# G04 — Brad / Outsource Access 75–100-piece daily adaptation blueprint **Status:** Design blueprint only. No account creation, publishing, deployment, outreach, or Marketing Machine edit. ## Executive decision Brad/OA should treat **75–100 platform-native pieces per day as a gated production target, not a day-one quota**. The defensible Gary/Vayner pattern is an authentic-source learning loop: capture → select → transform for platform/handle → accountable review → distribute → measure → learn. The evidence does not prove that Gary’s self-reported 343-post day was audited, efficient, or produced business outcomes. The recommended OA target model is 50–100 daily iPhone inputs, 20–22 selected moments, and 3.75–4.55 final variants per selected moment. It preserves immutable source-to-placement lineage, uses agents for proposals, deterministic workers for stateful execution, and humans for identity, rights, claims, quality, and public approval. This is **Sterling inference**, not Gary’s current org or demonstrated OA throughput. ## 1. Counting contract Keep three units separate: 1. **Source asset:** one original iPhone photo/video/audio capture with creator, time, context, rights, sensitivity, and hash. 2. **Platform-native piece:** one final approved creative variant with an intentional hook, copy, crop/edit, thumbnail/cover, format, or context for a defined audience. This is the **75–100/day target unit**. 3. **Placement:** one publication of a piece to a specific platform + handle. Stories, comments, replies, and community interactions are separate. Blindly reposting one binary to seven platforms creates seven placements, not seven native pieces. This follows Gary’s verified recommendation to change copy, thumbnail, hook, and context (G01 V-05) without using his self-reported 343 number as a benchmark (G01 S-01). ## 2. Capacity math All figures are **Sterling planning assumptions**, not observed Gary/OA performance. | Wave | iPhone assets | Selected moments | Selection | Variants/moment | Pieces/day | |---|---:|---:|---:|---:|---:| | Pilot | 20 | 8 | 40.0% | 3.00 | 24 | | Controlled scale | 50 | 16 | 32.0% | 4.00 | 64 | | Target low | 75 | 20 | 26.7% | 3.75 | 75 | | Target high | 100 | 22 | 22.0% | 4.55 | 100 | The system should select strong moments rather than force content from every input. ### Human capacity and review workload Planning assumptions: intake/context/rights 45–75 min/day; concept triage 40–60 min; adaptation/editing 2.5–5 min/piece; production QA 45–75 sec/piece; queue operations 30–45 min. - **75 pieces:** about **6.0–10.8 human hours/day**, including **0.94–1.56 QA hours**. - **100 pieces:** about **7.3–13.4 human hours/day**, including **1.25–2.08 QA hours**. - Size internal capacity for **90/day** at the 75 target and **120/day** at the 100 target (20% headroom). Production capacity never automatically raises public posting caps. - This requires a cross-functional pod with at least two trained humans covering editing, QA, and operations—not one unsupervised agent. **Brad review envelope:** during the pilot Brad reviews all 24 pieces, approximately 24–36 minutes/day at 60–90 seconds each. At scale, a trained OA reviewer still checks every piece. If 10% route to Brad at two minutes each plus a 10-minute batch decision, Brad’s load is approximately **25 minutes at 75 pieces** and **30 minutes at 100 pieces**. This reduced load is a proof gate; if exceptions exceed 10%, volume must hold. ## 3. Legitimate handle taxonomy Every handle needs a real accountable owner, truthful bio, editorial charter, authorized account, moderation owner, and retirement/recovery plan. | Class | Legitimate pattern | Control | |---|---|---| | OA corporate | Official Outsource Access brand | OA ownership and contact route visible | | Brad identity | Brad’s real professional identity | Brad named; posts attributable | | OA service line | OA + an actual named service/product | OA-operated; claims approved | | Industry vertical | “OA for [approved industry]” | OA ownership visible; never pose as an independent trade body | | Locality/community | OA/Brad in an approved real locality | Real relationship; no fake-resident voice | | Expertise/interest | Real Brad/OA expertise or recurring interest | Accountable identity and narrow charter | | Event/series | Real OA event, show, research series, or campaign | Dates, organizer, lifecycle, archive plan | Start with existing accountable identities. Add at most one new handle class per wave and only after separate explicit approval. Pause any handle after repeated corrections, unclear ownership, moderation backlog, recommendation/account-status problems, or no incremental qualified signal. **No fake people, fake residents, deceptive independent communities, or account farms.** ## 4. Roles, agents, and decision rights These are functional lanes, not Gary’s org chart. - **Brad/source owner (human):** authentic source and missing context; final authority on personal identity and sensitive exceptions. - **Managing editor (human):** priorities, exclusions, concepts, release window; can stop a batch. - **Intake/provenance worker (deterministic):** hashes, metadata, duplicates, state; can quarantine but not publish. - **Transcription/vision agent:** proposes transcript, scene tags, identities, moments, and confidence; suggestions only. - **Editorial agent:** proposes audience, persuasive/community job, concept, evidence needs, and contraindications. - **Platform adaptation agents:** draft variants against locked brand/format rules; no account token or publishing rights. - **Editor/producer (human):** media, pacing, captions, crop, design, and platform fit. - **Identity/handle steward (human):** ownership, locality/interest truth, disclosure, voice; veto right. - **Claims/rights reviewer (human):** claims, permissions, client/private data, logos, music, sensitive content; veto/quarantine. - **QA reviewer (human):** checks every public piece against source and brief. - **Approval service (deterministic):** exact version-bound receipts; invalidates approval after content/config change. - **Publisher adapter (deterministic):** prepares/schedules only approved immutable versions; live use remains separately approved. - **Analyst/learning agent:** recommends tests from outcomes, edits, and rejects; cannot alter rules or publish. - **Community operator (human):** replies, moderation, escalation; no generic automated comments, DMs, or invitations. **Boundary:** durable software owns identity, lineage, state, permissions, and audit; agents interpret and propose; deterministic workers execute validated jobs; humans own ambiguous judgment and public accountability. ## 5. Queue and state machine Canonical flow: `INGESTED → NORMALIZED → RIGHTS_CHECKED → SELECTED → CONCEPTED → VARIANT_DRAFT → EDITED → QA_REVIEW → APPROVAL_REQUIRED → APPROVED_VERSION → SCHEDULE_READY → PUBLISHED → OBSERVED → LEARNING_RECORDED` Side states: `DUPLICATE`, `QUARANTINED`, `REVISE`, `REJECTED`, `EXPIRED`, `PUBLISH_FAILED`, `TAKEDOWN`, `DEAD_LETTER`. No item skips rights, QA, or required approval. Editing approved copy, media, handle, timing, audience, or configuration invalidates the receipt. ### Queue controls - Intake may accept 50–100 assets/day, but selected-moment WIP is capped at **24**. - Draft capacity is **90–120 pieces/day**, depending on wave. - If QA/approval backlog exceeds one operating day, stop derivative generation and clear the queue. - Retry idempotent technical failures only; identity, rights, policy, and factual failures route to humans. - Per-platform/handle posting limits are configuration, never inferred from production capacity. - Expire stale context before publishing; a kill switch halts scheduling/publishing while preserving evidence. - Idempotency key: `campaign_id + source_id + moment_id + platform + handle_id + format + creative_version`. Media hashes and copy similarity provide secondary duplicate warnings. ## 6. Source-to-derivative lineage Canonical graph: `source_asset → source_segment/moment → canonical_concept → creative_variant → approved_version → placement → performance_observation → learning_record` Minimum fields: - **Source:** immutable ID/hash, original handle/path, capture time, creator, people/brands, location precision policy, rights, sensitivity, transcript/vision version. - **Moment:** exact frame/time range, quoted or visual basis, context, claim/evidence links, confidence, selection decision. - **Concept:** audience, job, claim, CTA class, prohibited interpretations, reviewer. - **Variant:** platform, handle, format, hook, copy, transformations, thumbnail/cover, disclosure, model/prompt/tool versions, parent IDs. - **Approval:** exact content/config hash, approver, scope, timestamp, expiration, decision, reason, invalidation link. - **Placement:** provider ID, handle, scheduled/published time, URL/status, failure/takedown receipt. - **Observation:** metrics with window and retrieval time; qualified/business events linked separately. - **Learning:** evidence set, recommendation, confidence, counterexamples, human acceptance/rejection, next experiment. Canonical truth remains outside agent memory and vendor-specific state. Derivatives may be regenerated; source evidence, receipts, and history cannot be overwritten. ## 7. Daily operating rhythm 1. Brad captures naturally and adds people/context/sensitivity/intended-use notes. 2. Deterministic ingest registers, deduplicates, extracts metadata, and screens private/client data. 3. The managing editor selects up to the wave cap and defines audience + job. 4. Agents propose a small variant matrix; no forced every-platform/every-handle fan-out. 5. Humans edit and QA every public candidate; exceptions route to identity/claims/rights owners. 6. Brad reviews the full pilot batch, then exceptions + release summary only after gates are proven. 7. The deterministic adapter prepares idempotent approved jobs; live publishing remains separately authorized. 8. Metrics are captured in defined early/later windows without overwriting old observations. 9. Learning review promotes, holds, or suppresses patterns; raw views never auto-rewrite policy or publish rules. ## 8. Failure modes and controls | Failure | Control / stop rule | |---|---| | Placement count masquerades as unique creative | Enforce source/piece/placement counting and reconcile daily | | Intake overwhelms QA | Halt generation when backlog exceeds one day | | Hallucinated person, role, place, or event | Source-linked claims; identity steward; quarantine low confidence | | Client/private data leak | Default-private ingest; automated flags + human crop/blur/reject | | Rights/music/logo misuse | Rights state required before selection; never “assumed safe” | | Fake/misleading handle | Reject; visible ownership/disclosure; separate creation approval | | Blind duplicate posting | Native-change requirement; hash/similarity warnings | | Approval drift | Content/config hash automatically invalidates receipt | | Wrong account/handle | Immutable provider resource-ID canary before every live write | | Timeout creates duplicate | Idempotency key plus provider receipt reconciliation | | Automated engagement error | Human operator; no auto-comment/DM/invite | | Eligibility/takedown warning | Pause affected queue; inspect first-party status; never evade enforcement | | Views rise but business quality falls | Balanced scorecard; no views-only volume promotion | | Brand fatigue/cannibalization | Frequency/novelty review, holdouts, retire weak variants/handles | | Model/prompt drift | Version everything, canary changes, rollback on correction spikes | | Provider outage/partial response | Pause, reconcile, retry only proven non-side-effect jobs | ## 9. Staged rollout and proof gates ### Phase 0 — Definitions and dry run (5 business days) Lock counting, source schema, handle registry, rights/sensitivity rules, states, receipts, idempotency, and dashboards. Run real private intake with **no live placements**. Canary one source → three variants; revise one; prove approval invalidation; simulate timeout; prove no duplicate job. **Gate:** 100% lineage, sensitive material quarantined, kill switch and invalidation pass, no public action. ### Phase 1 — 24 pieces/day on existing accountable handles (10 business days) 20 inputs → 8 moments → 3 variants. Brad reviews all; OA records every correction/rejection. Any live publishing requires separate approval and route proof. **Proposed Sterling gate:** ≥95% complete provenance; zero severe identity/privacy/rights incidents; factual/identity correction rate <2%; rework/reject <15%; Brad review ≤45 min/day; backlog <1 day. ### Phase 2 — 50–64 pieces/day and one controlled handle expansion (10 business days) 50 inputs → 16 moments → 4 variants. Add at most one separately approved legitimate handle. QA checks 100%; Brad reviews exceptions + release summary. Compare incremental qualified signal, not aggregate views. **Proposed gate:** Phase 1 safety holds; Brad review ≤30 min/day; 90% of jobs finish in the daily window; duplicate placements = 0; queue age <1 day; at least one theme/format shows incremental qualified attention. ### Phase 3 — 75–100/day target (minimum 20-business-day proof window) 75–100 inputs → 20–22 moments → 75–100 pieces. Keep 20% headroom. Add handles one at a time. Use holdouts and versioned experiments to distinguish topic, hook, format, platform, and handle effects. **Continuation gate:** zero severe incidents; clean route/idempotency reconciliation; Brad review ≤30 min/day; QA/rework and queue gates hold; qualified conversations/business signals justify marginal labor. Reduce volume if business signal does not improve, even when reach rises. ### Phase 4 — Selective automation only after proof Automate stable low-risk transformations and scheduling preparation—not identity/rights ambiguity or unsolicited interaction. Keep allowlists, caps, version-bound approvals, kill switch, receipts, and rollback. This blueprint authorizes **no** autonomous account creation, posts, comments, DMs, or invitations. ## 10. Scorecard Track daily/weekly by source, concept, piece, placement, platform, and handle: - Inputs: sources, selected moments, selection rate, rights/sensitivity. - Production: pieces, native-change dimensions, cycle time, editor/QA minutes, reject/rework, cost. - Governance: provenance completeness, approval invalidations, exceptions, incidents, duplicates prevented, failures/takedowns. - Attention quality: retention/watch quality, saves, shares, meaningful comments, profile visits—not views alone. - Community: substantive conversations, response time, escalations, negative feedback. - Business: identified inquiries, meetings, owned-audience growth, opportunity quality, attributable revenue with an explicit window. - Learning: promoted/suppressed patterns, confidence, counterexamples, holdout lift. ## 11. Marketing Machine read-only alignment The Marketing Machine state was read, not edited. Its completed foundation describes campaign/version objects, graph nodes/edges, evidence, lifecycle/health/approval signals, and a visual canvas. Pending stages already point toward immutable manifests/approval receipts (S08), bounded queues (S09), idempotent dry-run adapters (S10), retry/outage/kill-switch rehearsal (S11), route proof (S12), controlled deployment (S13), and a Brad approval packet (S14). Later handoff mapping: - source/moment/concept/variant/placement/observation/learning → campaign graph + evidence; - role lanes → agent/worker/decision/approval/wait/channel/goal nodes; - content/config hashes → manifest + approval receipt; - WIP/backpressure/idempotency/dead letter → durable queue; - private pilot/failure canaries → synthetic rehearsal; - live publishing/account creation → separately approval-gated. ## 12. Evidence, alternatives, and limitations - **Verified Gary guidance:** historical pillar-to-micro-content, platforms + handles, accountable locality examples, platform-native changes, community participation, and human checking of automated replies (G01; G02-01/-03/-04/-05; G03). - **Self-report/corporate position:** 343 posts, seven-to-eight-platform prescription, and VaynerMedia’s high-volume/local/social-first learning flywheel (G01 S-01/S-02; G02-11/-12/-17/-18). These are not audited OA benchmarks. - **Current platform boundaries:** Instagram/TikTok/YouTube/LinkedIn recommendation, eligibility, and connection guidance (G02-21–G02-24). They do not guarantee reach or authorize automation. - **Sterling inference:** all capacity ratios, labor, gates, roles, state machine, schema, rollout, and controls. The local technology-intelligence library was current enough to query, but the narrow Cloudflare query for content workflow approvals, queues, provenance, and media returned no directly relevant records. No Cloudflare capability was promoted as evidence. Keep architecture composable: canonical lineage in durable storage, media in object storage, deterministic queue/workflow workers, replaceable AI adapters, and platform-specific publishing adapters. Re-verify lifecycle, security, pricing, and account routes before implementation. The QVC/TikTok record (G02-15) is company-reported commerce precedent—not proof Brad should copy its scale or economics. No source proves OA can sustain 75–100 safely, that extra handles improve qualified outcomes, or that Brad’s review load will hold. The staged canaries are the proof mechanism. **Rejected shortcuts:** 343/day as target; every input becomes content; identical one-click cross-posting; agent-owned accounts/publishing; fake locality or independent-looking niche accounts; views-only optimization; or one vendor owning canonical truth and agent memory. ### Current source links - https://garyvaynerchuk.com/the-garyvee-content-strategy-how-to-grow-and-distribute-your-brands-social-media-content - https://garyvaynerchuk.com/content-marketing-strategy - https://garyvaynerchuk.com/instagram-for-business-180-strategy-grow-business-brand - https://garyvaynerchuk.com/here-are-five-reasons-why-automating-on-social-media-sucks - https://vaynermedia.com/social-first-marketing-models-vaynermedia - https://vaynermedia.com/regional-marketing-strategy-vaynermedia - https://about.instagram.com/blog/announcements/instagram-ranking-explained - https://newsroom.tiktok.com/en-us/how-tiktok-recommends-videos-for-you - https://support.google.com/youtube/answer/141805 - https://www.linkedin.com/help/linkedin/answer/a702683 ## Verification - Used one bounded evidence set: 29 prior G01/G02 records, G03 synthesis, one read-only Marketing Machine state, and one narrow technology-library query. - Public evidence was deduplicated once by canonical URL in G02 and not reacquired. - Included capacity math, legitimate handle taxonomy, roles/agents, queueing, lineage, review workload, failure modes, staged rollout, canaries, and Marketing Machine mapping. - Marked all design math and thresholds as Sterling inference. - Preserved the correction that `gary-attention-deck.pdf` is HTML, not a valid PDF; it was not used. - No email, external draft, post, comment, DM, invitation, deployment, spend, credential/account change, customer-data mutation, GitHub action, third-party contact, or Marketing Machine edit occurred.