A hardening patch, not a replacement architecture.

Run deadline and escalation evaluation on infrastructure that does not depend on user traffic or a serverless instance remaining warm.

Maintain an independent freshness monitor outside the scheduled workload. Alert after two missed cadences and on stale heartbeat.

Run an automated post-deploy scheduling probe and block or roll back deployment when the probe fails.

Expose HEALTHY, DEGRADED, FAILED and UNKNOWN scheduler states, last success, next expected run, backlog and oldest overdue item in Mission Control.

Make jobs idempotent, concurrency-safe and replayable. A retry must not create duplicate notices, transitions, fees or assignments.

Use a durable queue and dead-letter mechanism for asynchronous work; alert on age, depth and poison messages; provide controlled replay.

Store deadline source, rule version, timezone, created time, due time, pause/resume history and disposition. Test DST and clock-skew cases.

Define safe behavior for notification failure: the obligation stays visible, retry is bounded, and escalation is not satisfied merely because a send was attempted.

Perform a destructive non-production restore of database, object/document store, audit chain, identities/roles, configuration, secrets references and entity links.

Approve RPO and RTO for Trial 1 and measure achieved values. Restore success requires reconciliation, not merely service startup.

Encrypt backups, restrict restore/delete permissions, log access, and separate backup administration from routine application roles.

Test point-in-time recovery and corrupted or partial-backup scenarios. Maintain retention and geographic/account separation appropriate to risk.

Document dependency restoration order, DNS/credential recovery, communications and decision authority for disaster declaration.

Repeat restore drills on a defined cadence and after material storage or schema changes; track remediation to closure.

Require MFA for internal and privileged users. Prohibit shared privileged accounts; maintain break-glass accounts with alerting and review.

Use server-side authorization on every object and action. Deny by default; derive tenant context from trusted identity, never request parameters alone.

Provision real lender and vendor test identities. Test horizontal/vertical escalation, IDOR, search, export, attachments, APIs, errors and guessed identifiers.

Require step-up authentication and/or dual approval for AI kill-switch changes, Trial resume, privilege grants, vendor re-eligibility, destructive export, restore and future financial actions.

Define joiner/mover/leaver workflow, access expiration, quarterly review and immediate disablement. Log all permission changes.

Protect sessions with secure cookies/tokens, rotation, revocation, idle and absolute timeouts, and reauthentication for sensitive actions.

Represent lender, vendor and company hierarchies explicitly; scope users to organization, branch, portfolio and function as needed.

Ensure support impersonation, if allowed, is time-bound, approved, visibly indicated, fully logged and incapable of bypassing protected approvals.

Disable analytics initialization on authenticated and operational routes. Prefer a strict allowlist of approved public routes; default every new or unknown route to analytics off.

Remove assignment, lender, vendor, borrower, vehicle, document, incident and finance identifiers from URLs where practicable. Never put a VIN, lender reference or human-readable assignment reference in path, query or fragment.

Before any event leaves the browser or server, replace dynamic segments with stable templates such as /assignments/:id, drop query and fragment, suppress page title and referrer, and reject properties outside an allowlist.

Remove or disable hosting-injected Plausible on the application origin, or prove through configuration and captured network traffic that it cannot collect sensitive routes.

Use separate analytics properties for public marketing and any approved product telemetry; restrict access, retention, export, subprocessors and cross-environment mixing.

Add automated route-classification tests to CI. A new route fails closed until explicitly classified public and non-sensitive.

Document lawful purpose, notice/consent position, retention and deletion; prohibit form payloads, record content, free text, identifiers, tokens and secrets in telemetry.

Scan uploads before normal processing. Quarantine suspicious or unknown files; prohibit rendering, indexing or AI ingestion until cleared.

Preserve the original byte stream with cryptographic hash, uploader identity, trusted receipt time, source, entity link and content type.

Never silently replace evidence. Corrections create a new version linked to the prior version with reason, actor, time and approval where required.

Bind material dispatch and eligibility decisions to the exact document versions and credential facts relied upon at decision time.

Validate file signatures rather than extensions; limit type, size and count; sanitize filenames; isolate object keys; use short-lived authorized download links.

Record view, download, export and delete-request events for sensitive evidence. Test cross-tenant file URL and thumbnail isolation.

Retain original photo metadata subject to policy, note transformations, generate derivatives without changing originals, and display provenance.

Define legal and contractual retention, hold, correction and deletion rules. Deactivation must not make historic assignments unintelligible.

Treat document text as untrusted AI input. Separate content from instructions and test indirect prompt injection and malicious embedded links.

Record actor/service/AI identity, delegated user, action, target, before/after or event facts, reason, authority basis, correlation ID, trusted time and rule or model version.

Provide a controlled export that reconstructs assignment chronology, eligibility basis, evidence versions, human approvals, AI recommendations and exceptions.

Restrict and log audit access; redact secrets and tokens and minimize unnecessary sensitive payloads without losing decision proof.

Synchronize clocks and define timestamp semantics. Preserve source time and ingestion time for external events.

Define audit retention, verification ownership, incident procedure and evidence-of-review cadence.

Retain the controlled lifecycle and guard model; all new transitions require authority, guard, audit and regression-test definitions.

Model immediate stop-work triggers — cure, cancellation, recall, bankruptcy, legal or law-enforcement hold, identity conflict and duplicate recovery — as gates, not notes.

Preserve eligibility snapshots and add exact credential, rule and document versions, geographic authority, lender approval and capacity inputs used at dispatch.

Require reason, permission and sometimes second approval for overrides; never permit override of a legally or non-waivably prohibited action.

Create first-class exception cases with type, severity, owner, due date, evidence, permitted actions, resolution and linkage to affected assignments.

Version lender rules, fee schedules, SLAs and compliance policies with effective dates; historic assignments retain the applicable version.

Make routing explainable: list included and excluded vendors with reasons; preserve the candidate set and scoring inputs.

Model vendor geography, credentials, capacity, conflicts, specialization and lender constraints explicitly; avoid free-text eligibility.

Tie communications, notices, approvals and acknowledgements to the assignment chronology; email alone is not the system of record.

Keep domain concepts extensible — asset, assignment, client, provider, credential, event — without building non-vehicle workflows in Trial 1.

Retain the architectural prohibition on AI protected-state writes, autonomous dispatch, privilege elevation and money movement.

A central kill switch must stop AI calls and tools without stopping deterministic safety controls; restrict, log and test kill and resume.

Treat documents, messages and web content as untrusted data, not instructions; test direct and indirect prompt injection.

Allowlist tools and parameters; enforce tenant scope and authority outside the model; validate every tool request server-side.

Instrument recommendation → confidence → human decision → disagreement reason → operational outcome.

Exercise each Trial-used employee on representative, boundary, adversarial, outage and abstention cases before reliance.

Maintain golden evaluation sets and task-specific metrics: field accuracy, calibration, abstention, override and critical-error rate.

Version and approve model, prompt and policy changes; regression-test before deployment and support rollback.

Cloud accounts, billing, DNS, certificates, data stores, object storage and backups are company-controlled. Proof: account inventory, owners, billing/recovery contacts and access review.

Secrets, signing keys, integration credentials and encryption keys are inventoried and recoverable without a personal account. Proof: secret-store inventory, rotation and break-glass drill.

CI/CD, build artifacts, infrastructure-as-code, migrations and deployment runbooks are versioned and company-owned. Proof: successful clean-environment deployment.

Observability, alerting, status, incident and support systems use company accounts and named responders. Proof: on-call test and escalation evidence.

Architecture, data model, state/guard catalog, API contracts, security model, AI contracts, runbooks and ADRs are current. Proof: documentation index with owner and last-reviewed date.

A second competent engineer can deploy, diagnose, pause and restore the platform using documentation. Proof: observed bus-factor exercise with gaps closed.

IP assignment, contractor contributions and third-party license obligations are documented for counsel review.