Mechanism cases
Validate source-bound mechanism fixtures with positive and omission controls.
The evidence corpus has 42 candidate mechanisms. Each has a separate executable
reference fixture in src/catalog/mechanism-cases/, using the production engine
and its independent verifier. These fixtures do not enter the public playable task
inventory, public task listing or benchmark suites. Their purpose is to check a
specific causal claim without giving a retention or contractual case an
inappropriate recovery score.
compileMechanismCases(companies, tasks) builds the fixtures.
probeMechanismCase(testCase, seed) runs their declared counterfactual branches.
admitMechanismCase(testCase, seeds) requires at least two distinct seeds and
produces the standard validation manifest, assertion results and a canonical
mechanism.json artifact. Branch episode IDs differ; semantic randomness stays
paired. Every completed branch must independently reconcile primitive receipts,
dated costs, customer entitlements and all business-day financial records.
Certificates bind the source mechanism, full company/task/branch configuration,
assertion text, executable code and evidence corpus. Source changes during a run
make its results inconclusive. A missing or failing certificate does not become
a pass because the case exists. Catalog inventory checks therefore still report
mechanism admission as not_run; explicit certificate artifacts carry the
executed results.
What the cases establish
| Case family | Executable comparison |
|---|---|
| Cash, annual billing and price units | Paid invoice terms, service periods, collected cash, recognized and deferred revenue, and the units behind recurring value are reconstructed separately. |
| Payroll, reserves and commitments | Severance and dated costs occur before savings; future receipts cannot fund a current commitment; financing preserves its repayment liability. |
| Procurement and product readiness | Typed approval and product prerequisites govern collection or downstream work; early hiring incurs payroll while those prerequisites remain blocked. |
| Exposure and dependencies | Equally many affected groups can have unequal actual customer work; shared paid-product dependencies propagate failures; customer batches recover independently. |
| Recovery work and automation | Capacity on the binding queue changes throughput; unrelated capacity does not. Automation consumes work and cash before increasing later throughput. |
| Access and support | Work blocked by an unavailable route stays blocked until a separately resourced fallback executes. Paid work and completion receipts remain auditable. |
| Restore points and preparedness | Older restore points sacrifice explicit irreversible work for speed; rehearsal unlocks a validated route; stale preparation can expire before a later incident. |
| Replanning and deletion | Cancelling an actual scheduled deployment prevents its action-triggered regression. A stale revision can undo recovery. A staged deletion has a finite reversal window. |
| Trials and causal attribution | Beneficial, null and harmful authored arm effects change future paid customer outcomes after pilot maturity; a concurrent demand increase does not identify a policy effect. |
| Cohort retention | Opening customer membership stays fixed; gross retention caps expansion per customer; corrected denominators supersede mistaken observations without changing ledger truth. |
| Onboarding and contract exceptions | Actual support work makes paid service usable; activation evidence arrives later. Approved opt-outs prevent renewal while preserving paid-through service. |
| Repair, prevention and roadmap choices | Current recovery and future protection consume shared resources but affect different dates; independent roadmap work can remain feasible during an external platform block. |
Some fixtures preserve the full catalog population. Controlled fixtures use smaller deterministic populations and remove unrelated churn, payment failures or onboarding to isolate the tested path. Their full altered company state is bound into the certificate; they are not counted as additional companies. Numeric costs, rates, durations and work-loss units are explicitly authored stress settings, not estimates transferred from a blog post or aggregate report.
Limits that remain explicit
The certificate admits only its registered assertion. It does not establish the whole real-world phenomenon represented by the source candidate. In particular:
- Backup states and lost work are explicit simulator state, not stored database bytes or actual restore tooling.
- Recovery throughput is an abstract work resource, not measured disk bandwidth.
- Product dependency classes stand in for deployment architecture differences; this is not a network topology emulator or a universal canary-size estimate.
- Support and communication cases validate work-path reachability and truthful completion prerequisites. They do not measure external delivery, recipient receipt, subjective trust or communication-caused churn.
- Trial confounding isolates concurrent acquisition volume; demographic selection, policy duration changes and real-customer causal identification remain outside that assertion.
- Public probe seeds and privileged scripted plans establish reproducible mechanism examples. They do not establish model superiority, statistical power, strategic diversity, confidential holdout performance or empirical causal calibration.
These limits accompany each machine-readable certificate. Public source provenance, excerpts, reuse constraints and quarantined claims remain in the evidence corpus. Synthetic educational billing data remains an ingestion fixture and never supplies behavioral calibration.