooses	Harness enforces
Provider audit	dry-run/submission evidence paths	exact prepared workflow bytes and manifest order
Capture	batch-capture path и raw root	capture-v1 exact keys, unique completed jobs, canonical base64, raw byte equality
Transport	fresh output root	`transport.json`, copied `batch-capture.json`, one job metadata file per request
Semantic	fresh output root и separate review path	`semantic-private-map.json`, `semantic-review-packet.json`, exact blind join
Finalization	fresh output root	screening evidence, factorial analysis, selection; confirmation evidence/verdict when eligible

Trust boundaries#

Model ID должен быть exact live catalog match.
Preparation отклоняет invalid catalog evidence и non-empty output.
Prompt hashes, workflow hashes, config, language, reasoning и seed записаны в prepared manifest.
medium передаётся в dynamic node input model.reasoning_effort; это не текстовая подсказка.
Terminal capture хранит exact rawResponseBase64; эти bytes являются authoritative source, а внешний raw-файл обязан совпасть byte-for-byte.
Transport-v2 связывает prepared request, workflow, payload, batch/job, captured raw, job metadata и доступную telemetry.
Scoring заново вычисляется из captured raw через frozen evaluator и должен byte-for-byte совпасть с сохранённым canonical scorecard.
Prepared, transport, semantic и finalization bundles публикуются атомарно, без overwrite и с отказом на symlink target/ancestor.
Selector не принимает средний score вместо all-cell hard gate.
Confirmation запрещена без selected eligible config и повторной проверки полного screening source chain.

Ограничения#

Comfy node не предоставляет hard output token cap. Seed является hint. Reasoning может увеличить output-token cost. Недоступные telemetry fields остаются Unknown, а не оцениваются.

Условие production-переноса#

Только successful all-role confirmation позволяет перенести выбранную role-specific prompt/reasoning policy в отдельную production-задачу. Harness, research fixtures и provider adapter не становятся production authority.