МАТЧАСТЬ / MASTER IMPLEMENTATION PLAN / DOCUMENT 44 / 02.09.2026

MASTER PLAN реализации

Исполняемый план сборки mathchast.com после завершения исследовательской фазы. Документ переводит стратегию из Docs 1–43 в конкретную структуру репозитория, модули, миграции, API, страницы, workers, AI‑агентов, policies, тесты, observability и последовательность задач. Его задача — дать Codex/Qwen/другим ИИ‑разработчикам достаточно контекста, чтобы они могли по одной атомарной задаче собирать систему без постоянного переосмысления архитектуры.

44 → codeдальше не исследование, а последовательная реализация

1 task = 1 outcomeкаждая задача ограничена файлами, контрактом, тестами и DoD

AI-operated productroutine editorial/analytics выполняются агентами и rules

Paid slice firstпервая цель — реальная компания, реальная публикация, реальная оплата и реальный отчёт

0. Правило запуска

После этого документа не добавлять новые исследовательские блоки до появления первой рабочей вертикальной цепочки. Любая новая идея попадает в backlog, если она не блокирует путь Company → Publication → Proof → AI Visibility → Report → Next Action.

1. Финальный стек

HOST Ubuntu Server 24.04 Docker Compose nginx WEB Next.js 16 Active LTS TypeScript SSR / Server Components API FastAPI SQLAlchemy 2 Alembic Pydantic DATA PostgreSQL 18.x pgvector pg_trgm Redis Redis Streams MinIO AUTH Authentik / OIDC AI Local Qwen 30B External approved model adapters Prompt Registry Typed Tool Gateway OPS OpenTelemetry Prometheus Grafana Loki Blackbox Exporter pgBackRest off-host backup.

2. Monorepo

mathchast/ ├─ apps/ │ ├─ web/ │ ├─ api/ │ └─ worker/ ├─ packages/ │ ├─ contracts/ │ ├─ design-tokens/ │ └─ config/ ├─ db/ │ ├─ migrations/ │ └─ seeds/ ├─ policies/ │ ├─ editorial/ │ ├─ commercial/ │ ├─ security/ │ ├─ ai-tools/ │ └─ retention/ ├─ prompts/ │ ├─ editorial/ │ ├─ verification/ │ ├─ measurement/ │ └─ recommendations/ ├─ evals/ │ ├─ editorial/ │ ├─ security/ │ └─ ai-visibility/ ├─ infra/ │ ├─ compose/ │ ├─ nginx/ │ ├─ monitoring/ │ ├─ backups/ │ └─ scripts/ ├─ docs/ │ ├─ adr/ │ ├─ runbooks/ │ └─ product/ ├─ scripts/ ├─ AGENTS.md ├─ README.md └─ Makefile

3. AGENTS.md — обязательный первый файл

До генерации кода любой ИИ‑разработчик должен читать один короткий authoritative файл с правилами проекта.

# AGENTS.md PROJECT What Mathchast is CORE LOOP Company → Publication → Proof → AI → Action ARCHITECTURE Modular monolith Postgres truth Async workers FORBIDDEN No microservices No Kafka No Neo4j No unrestricted AI shell No cross-workspace access No hidden paid links MODULE OWNERSHIP identity content trust commerce measurement workspace TESTS required commands MIGRATIONS Alembic only SECURITY least privilege RLS typed AI tools DONE tests + metrics + audit + docs.

4. Команды разработчика

make dev make test make lint make typecheck make db-up make migrate make seed make e2e make security-test make eval make build make staging make smoke make backup-test

5. Definition of Done глобальный

FEATURE DONE only if: ✓ implementation works ✓ unit/integration tests ✓ auth checked ✓ tenant scope checked ✓ audit where needed ✓ metrics/logging ✓ error/loading/empty states ✓ migration included ✓ rollback path ✓ no secrets ✓ documentation ✓ AI eval if AI feature ✓ public SEO checks if public page.

6. Task format для Codex/Qwen

TASK ID: M44-XXX GOAL: one concrete outcome CONTEXT: relevant architecture/policy FILES: allowed paths DEPENDENCIES: previous task IDs IMPLEMENT: exact behavior DB: tables/migrations API: routes/contracts UI: screens/components TESTS: required cases SECURITY: checks OBSERVABILITY: metrics/logs DONE: acceptance criteria DO NOT: explicit exclusions.

7. Правило размера задачи

Одна задача должна помещаться в один осмысленный pull request.

GOOD: "Add organizations table, aliases table, repository and GET /companies/{slug}" BAD: "Build Entity Graph" GOOD: "Implement publication state transition DRAFT → REVIEW" BAD: "Build CMS".

8. Этапы реализации

Этап	Результат	Business milestone
0 Foundation	Repo, infra, auth, DB, observability	Можно безопасно писать продукт
1 Identity	Company/entity/source	Компания существует в системе
2 Publishing	Draft → article → SSR	Матчасть выглядит как медиа
3 AI Editorial	Writer/critic/verifier/policy	Routine content автоматизирован
4 Commerce	Order/payment	Можно взять первые деньги
5 Proof	Publication Health/Search	Есть доказательство результата
6 AI Visibility	Prompt runs/citations	Ключевая дифференциация
7 Report/Action	Before/After + NBA	Замыкается repeat loop
8 Hardening	Security/backup/monitoring	Можно масштабировать пилот
9 Agency	Multi-client	Мультипликатор продаж

9. Этап 0 — Bootstrap

M44-001 — Инициализировать monorepo

Create: apps/web apps/api apps/worker packages/* infra/* policies/* prompts/* evals/* docs/* scripts/* Add: README AGENTS Makefile .editorconfig .gitignore CI skeleton.

M44-002 — Docker development stack

Services: postgres redis minio api web worker mail catcher optional Acceptance: make dev starts clean environment.

M44-003 — Environment/config package

Typed config: APP_ENV DATABASE_URL REDIS_URL MINIO_* OIDC_* PUBLIC_URL Fail fast on missing config.

M44-004 — CI baseline

Run: Python lint/type/test TS lint/type/test migration sanity container build secret scan.

10. Этап 0 — Database foundation

M44-005 — PostgreSQL cluster conventions

Create schemas: identity graph content trust workflow commerce measurement workspace audit system.

M44-006 — UUIDv7 / timestamps / enums conventions

Rules: UUIDv7 primary IDs timestamptz UTC created_at updated_at explicit lifecycle enum.

M44-007 — Audit Event base

audit.audit_events: id actor_type actor_id workspace_id action object_type object_id before_json after_json reason created_at.

M44-008 — Transactional Outbox

system.outbox_events: id event_type aggregate payload created_at published_at attempts error.

11. Этап 0 — Auth / Workspace

M44-009 — Authentik OIDC client

Create: prod/staging app callback session logout Test: login/logout/session expiry.

M44-010 — User profile

workspace.users: id auth_subject email name status created_at.

M44-011 — Workspace / Membership

workspaces memberships roles Initial roles: OWNER ADMIN EDITOR VIEWER.

M44-012 — Permission service

Function: can(user, action, resource) No: permission logic duplicated in routers.

M44-013 — RLS baseline

Private tables: workspace_id RLS policy Test: A cannot read B direct UUID attack.

12. Этап 0 — Observability baseline

M44-014 — Structured logging

JSON logs: service request_id trace_id job_id level event error_code. Automatic secret redaction.

M44-015 — /health and /metrics

/health/live /health/ready /metrics Prometheus scrape.

M44-016 — Prometheus/Grafana/Loki

Initial dashboard: web api postgres redis workers.

M44-017 — external black-box

Probe: / known canary page TLS DNS.

13. Этап 1 — Entity / Company

M44-020 — Organization table

identity.organizations: id canonical_name legal_name description country status current_slug.

M44-021 — Entity aliases

entity_aliases: entity_id alias language alias_type active.

M44-022 — Official domains

official_domains: organization_id domain verification_state verified_at method.

M44-023 — Entity identifiers

identifiers: entity_id type value source verified.

M44-024 — Sources

graph.sources: id url publisher source_type retrieved_at status hash.

M44-025 — Claims

claims: subject_entity predicate value state valid_from valid_to claim_sources: claim_id source_id support_type.

M44-026 — Relations

relations: subject relation_type object state sources.

M44-027 — Public company API

GET /api/v1/public/companies/{slug} Returns: identity domains verified facts experts publications.

M44-028 — Company SSR page

Route: /companies/{slug} Includes: name description verification facts sources publications JSON-LD.

M44-029 — Claim company

Authenticated: request claim email/domain proof state machine: PENDING VERIFIED REJECTED DISPUTED.

14. Этап 1 — Verification automation

M44-030 — Domain verification

Methods: email domain DNS token optional website token optional Auto: verify audit.

M44-031 — Public-source verifier agent

Input: company/domain Agent: find official pages extract identifiers extract key claims Output: candidates only.

M44-032 — Claim evidence scorer

Deterministic: source type official status conflict freshness States: VERIFIED SUPPORTED UNVERIFIED CONFLICTED.

M44-033 — Verification exception queue

Only: conflicts ownership disputes low confidence high-risk identity changes.

15. Этап 2 — Publication data model

M44-040 — Publications

content.publications: id type title slug status author company published_at current_version_id commercial_classification.

M44-041 — Publication versions

publication_versions: publication_id version content_json change_reason created_by created_at.

M44-042 — Content blocks schema

Blocks: paragraph heading quote image chart table callout source entity code list.

M44-043 — Publication entity links

publication_entities: publication entity role confidence confirmed.

M44-044 — Topic model

topics publication_topics topic relations later.

M44-045 — Citation/link model

links: url target_type commercial_relation rel_policy source_reference.

16. Этап 2 — CMS

M44-046 — Draft API

POST /workspace/publications GET PATCH version preview submit.

M44-047 — Structured editor

P0: title dek blocks sources entities hero No: full WordPress clone.

M44-048 — Version diff

Show: text changes source changes commercial/status changes.

M44-049 — Preview page

Signed/private noindex same renderer as public article.

M44-050 — Publication state machine

DRAFT → AI_REVIEW → POLICY_CHECK → EXCEPTION optional → APPROVED → PUBLISHED → UPDATED → ARCHIVED.

17. Этап 2 — Public article

M44-051 — Article SSR

/articles/{slug} Render: headline dek author company disclosure body sources entities related.

M44-052 — Canonical / redirects

slug_history 301 old slug one canonical.

M44-053 — JSON-LD

Article Organization Person Breadcrumb as applicable Same PageViewModel as HTML.

M44-054 — Sitemap generator

canonical public indexable substantive lastmod.

M44-055 — OG/social renderer

1200×630 template version title format brand.

18. Этап 3 — AI Tool Gateway

M44-060 — AI provider abstraction

Adapters: LOCAL_QWEN EXTERNAL_A EXTERNAL_B Uniform: generate() structured_generate() embed().

M44-061 — Prompt Registry

prompt_id task_type version template input_schema output_schema provider_policy active.

M44-062 — AI run ledger

model prompt version input hash output usage latency cost status workspace data class.

M44-063 — Typed Tool Gateway

Tools: read_publication read_sources write_precheck suggest_entities create_brief No shell. No arbitrary SQL.

M44-064 — Tool policy engine

task service principal allowed tools workspace bound risk class audit.

M44-065 — Data routing policy

PUBLIC: external/local CLIENT_PRIVATE: local default SENSITIVE: local/redacted SECRET: never model.

19. Этап 3 — AI editorial agents

M44-066 — Researcher agent

Input: brief + sources Output: facts source map unknowns questions.

M44-067 — Writer agent

Input: verified fact package style guide content type Output: structured draft.

M44-068 — Critic agent

Independent: clarity claims unsupported statements commercial bias duplication.

M44-069 — Source verifier

Every material claim: source relation evidence state flag unsupported.

M44-070 — Policy classifier

Commercial? Advertising? Partner? High-risk? Personal data? Regulated? Returns: classification + rule IDs.

M44-071 — AI editorial orchestrator

research → writer → critic → verifier → policy → deterministic validators → auto/exception.

20. Этап 3 — Deterministic validators

M44-072 — Link policy validator

commercial link → sponsored identity/editorial → correct rel.

M44-073 — Disclosure validator

commercial classification must match visible disclosure.

M44-074 — Source coverage validator

Material factual claims must be supported or explicitly framed.

M44-075 — Duplicate/content similarity

FTS/vector candidate → duplicate score → policy threshold.

M44-076 — Security input validator

prompt injection patterns PII/secrets unsafe embeds malicious URLs.

21. Этап 3 — AI evals

M44-077 — Editorial gold set

20–50 examples: good/bad unsupported commercial case research style.

M44-078 — Security eval set

prompt injection tool abuse PII secret cross-tenant fake citation.

M44-079 — Model change gate

Any prompt/model update: run eval compare canary rollback.

22. Этап 4 — Commerce

M44-090 — Product SKU

PUBLISH price currency included features active_from.

M44-091 — Orders

order: workspace client_entity payer advertiser status price_snapshot.

M44-092 — Payment adapter

Provider abstraction: create_payment verify_webhook refund later.

M44-093 — Payment webhook

signature idempotent captured failed refunded.

M44-094 — Order workflow binding

PAID → publication entitlement/workflow No payment: cannot final commercial publish.

M44-095 — Refund exception

P0: human/high-risk reason audit provider action.

23. Этап 5 — Publication Health

M44-100 — Health crawler

Checks: HTTP canonical robots noindex schema assets disclosure links.

M44-101 — Health snapshot

publication_id checked_at check_name state details severity.

M44-102 — Publish event orchestration

PublicationPublished → cache invalidation → sitemap → IndexNow → OG → health → measurement schedule.

M44-103 — IndexNow adapter

event-driven response logged retry does not imply indexed.

M44-104 — Health timeline UI

Published Technical Ready Crawler Observed Search Observed.

24. Этап 5 — Search Proof

M44-105 — GSC connector

URL Inspection Search Analytics source freshness quota data cutoff.

M44-106 — Yandex Webmaster connector

URL state search stats reindex where appropriate quota/freshness.

M44-107 — Search observation model

engine URL observed_at state impressions clicks queries source metadata.

M44-108 — Search Proof UI

No SEO score. Show: state timeline impressions clicks queries last sync.

25. Этап 6 — AI Visibility

M44-120 — Prompt Set

prompt_sets: id entity version language region locked_at.

M44-121 — Prompt records

text type intent brand_status priority source active.

M44-122 — Competitor Set

fixed version entity IDs valid from.

M44-123 — Provider adapters

Start: 2–3 providers Each: request normalize citations metadata error/freshness.

M44-124 — AI Run scheduler

baseline weekly/daily later replicates priority quota.

M44-125 — Answer store

answer provider model/mode run eligible error class raw policy.

M44-126 — Mention resolver

aliases context entity match false-positive eval.

M44-127 — Citation extractor

URL domain watched URL? Mathchast? source class.

M44-128 — Recommendation prominence

PRIMARY SHORTLIST MENTION CAVEAT CITATION_ONLY.

M44-129 — Metric engine

Mention Rate Recommendation Rate SoV Citation Rate Mathchast Citation Rate coverage.

M44-130 — AI Visibility dashboard

Top: core metrics Then: platform matrix sources watched URLs prompts.

26. Этап 7 — Before / After

M44-140 — Measurement Protocol

scope intervention prompt set version competitor set platforms region language primary outcome baseline/post windows.

M44-141 — Intervention event

PUBLICATION SUBSTANTIVE_UPDATE DISTRIBUTION PROFILE_VERIFICATION.

M44-142 — Metric Delta

pre post absolute pp relative optional numerator denominator.

M44-143 — Prompt transitions

ABSENT→PRESENT PRESENT→PRESENT PRESENT→ABSENT ABSENT→ABSENT.

M44-144 — Evidence grade

OBSERVED TEMPORALLY_ASSOCIATED COMPARATIVE QUASI_EXPERIMENTAL EXPERIMENTAL.

M44-145 — Confounder annotations

PR campaign product launch model update search update competitor action site deploy.

27. Этап 7 — Report

M44-146 — Report snapshot

immutable: data cutoff methodology version metrics sources limitations generated_at.

M44-147 — HTML report

Sections: scope what changed Search Proof AI Visibility sources limitations next action.

M44-148 — PDF renderer

async print-safe stored in MinIO versioned.

M44-149 — Report data-quality gate

Before send: coverage freshness provider changes missing source math validation.

28. Этап 7 — Next Best Action

M44-150 — Gap detectors

absence gap citation gap source gap fact gap case gap expert gap.

M44-151 — Recommendation rules

Rule: conditions evidence action_type priority blocking conditions.

M44-152 — Recommendation types

CREATE_ARTICLE CREATE_CASE UPDATE_PAGE VERIFY_CLAIM ADD_EXPERT EARN_MEDIA FIX_FACT.

M44-153 — Recommendation explainer

LLM: turn rule result into clear rationale/brief Cannot: change evidence/priority silently.

M44-154 — Recommendation UI

Action Why Evidence Needs Expected measurable outcome [Create brief].

29. Этап 8 — Automation exceptions

M44-160 — Human Exception model

type risk reason evidence recommendation deadline decision resolved_by.

M44-161 — Exception dashboard

Only: items needing decision Filters: security legal identity content billing.

M44-162 — Adaptive sampling

Sample auto-approved output by: content type model version risk error history.

M44-163 — QA outcome taxonomy

FACT_ERROR SOURCE_ERROR STYLE POLICY ENTITY SECURITY NO_ERROR.

M44-164 — Eval feedback loop

QA error → gold set → prompt/rule fix → canary → rollout.

30. Этап 8 — Backup/security

M44-170 — Production network hardening

Only 80/443 public SSH via VPN/allowlist DB/Redis/MinIO private.

M44-171 — Container hardening

non-root cap_drop no privileged no Docker socket read-only FS where feasible.

M44-172 — Upload quarantine

type size malware parser isolation image re-encode.

M44-173 — SSRF protection

block private IP redirect recheck port allowlist timeouts size limits.

M44-174 — Backup

pgBackRest: WAL daily diff weekly full repos: local HDD off-host encrypted.

M44-175 — Automated restore test

monthly: restore isolated run integrity smoke app record result.

31. Этап 8 — Monitoring

M44-180 — Queue metrics

depth oldest age retry dead-letter processing.

M44-181 — Provider health

success latency 429 5xx schema coverage cost.

M44-182 — Canary article

External monitor: HTTP content marker canonical TLS.

M44-183 — Alert rules

site down DB down backup stale restore failed queue stale provider malformed disk forecast payment invariant.

M44-184 — Auto-remediation

restart worker reclaim job circuit breaker safe cleanup rollback bad deploy.

M44-185 — Incident record/runbooks

auto timeline severity actions verification postmortem.

32. Этап 9 — Agency Lite

M44-200 — Agency organization

agency members roles client workspace links.

M44-201 — Delegation

client entity agency scope granted revoked.

M44-202 — Client switcher

recent pinned needs attention.

M44-203 — Shared credits

batch allocation reserve consume return.

M44-204 — Agency report branding

logo cover intro Mathchast provenance retained.

M44-205 — Pitch workspace lite

prospect domain small AI baseline report promote to client.

33. Frontend implementation order

1 Design tokens 2 Public shell/header/footer 3 Company page 4 Article 5 Business landing 6 Login/workspace shell 7 Publication editor/status 8 Publication Health 9 AI Visibility 10 Report 11 Next Action 12 Pricing/order 13 Admin exceptions 14 Agency later.

34. Design tokens P0

color: Paper / Ink / Graphite / Orange type: Onest + IBM Plex Mono space: 4/8/12/16/24/32/48/64 radius: 4/8/12/18 semantic: bg/text/border/action/status.

35. Core frontend components

Button Input Select Badge Status Card Table Tabs Dialog Toast SourceCard EntityChip VerificationBadge EvidenceBlock Metric ChartShell RecommendationCard.

36. No custom components before need

Если конкретный экран можно собрать existing primitives, не создавать новый design-system component.

37. Public SEO gate

Every public route: SSR canonical metadata structured data if applicable robots correct sitemap accessible heading hierarchy server-visible content.

38. API conventions

/api/v1/public/* /api/v1/workspaces/{workspace_id}/* /api/v1/admin/* /api/v1/webhooks/* Errors: code message details request_id.

39. API auth convention

Public: anonymous Workspace: session + membership Admin: privileged role + MFA policy Webhook: signature/idempotency.

40. Database migration conventions

Additive first backfill switch reads/writes constraint remove later No: large destructive migration inside deploy blindly.

41. Background job convention

Job: id type payload workspace priority attempt idempotency timeout Worker: validate execute record ack.

42. Queue groups

stream:content stream:crawler stream:measurement stream:ai stream:report stream:integration.

43. Priority classes

P0: security/payment/publish P1: client report/high-priority AI P2: crawler/normal measurement P3: backfill/embeddings/experiments.

44. n8n use

Allowed: CRM sync Telegram email workflow lead enrichment internal admin integration Forbidden: core publication state payment truth verification truth AI metric truth.

45. Qwen use in production

Default: private drafting source extraction classification summarization recommendation explanation code/internal ops assistant External model: only when quality gain and policy allows.

46. Qwen use in development

Routine: tests CRUD docs fixtures refactor lint fixes Strong external coding model: architecture-critical complex bugs security-sensitive review.

47. Multi-agent development pattern

Coordinator: selects task Agent A: backend Agent B: frontend Agent C: tests/security Agent D: docs/evals Merge only: after shared contracts and CI green.

48. Do not parallelize database ownership

Two agents editing same migration chain or same domain aggregate without coordination is forbidden.

49. Branching

main: always deployable feature: small short-lived PR: tests self-review AI review merge.

50. Commit discipline

One coherent change clear message migration with feature tests with feature No: "misc fixes" 100-file random commit.

51. Staging

Separate: DB Redis buckets OIDC client provider keys domain Never: prod private data copied casually.

52. Production release checklist

CI green migration reviewed backup healthy staging smoke security gate deploy canary metrics rollback available.

53. First deployment milestone

Own editorial article on mathchast.com rendered from real CMS/database.

54. Second deployment milestone

Real company profile linked to real article and sources.

55. Third deployment milestone

Paid pilot company passes end-to-end workflow.

56. Fourth milestone

+30 report produces an evidence-backed next action.

57. First 30 days

Week	Build	Visible result
1	Repo, DB, Auth, Workspace, Ops	Login + empty workspace
2	Company, sources, claims	Public company page
3	Publication model/editor/SSR	First real article
4	AI editorial + policy	AI-assisted publication

58. Days 31–60

Week	Build	Visible result
5	Commerce	Paid order possible
6	Publication Health	Proof timeline
7	GSC/Yandex	Search Proof
8	AI Visibility adapters	Prompt dashboard

59. Days 61–90

Week	Build	Visible result
9	AI metrics/source analysis	Baseline report
10	Before/After	+30 comparison
11	Next Best Action	Action brief
12	Security/monitoring/pilot polish	5–10 pilot clients ready

60. Sales runs in parallel

Week 1–2: build 200-account list Week 3–4: 20 discovery calls Week 5: sell founding pilots Week 6–8: first paid workflows Week 9–12: reports/repeat.

61. Content runs in parallel

Goal by public launch: 20–50 strong pieces minimum Goal by commercial acceleration: 50–100 Generated: AI-first QA: sample + evidence gates.

62. Brand runs in parallel

Week 1: tokens/wordmark prototype Week 2: article/company Week 3: workspace/report No: months of logo polishing.

63. The first pilot offer

FOUNDING VISIBILITY PILOT Company profile 1 publication Search Proof AI baseline/post +30 report Next Best Action Real payment. Limited founding bonus.

64. Pilot intake

company domain contact goal topic facts sources expert advertiser approval contact.

65. Intake automation

Domain → company resolution → source discovery → public facts → suggested brief → missing data questions.

66. Brief generation

AI asks only: what cannot be found or safely inferred. Goal: 5–10 focused questions not 40-field form.

67. Publication generation

facts/sources → research → draft → critic → source validation → policy → publish candidate.

68. Publication monitoring

publish → health → search → AI observation → report → action.

69. Pilot completion condition

NOT: AI visibility must increase YES: publication delivered measurement delivered methodology valid client sees next decision.

70. First 10 customers — build freeze rule

После каждого клиента новые feature requests не идут сразу в код.

Classify: BLOCKER REPEATED PAIN ONE-OFF NICE TO HAVE WRONG ICP Only: blocker/repeated pain enters near-term roadmap.

71. Product feedback schema

customer stage problem frequency workaround severity requested solution actual need decision.

72. Automation metrics from day one

human_minutes/publication AI_cost/publication exception_rate QA_error_rate time_to_publish report_cost AI_run_cost.

73. Gross margin instrumentation

Order: revenue Costs: AI API GPU estimate payment fee human exception minutes external services. Contribution: computed.

74. Why cost tracking is P0

Automation economics is one of the project's core hypotheses; it must be measured, not assumed.

75. Security metrics from day one

privileged MFA % cross-tenant failures blocked SSRF blocked tool calls secret detections backup age restore age.

76. Data-quality metrics

claims verified % AI run coverage citation parser success provider freshness reports held for data issues.

77. Reliability metrics

availability p95 queue age incidents auto-remediation % backup restore.

78. Product metrics

paid workspaces published report opened repeat monitoring attach next action accepted agency second client later.

79. What to show founder daily

PAID 7 PUBLISHED 11 HUMAN EXCEPTIONS 2 AI COST TODAY ₽... QUEUE healthy PROVIDERS 2/3 BACKUP 47m RESTORE 12d INCIDENTS 0.

80. What NOT to show founder daily

hundreds of normal jobs raw logs all prompts all health checks all crawler requests.

81. Admin exception types

IDENTITY CONTENT LEGAL SECURITY BILLING PROVIDER DATA_QUALITY.

82. Exception SLA

Security: immediate Payment: same day Content: working hours Low-risk identity: 1–2 days But: mostly automation.

83. AI prompt version control

prompts/ task_name/ v001.md v002.md metadata: schema model policy eval score.

84. Policy version control

policies/ editorial.yaml commercial.yaml security.yaml ai-tools.yaml retention.yaml Changes: PR + tests.

85. Rule example

rule_id: COMMERCIAL_LINK_001 when: publication.commercial = true link.external = true then: require_rel: - sponsored severity: blocking.

86. AI tool rule example

task: editorial_precheck allowed: - read_publication - read_source - write_precheck forbidden: - publish - refund - shell - database_admin.

87. Recommendation rule example

rule_id: NBA_CASE_GAP_001 if: high_value_prompts_absent >= 3 competitors_case_citations >= 2 verified_client_case_exists = true then: action = CREATE_CASE priority = HIGH.

88. No opaque rules

Every automated decision visible internally with rule IDs and evidence.

89. Testing architecture

UNIT domain rules INTEGRATION Postgres/Redis/MinIO CONTRACT provider/API E2E user loop SECURITY tenant/tool/SSRF/upload EVAL AI quality.

90. E2E test #1

Create company → verify → create draft → AI review → publish → public article works.

91. E2E test #2

Create order → payment webhook → entitlement → publish.

92. E2E test #3

Published article → health → AI run → report → recommendation.

93. E2E test #4

Agency A user tries Client B object → denied.

94. E2E test #5

External source contains prompt injection → tool gateway blocks → no side effect.

95. E2E test #6

Provider returns missing data → report marked partial → no zero substitution.

96. E2E test #7

Kill worker → auto restart → queue resumes.

97. E2E test #8

Restore backup → app starts → critical records match.

98. Seed database fixtures

company_acme company_long_name_ru expert_1 article_editorial article_commercial article_case prompt_set_40 competitors_5 report_partial report_full.

99. No production secrets in fixtures

Obvious, but must be in AGENTS.md.

100. Search architecture implementation

P0: exact prefix trigram FTS P1: vector semantic Order: entity exact first.

101. pgvector first use

duplicate candidate detection topic similarity source similarity Not: critical identity truth.

102. Crawler P0

HTTP fetch robots policy timeouts redirects content hash canonical source metadata.

103. Browser fallback P1

Only for public pages that truly require JS.

104. Media pipeline P0

upload validate quarantine image re-encode AVIF/WebP OG crop MinIO.

105. AI visual pipeline P1

prompt spec Comfy/model render provenance editorial approval/rule derivatives.

106. Procedure graphics P1

static generator first motion second Inputs: topic/data Outputs: SVG/canvas/rendered video.

107. Email P0

transactional: login/invite if needed approval report security Provider adapter not self-host mail server.

108. Telegram P0/P1

Internal: critical alerts ops notifications External: editorial distribution later.

109. Sitemap strategy

sitemap index articles companies experts topics Generate: event-driven static serve.

110. Robots strategy

Public: allow indexable Preview: noindex Staging: blocked/noindex AI crawlers: explicit policy documented.

111. Launch checklist — technical

✓ domain/TLS ✓ SSR ✓ canonical ✓ sitemap ✓ robots ✓ JSON-LD ✓ auth ✓ DB backup ✓ restore test ✓ external monitor ✓ payment ✓ health ✓ AI provider policy ✓ audit ✓ privacy pages.

112. Launch checklist — business

✓ offer ✓ one SKU ✓ invoice/payment ✓ founding customer terms ✓ editorial policy ✓ commercial policy ✓ methodology ✓ support contact ✓ first 5 prospects.

113. Launch checklist — content

✓ homepage not empty ✓ 20+ strong pieces minimum ✓ several company profiles ✓ methodology ✓ flagship research/case ✓ topic structure.

114. Launch checklist — automation

✓ writer ✓ critic ✓ source verifier ✓ policy classifier ✓ tool gateway ✓ eval set ✓ exception queue ✓ audit.

115. Launch checklist — reliability

✓ site external probe ✓ queue monitor ✓ provider health ✓ disk alerts ✓ backup alert ✓ second critical channel ✓ runbooks.

116. Founder daily operating loop

09:00 exception queue 09:15 sales pipeline 09:30 product/reliability health Then: build/sales No: manual article production queue.

117. Weekly product review

paid repeat exceptions AI QA errors cost incidents feature blockers client feedback.

118. Weekly AI review

model performance prompt changes eval score security failures API cost local GPU usage automation rate.

119. Weekly reliability review

availability error budget queue backups restore incidents pages/noise.

120. Monthly roadmap review

Which hypothesis passed? Which failed? What repeated? What should be deleted? What P1 is now justified?

121. The backlog categories

NOW NEXT LATER PARKED REJECTED Not: everything "planned".

122. Feature admission rule

A feature enters NEXT if: blocks revenue or 3+ ICP customers need it or reduces major human toil or reduces major risk.

123. Otherwise

Park it.

124. Stop conditions

STOP scaling if: repeat weak automation poor margin bad security unreliable backup untested human queue growing AI metrics distrusted.

125. Acceleration conditions

ACCELERATE if: clients repeat agency adds second client monitoring renews human exceptions fall gross margin healthy SLO healthy.

126. When to add first employee

Not because: "startup needs team" Add when: founder is bottleneck in repeatable commercial/operator work and revenue supports role.

127. First likely hire

Commercial/operator or account/product generalist is likely more useful than a large editorial hire, because AI should absorb editorial throughput.

128. When to add engineer

When: AI-assisted solo build becomes bottleneck or operations/quality needs ownership.

129. When to add human editor

Only if: exception volume and quality risk justify Role: policy/quality lead not article factory.

130. Agency phase entry gate

Single-client flow: works repeatedly Tenant isolation: tested Credits: needed At least: 2 real agency design partners.

131. Reputation phase entry gate

Need: verified entities claim/source graph dispute workflow moderation automation real demand.

132. Rating phase entry gate

Need: independent reviews at scale fraud control methodology legal maturity Likely P3.

133. International phase entry gate

Russian market: working economics Then: English payments legal AI regions localization.

134. Infrastructure scaling gate

Measure first: CPU DB queue GPU availability Then: extract worker/host.

135. First likely infrastructure extraction

GPU worker or crawler/browser to second host.

136. No architecture rewrite before metric

No «нам скоро нужен Kubernetes» without concrete failure/scale requirement.

137. Documentation hierarchy

Docs 1–43: product/research reference Doc 44: implementation authority AGENTS.md: daily developer authority Policies: runtime authority ADR: architecture decisions.

138. Conflicts between docs

Priority: latest explicit decision → Doc 44 → relevant specialized doc → older research. Record: ADR/policy update.

139. Code should not depend on reading 44 HTML docs

Extract runtime-critical decisions into repository files before coding each area.

140. First extraction set

AGENTS.md architecture.md editorial.yaml commercial.yaml security.yaml ai-tools.yaml data-classification.yaml retention.yaml metrics.md.

141. Recommended first prompt to coding AI

Read AGENTS.md and docs/architecture.md. Do not change architecture. Implement task M44-001 only. Before coding: list files to create/change. After coding: run required tests. Report: files changed, tests, remaining risks. Do not implement future tasks.

142. Coordinator prompt pattern

You are the implementation coordinator. 1. Read current task. 2. Verify dependencies. 3. Inspect existing code. 4. Produce minimal plan. 5. Assign bounded subtask if useful. 6. Implement. 7. Run tests. 8. Run security checks. 9. Update task state. 10. Stop.

143. Why explicit stop

AI agents tend to «helpfully» expand scope. Stopping after the task protects architecture and reviewability.

144. Progress tracking

tasks/ M44-001.md M44-002.md ... Fields: status owner/agent started completed commit tests notes.

145. Status enum

TODO READY IN_PROGRESS BLOCKED REVIEW DONE DEFERRED.

146. Dependency graph in task metadata

id: M44-067 depends_on: - M44-061 - M44-066 blocks: - M44-071.

147. Automated task selection

Coordinator: find READY tasks whose dependencies DONE sort by critical path assign agent.

148. Critical path tasks

001 → 005 → 009 → 020 → 040 → 050 → 060 → 071 → 090 → 100 → 120 → 146 → 150 → PILOT.

149. Parallelizable tasks

Design tokens Monitoring infra Seed content GTM prospecting Prompt/eval preparation Legal/policy text can run alongside core.

150. What cannot block pilot

perfect logo advanced search agency white-label ratings mobile app 7 AI providers fancy procedural motion.

151. What can block pilot

wrong data broken auth payment failure unsafe AI tool no backup no public article no measurement no report.

152. First paid customer target date

Planning target: as soon as Sprint 5–7 core is usable; do not wait for full 90-day polish. A concierge pilot can use internal/admin tooling while public/client surfaces remain narrow.

153. First 10 customers target

Aim: ~90–120 days if build + sales cadence holds But: quality/repeat > calendar.

154. What we learn from first 10

best ICP best first SKU what report matters monitoring value automation cost agency potential actual objections.

155. P1 backlog after first 10

Agency Lite Recurring Billing References Case confirmation Reader follow/save Scheduled reports Better distribution More Next Actions.

156. P2 backlog after first 30

API SSO/SCIM advanced agency white-label fact accuracy advanced reputation read replica CDN external integrations.

157. P3 backlog after first 100

ratings rankings benchmarks partner directory advanced research data products international multi-region HA enterprise.

158. Final anti-scope list

DO NOT BUILD NOW: social network open comments third-party placement marketplace 100 industries full CRM mobile app Kubernetes Kafka Neo4j Elasticsearch ClickHouse 7 provider connectors ranking score reviews marketplace autonomous root agent.

159. Final technical success criterion

One command can deploy a version; one test suite can prove key invariants; one backup can restore the product; one company can complete the entire paid loop without founder editing database rows.

160. Final business success criterion

After receiving the first report, a client decides to take a second paid action because Mathchast produced useful evidence.

161. Final automation success criterion

Routine publication and measurement work passes through AI + rules with humans seeing only exceptions and QA samples.

162. Final product success criterion

Company ↓ Article ↓ Proof ↓ AI ↓ Report ↓ Action ↓ Repeat.

163. Stop reading, start building

Docs 1–43 дали нам достаточно информации, чтобы перестать проектировать платформу в теории. Следующий правильный артефакт — не ещё одно исследование, а репозиторий с M44‑001.

164. Рекомендуемый порядок первых 15 задач

M44-001 Monorepo M44-002 Docker dev M44-003 Config M44-004 CI M44-005 DB schemas M44-006 ID conventions M44-007 Audit M44-008 Outbox M44-009 Authentik M44-010 User M44-011 Workspace M44-012 Permissions M44-013 RLS M44-014 Logging M44-015 Health/Metrics

165. Затем

M44-020 Company M44-024 Sources M44-025 Claims M44-028 Company page M44-040 Publication M44-046 Draft API M44-050 State Machine M44-051 Article SSR M44-060 AI adapter M44-071 Editorial Orchestrator.

166. Затем

M44-090 Commerce M44-100 Health M44-105 Search Proof M44-120 Prompt Set M44-123 AI Providers M44-129 Metrics M44-146 Report M44-150 Next Action M44-174 Backup M44-183 Alerts.

167. Decision

Утвердить Doc 44 как implementation authority и начать разработку с M44‑001. Все будущие решения должны либо помогать critical path к первому paid end-to-end loop, либо откладываться. Основная организационная единица разработки — атомарная задача с ограниченным scope, dependency list, tests, security, observability и Definition of Done. Qwen/Codex работают через repository rules and typed tasks; routine coding, editorial production, measurement and operations максимально автоматизируются. Humans remain responsible for architecture/policy changes, exceptional legal/security situations and sampling QA. После появления первого working vertical slice начинается paid pilot before broad platform expansion. Это завершает проектирование «Матчасти» и переводит проект в фазу реализации.

168. Первый реальный шаг после документа

CREATE REPOSITORY ↓ WRITE AGENTS.md ↓ IMPLEMENT M44-001 ↓ M44-002 ↓ ... ↓ FIRST ARTICLE ↓ FIRST CLIENT ↓ FIRST REPORT ↓ FIRST REPEAT.