Business mechanics

worldseq evaluates consequential CEO decisions: diagnosis, causal reasoning, experiments, allocation, growth economics, turnarounds, recovery, replanning, long-horizon commitments, risk, stakeholders and product portfolios. Participants choose commitments and assessments. The engine executes the business consequences.

Company definitions and task scenarios are strict JSON. The reusable kernel lives in src/engine/business, independently of model providers and evaluation adapters. HTTP, MCP, the dashboard and evaluation workers use the same episode transitions.

Operating state

An operating company defines eight possible departments, their weekly capacity and monthly compensation, opening receivables and obligations, customer cohort ages, and independently billed addon products. Department hours reconcile with total staff capacity. A commitment consumes both its department’s capacity and the company’s remaining capacity. Work, contractual lead time and cash are independent prerequisites; unused hours from another department do not become engineering hours.

Named interventions publish their description, cost, staff hours, lead time and prerequisite actions. Their causal effects remain private. They can change staffing, overhead, compliance, product adoption, incident recovery, financing or a measured trial policy. Actions cannot complete twice. Cancelling work preserves sunk costs. Hiring cannot contribute capacity before completion; layoffs preserve severance.

Generic sales_initiative work publishes a total effort budget (hours) and an opportunity ceiling (targetAccounts). Completed work produces at most one opportunity per four sales hours, rounded down and capped by that ceiling. Eight hours can therefore produce at most two opportunities even when the target is 100; 100 opportunities require at least 400 hours and USD 10,000 in incremental execution costs, plus any daily rounding and recurring payroll. Work shares sales capacity with other commitments. Failed or cancelled work retains its costs and produces no opportunities. Prospects still face procurement, trial conversion and payment rules; opportunities are not paying customers. The four-hour productivity constant is an authored modeling assumption, not a measured startup sales benchmark. Finite, segment-specific campaign reach and channel saturation are modeled separately by growth experiments.

Opening cohort ages are sampled by stable customer identity, up to the declared maximum age. Opening paid service is prorated to the remaining current billing period. It creates entitlement and deferred service, but no new episode cash. The original calendar anchor survives short months and leap years. An optional annual-contract share selects one-month or twelve-month base subscriptions by stable customer identity. Annual invoices prepay twelve months at the agreed monthly rate; changing the list price does not reprice a paid contract’s MRR. Addon subscriptions currently use monthly terms.

An addon’s prepaid entitlement lasts until its own paid-through date, including when core payment is being retried or the core subscription has been cancelled. Fulfillment still requires paid core eligibility and the other service prerequisites: an unpaid core term receives no successful addon usage, and failed days continue to accumulate. Core recovery can restore fulfillment within the existing addon term without charging again or clearing those failed days. At the addon’s term end, renewal requires paid core eligibility and the usual retention and collection checks; otherwise the addon terminates. Terminated account/product pairs do not automatically readopt. This preserves paid obligations without granting new free service periods.

Financial truth

All money uses integer cents. The financial verifier is a separate computation path over primitive events and authored contractual terms. It checks every daily record, including day zero, rather than trusting a final cached balance.

Daily objective MRR includes core and addons. Customer counts use unique base economic identities. Buying another product never creates another customer. Opening receivables are already earned: collecting them cannot earn revenue again. Financing retains principal and interest after the episode deadline. A full refund reverses episode-earned service and releases the associated deferred liability.

Day zero initializes historical balances. Later days recognize service on the declared paid interval, with integer remainder handling. Core payment retries retain their existing schedule. Addon subscriptions currently cancel after their first failed renewal; they do not implement the core product’s retry sequence.

Public reports include dated obligations and opening receivables, amounts and settlement status. Receivable due dates are known commitments, not guaranteed collection dates. Their hidden collection-delay coefficient is not published.

Retention fixes the cohort to customers paying on day zero. Its denominator is their original core-plus-addon contracted MRR. Net revenue retention includes expansion within that cohort. Gross revenue retention caps each customer’s current MRR at that customer’s opening amount before summing, so another customer’s expansion cannot conceal contraction. New customers never enter the denominator; a zero denominator gives an unavailable ratio. Three integer-cent metrics expose the denominator and both numerators for objective/forecast evaluation. A separate verifier reconstructs them from customer, invoice, payment and product receipts.

Customer readiness and commitments

Optional onboarding requirements bind specified customer segments to a declared workload. Paid accounts can owe service before they are ready to use it. A support commitment chooses ticket resolution or onboarding and consumes the same department and company capacity; hours cannot count for both. Completion creates readiness, not a payment or a new subscription. Trial demonstrations remain available. Revenue recognition follows the paid license interval; it is deliberately separate from the customer’s readiness to activate the product.

Approved contractual opt-outs disable renewal for a bounded eligible cohort. They preserve service through its paid boundary and cannot create an early refund, erase deferred service or authorize another invoice. Staff-day records freeze the budget before work; an independent verifier reconstructs allocations, prerequisites, readiness and renewal terms chronologically.

Events, uncertainty and observation

Events have an explicit trigger: a dated occurrence with bounded jitter, a delay after another event, after an executed action, or a threshold on the preceding closed business state. Action-triggered failures can require another incident to still be active, making an actual release freeze consequential. Dependencies are acyclic. A zero-delay cascade activates in dependency order, independently of the JSON array order. Separate causes can overlap and recover independently. Recovering service does not erase accumulated support work or losses.

Effects include demand, service availability, payment availability, department availability, addon availability, support pressure and reporting delay. Shared-cause duplicates are combined by their declared effect and target; independent causes compose. Product dependencies propagate per customer’s actual paid entitlement, service capacity and availability, rather than merely changing a dashboard flag. Availability may target declared customer segments. Recovering one segment does not restore another; product dependencies retain the same customer’s exposure.

Preventive actions work only when completed before a trigger and within the declared freshness window. Stale preparation can fail. Event and action effects can irreversibly lose work units; service restoration cannot undo that loss. The loss verifier independently checks receipts, action prerequisites and prevention timing. Lost work is a bounded business-impact measure, not a simulated database.

Official campaign seeds are unpredictable and private. HMAC-addressed random draws use causes, customers, days and semantic decisions. A different run ID or extra read does not reroll the world. Paired policies share the same hidden exogenous schedule. Their decisions can still create different endogenous consequences. Public source code and public task partitions remain learnable development material.

Every observation records its effective day and publication day. Reporting errors modify only an observed metric, never authoritative accounting. Corrections identify the observation they supersede. A delayed correction to an old day cannot replace a newer effective snapshot. Costly diagnostics and experiment results arrive only after their work and reporting delay. The observation layer can also misreport the opening retention denominator; corrected reports preserve the original effective day and unchanged financial truth.

Experiments and assessments

Experiments sample a bounded prospective population from authored control and treatment response rates. They report successes and denominators after completion. These are simulated pilots, not empirical samples of actual SaaS customers. An experiment_rollout intervention deploys the selected arm’s response rate into future real simulator trial conversions; price, cohort and service modifiers still apply. Rollout requires the experiment result to have been published. Thus choosing an ineffective policy has real business consequences. Transfer from the pilot population to the operating population remains an explicit modeling assumption.

Categorical assessments require probabilities for every public candidate, summing to one. The latest eligible answer is scored over the registered observational equivalence groups with grouped Brier loss divided by two. A hidden distinction inside an equivalence group cannot penalize a participant for choosing the other indistinguishable explanation. The unresolved flag is diagnostic by default. A question explicitly declaring requirement.kind: "supported_conclusion" also requires the submitted conclusion set and uncertainty to match the supported class for eligibility. This is separate from the unchanged Brier score; no probability threshold is applied. Current catalog tasks include mandatory conclusions; consult each public question contract. See required assessment contracts.

Forecasts refer to a named metric and an independently verified future day. Their loss is squared error divided by the square of the registered scale. Negative scores remain possible; the score is not clipped. Public schemas bound the input and scale so every accepted forecast has a finite score. Answer deadlines are exclusive: submit before the due day. An exact idempotent retry returns its original receipt. Missing, pending and invalid assessments stay visible as separate outcomes.

Scoring and interpreting results explains the raw primary metrics, assessment losses, eligibility gates and Harbor reward, with worked examples.

Objectives and quality claims

Task objectives are authored before the episode. Participant milestones and rationale cannot replace them. Stability and recovery objectives inspect every day in their inclusive window. A missing day or unfinished window cannot pass. Recovery also measures cumulative undelivered service, so a last-day repair does not undo the customer impact of earlier failure.

Task feasibility, useful interventions, strategic diversity, causal identifiability and real-world calibration are separate claims. An executable scenario is not automatically admitted. Public blogs, disclosures, product documents and incident reports support particular rules and cases. Authored adoption, churn, hazard and response rates remain sensitivity inputs. Source-derived candidate coverage, executed mechanisms and successful behavioral checks are reported separately.

Read evidence and licensing, validation campaigns, and release qualification for evidence, registered comparisons and the acceptance protocol. Integrity checks establish specific tested boundaries; they do not establish universal immunity to reward manipulation.