Validation: DeepSeek core instrument factorial

Текущее состояние#

Study status: complete
Outcome: fail
Paid screening calls: 24 / 24
Paid confirmation calls: 0 / 18, not run by preregistered stop
Selected config: none; no_eligible_config
Operational blocker: none; confirmation is forbidden by the frozen selector

Criteria matrix#

Criterion	Состояние	Evidence
AC-01	pass before paid calls	protocol-v2 + identity, canonical preflight-v2 audit, prompt/scorer/rubric/source hashes frozen; v1 preserved with 0 paid calls
AC-02	pass	24 unique submitted jobs, 24 terminal ready, 24 capture-authoritative raw files, 0 missing/failed; full transport binding recomputed
AC-03	pass	frozen scorer evaluated 24/24; independent metadata-hidden review supplied 24 exact aliases/hashes
AC-04	pass	prompt, reasoning and interaction effects published from the complete 2×2 design
AC-05	pass	full-chain selector deterministically returned `no_eligible_config`; no reduced denominator
AC-06	not run — preregistered stop	no config met deterministic 6/6 + semantic floor 6/6 + zero critical failures, so confirmation was forbidden
AC-07	pass with explicit null telemetry	model/prompt/reasoning/seed/workflow/job/raw hashes bound; unavailable credits/tokens/latency/retries stored as `null`; provider label drift preserved

Frozen protocol evidence#

Проверка	Фактический результат
Protocol v2 file SHA-256	`1806fee835210fcd1c902b3bea12c827145acc59eba67a0e9d33d732b149b223`
Factorial manifest hash	`3a36376394e7723adfc969ed9b1c991778a7b5003b7b7f8e2152ce0eff495482`
Canonical prepared manifest file	`034aa6a40bbffe798fdbc64522ca44f17e1832c526b4be1d5b468a7eb88af82c`
Canonical prepared tree	`6ba93cc0137dee0962b910ce520aef243455a6b5eabb7e92c0a8468415fc5c35`
Preflight v2 audit file	`ced0c45798ccf7a8006b5de7a418dc5bb37d671579094e3af6411f4ed90a618a`
Live catalog exact-model status	`confirmed_exact_match`; 33 entries; options hash `aa6ec51e7958356c88c214ac63f35ae7a44aa5dcb6bb6946bd94871b81ab73b3`
Semantic invariant block SHA-256	`15fc3866392d1cfb1a3143af62fafef0a691f7565155719e67c30c049f391c47`
RU baseline / treatment prompt	`07fedbf8ec56e8f8b85db60338561294f0418b9ef8bb09c1a096cec7643d1c7b` / `ecd7a19207acb34b28668bb10bbfa3e1fd6716cefec93ef3ef1d09d087c87958`
EN baseline / treatment prompt	`cb995c1879d845b2ed2690bcde5c18e4c2b4803fddc882e5778c16213883ef03` / `2188edf9c458bad58a0b307ab47491a960898f1dc3c102534af99d73413282e1`
Frozen evaluator file	`22f2877444133b149624018cea155efac0749025eaf70343a1e2811dd4f9f425`
Frozen semantic rubric	`3d90905f2649ad52663775805cfd7152313d272306fbb38d9ee3f4cbd35cb3f3`
Repository tests	139 technical + 5 research-site, 0 failed
TypeScript	`npm run typecheck`: exit 0

Pre-paid checks#

TECHNICAL: 139 tests passed, 0 failed
RESEARCH SITE: 5 tests passed, 0 failed
TYPECHECK: exit 0
PREFLIGHT: 72/72 prompt/workflow/payload bindings
REPRODUCTION: 50/50 files, 0 byte mismatches

Проверки доказывают:

exact 24-cell matrix и unique request IDs;
prompt delta не меняет output contract;
off/medium отличаются только dynamic reasoning input;
selector сохраняет all-cell hard gates;
capture-authoritative raw и transport-v2 не допускают подмену source bytes;
fresh recomputation и full-chain finalization отклоняют forged scorecard;
все evidence bundles публикуются atomic/no-overwrite и отклоняют symlink targets/ancestors.

Terminal screening evidence#

Проверка	Фактический результат
Submission	один confirmed batch; 24 unique request-to-job bindings
Terminal provider state	24 ready, 0 pending, 0 failed
Raw capture	24 exact files, 191730 bytes, no extras
Transport	24 records; SHA-256 `f65f9452a33f4c50c27e1c6afa877f7cb94fd0fc0564c68a832a726d11d83b3b`
Deterministic config passes	baseline/off 4/6; semantic/off 4/6; baseline/medium 3/6; semantic/medium 5/6
Blind semantic review	24 reviews; score 0: 12, score 1: 9, score 2: 3; critical: 12
Selector	`no_eligible_config`; confirmation not materialized

Config-level verified join:

Config	Observed	Deterministic	Semantic ≥1	Semantic 2	Semantic total	Critical
`baseline-v1--off`	6	4	1	0	1	5
`semantic-v2--off`	6	4	4	1	5	2
`baseline-v1--medium`	6	3	3	1	4	3
`semantic-v2--medium`	6	5	4	1	5	2

Counts reconcile: 4 × 6 = 24, deterministic passes 16/24, semantic-floor passes 12/24, semantic-2 cells 3/24, critical failures 12/24.

Terminal artifact bindings#

Artifact	SHA-256
batch capture	`253197e4f7e12c639f218e2dbd05d0e0292145a255d804ccde4b31ee661cc832`
provider transport audit	`00ee107c4275cd8bbc611adf8251cc2060e71009cc2bc010eba429ef39ed858d`
scorecard	`dac5eb5d71bcd750c4cbabdd3833d0533261be4fdf8a83e9cb7d4f69e492f86c`
blind semantic review	`ed5df980bda88c12af6dffa53e622240dd1e6afde1bdbd39829eec374c72e536`
screening evidence	`0d57129271a97b59e0eac9ed96a4922e45b01d380b0694ce1e0d75fbb9bdb074`
factorial analysis	`87e5e93c8e7db734a4b5a2b69e02741e8ce7af261226523a28510ef5686f6d22`
selection	`1654c760effda3f3cb6f8bb5605f7d241a68b508d54ad8921d261ac5e61b4e73`

MCP не сообщил actual credit delta, token counts, per-job latency или retry counts. Capture фиксирует эти пять полей как null; значения не оценивались.

Submitted и terminal batch tokens не byte-equal: provider добавил один дефис в label последнего job. IDs всех 24 jobs, terminal set и output set совпали exactly, поэтому request-to-raw binding остаётся однозначным. Расхождение сохранено в provider transport audit.

Fresh terminal verification#

Проверка	Результат
Full-chain recomputation	fresh `finalize-screening` из canonical raw bytes повторил все три output SHA-256 byte-for-byte и снова вернул `no_eligible_config`
Independent integrity audit	24/24 exact request → job → raw → score → blind review → evidence bindings; 9/9 terminal artifact hashes и 17/17 header/file bindings совпали
`npm run check`	5/5 research-site и 139/139 technical tests; typecheck exit 0; 13 studies valid; 109 documents built; 309 HTML pages, 32435 local links и 3764 anchors checked; 0 failures
Local Chrome E2E	library search → click `TECH.IDX` → rendered report; `complete`, `fail`, `24/24`, `5/6`, `no_eligible_config` и confirmation stop видимы; console errors 0
Cloudflare Pages	`npm run deploy` exit 0; canonical production report повторно проверен в браузере
390×844 production E2E	report → click `Проверка и итог` → terminal validation; document width 390 px, horizontal overflow false; console errors 0
Design detector	bounded `impeccable detect` завершён; три non-blocking findings: inherited black skip-link color, em-dash cadence и AC-01…AC-06 sequence

Verdict#

fail. Полный verified screening показал полезный средний prompt effect, но ни один config не прошёл all-cell hard gate. Лучший deterministic результат semantic-v2--medium равен 5/6, с semantic floor 4/6 и двумя critical failures. Confirmation и production transfer запрещены; следующий шаг требует отдельного исследования FactLedger или bounded repair без ослабления gates.