AI-Revenue-OS

Autonomous AI Revenue Ecosystem

src/revenue_os/ecosystem/ is the “brain” that turns real, discovered opportunities into an executable task chain, reusing the existing execution stack (opportunity_store, opportunity_state, execution queue, worker, acceptance, PayPal payment, SMTP delivery, revenue ledger) unchanged.

real sources ─▶ DiscoveryEngine ─▶ Opportunity(origin="real")
                     │
                     ▼
              verification.verify()   ─▶ DISCOVERED..QUALIFIED / REJECTED / HUMAN_REQUIRED / BLOCKED
                     │
                     ▼
            profitability.evaluate()  ─▶ expected profit / profit-per-hour / risk   (every number ESTIMATE)
                     │
                     ▼
              strategy.select()       ─▶ TASK | PRODUCT | AFFILIATE | ECOMMERCE | SERVICE | OTHER
                     │                    (SERVICE is never the default - spec §29)
                     ▼
              pipeline.plan()         ─▶ PRODUCT -> the existing acceptance chain
                                         other   -> a prepared plan, HUMAN_REQUIRED for the external step

Nothing here spends money, posts, contacts anyone, logs in, or creates an account. Every real action still flows through action_class / autonomous_context() / approvals.

Data model

New namespaces on each opportunity_store record (all optional, back-filled on load):

field meaning
origin "synthetic" (test data) or "real" (discovered from a source). Ratchets synthetic→real, never back.
discovery {source, source_url, source_id, access_method, policy_status, evidence[], opportunity_type, verification:{status,reasons,checks}, demand_hint, ...}
evaluation the deterministic profitability projection; carries is_estimate: true
strategy {recommended, options:[{strategy,score,...}], reason, plan:{...}}

Store writers: record_discovery (merge), record_evaluation, record_strategy (replace).

Sources (ecosystem/sources.py)

Every source declares SourceMeta (spec §6): source, source_type, source_url, access_method, automation_allowed, requires_login, requires_human, policy_status.

name real? access notes
synthetic no generated reuses opportunity_engine archetypes; deterministic; stays origin="synthetic"
hn / hackernews yes official keyless API HN demand threads (Ask HN / hiring / “I will pay …”); read-only; the fleet never posts to HN
remoteok yes official public JSON API remote job / gig listings → real freelance demand; read-only, descriptive UA
file yes curated local JSON a human-vouched signal list; fully offline
upwork, fiverr, amazon_associates, shopify — — HUMAN_SETUP_REQUIRED: yield nothing until a human wires an account / API key. The fleet never self-provisions.

Real sources take an injectable fetch_json callable (tests replace it - no network in the suite), same pattern as paypal.py / deploy.py.

Verification (ecosystem/verification.py)

Pure gate. Fail closed. Checks provenance, policy status, evidence, title, type + fleet capability, pay realism. Only a QUALIFIED opportunity can be planned into a real chain.

Profitability (ecosystem/profitability.py)

Deterministic projection. Every output carries is_estimate: true. Headline comparator: decision_value = profit_per_hour × success_prob × (1 − 0.5·risk) × (0.5 + 0.5·automation). A small fast likely task can out-score a big slow uncertain service (spec §9 example).

Strategy (ecosystem/strategy.py)

Scores every viable strategy on: capital-light, speed, automatable, low-platform-risk, repeatable, scalable (spec §28 weighting) plus the economic term and a demand lift. SERVICE carries a 0.80 handicap (spec §29) - it can still win, but only clearly. A non-positive projected profit → no recommendation.

Distribution channel classes (strategy reads them; execution unchanged)

OWNED / SEARCH / MARKETPLACE / AFFILIATE / ADS / COMMUNITY / DIRECT / PARTNER / OTHER - real distribution still runs through the existing distribution.py (Null adapter by default; owned-web only).

Autonomy levels (ecosystem/autonomy.py)

Presentation over action_class (no new policy). Maps an activity to a class (READ_ONLY / RESEARCH / BUILD / DRAFT / PUBLISH / CONTACT / SELL / DELIVER / BUY / PAY / ADVERTISE) and a verdict (AUTONOMOUS_ALLOWED / HUMAN_APPROVAL_REQUIRED / HUMAN_REQUIRED / BLOCKED). Unknown → BLOCKED.

Learning (ecosystem/learning.py)

OutcomeStore (data/ecosystem_outcomes.json, append-only). After an opportunity settles: {strategy, source, category, type, channel, time, cost, revenue, success, failure_reason}. aggregate() rolls it up; priority_weights() = win-rate ÷ overall-win-rate, clamped [0.5, 1.6], only once ≥ 5 outcomes have settled. Plain ratios - not ML.

Simulation (ecosystem/simulation.py)

simulate(n, seed) runs the whole loop over N synthetic opportunities with zero external side effects (no money, network, messages, ads, orders, accounts, or writes to real stores). Deterministic: (n, seed) → byte-identical report. Reports discovery / verification / strategy mix / executed / successes / failures / simulated revenue + profit + per-category analytics.

ExecutionTask integration (spec §24)

New task types DISCOVER, VERIFY, EVALUATE, SELECT_STRATEGY (all SAFE_AUTONOMOUS via task_class). Adapters in ecosystem/task_adapters.py, registered into default_registry(). The worker runs them inside autonomous_context() like any other task.

CLI

revenue_os discover [--source synthetic|hn|remoteok|file|<setup-name>[,...]] [--limit N] [--source-path P]
revenue_os evaluate <OPP_ID>
revenue_os select-strategy <OPP_ID>
revenue_os plan-strategy <OPP_ID>          # PRODUCT -> acceptance chain; other -> prepared/HUMAN_REQUIRED
revenue_os simulate [--n 1000] [--seed 42]
revenue_os ecosystem-status

Safety invariants (unchanged)

Not yet built (next phases)