# World Sequencer validation — completed release

**17/17 canonical release gates passed; all794 registered instances independently checked with no exclusions.** This summary is bound to final public source `058f62dc`, execution `9a08e349aff64eb256c3bd3d34c68b7c7b18b83dcd80869bad87fc3a14d5a821` and release inputs `63e19c0c62f231aa7989edec92bdd689f3be184a3dcccfbe6d5eb1169750bdf9`. The final read-only audit completed on October12,2026 at11:35:09 UTC.

The release scope is **349 public tasks across 24 companies**, plus **48 separately governed private reference tasks across 2 fictional companies**. Each task has 2 preregistered realizations: **698 public-task instances +96 private-task instances =794 required instances**. The private tasks cover 7, 30, 90, 180, 365 and 730 days, with 4 tasks per horizon per company. Their task packages and protected realization data are not part of the public download.

The successor uses the original registered instance identities, seeds and gameplay budgets; a source change does not authorize drawing easier cases. Certification under an earlier execution remains historical. Each current realization has a matching independently checked final certificate, including trusted input binding, legal command replay, independent reconciliation, constraints, assessments and required causal branches.

## What the checks establish

- **Bounded business consistency.** Independent reconstruction checks cash, earned revenue, receivables, obligations, service, shared work, payroll, usage and contractual consequences against primitive receipts. Passing results are limited to the checked implementation and inputs.
- **Legal solutions.** A retained author command path establishes feasibility for its exact registered realization and budget. The evaluator accepts other legal paths; matching the author's commands is not a scoring requirement. Feasibility is distinct from the player's available information and from model performance.
- **Information fairness.** The private suite has 48 reviewed case gates covering 20 mandatory questions, with 17 fresh public-information probes on execution `9a08e349aff64eb256c3bd3d34c68b7c7b18b83dcd80869bad87fc3a14d5a821`. These inspect disclosed evidence, deadlines and an indistinguishable-cause control. They do not establish universal identifiability for every possible interaction.
- **Adversarial checker controls.** The required semantic mutation set has 31 controls:12 core, 14 contractual/workforce and 5 usage/funded-growth. The matching current-release mutation gate passed; its receipt and retained artifacts were rehashed in this audit. A detected mutant must fail the intended assertion while its unmodified control passes.
- **Transport bounds.** The tested storage implementation accepts the two preserved 555 MB/574 MB sources without logical source/seal/journal changes, still rejects an explicit 512 MiB reader bound, and rejects actual inputs above the 1 GiB expanded and 128 MiB compressed writer bounds without replacing earlier evidence. Peak full-source transport memory was about 7.84 GiB on the measured Bun 1.3.14/Linuxx64 host. These are bounded transport checks, not permission to reuse old certificates under a new execution identity.
- **Distribution integrity.** The final handoff must bind actual archive and expanded file hashes. Known protected seeds, instance IDs, private task IDs, forbidden runtime/key paths and private-key material patterns are checked for exclusion. This is not a universal secret detector. The host, operator and Docker daemon remain trusted components.

## Evidence that remains separate

Private development campaigns, omission controls, stress sensitivities, earlier release tests and failed attempts remain versioned evidence. A deliberately ineligible omission control can be a successful mechanical test. An authored-rate sensitivity failure reports a model robustness limit; it is not silently converted into a certificate pass or failure. Historical private development included two Moraine stress-outcome failures, which remain disclosed.

The separate current Harbor SDK companion archive is `worldseq-harbor-sdk-examples-058f62dc.tar.gz` (SHA-256 `f7f4c2a47d3304826a38dc159c262657b1d9139187d77cf093f7afb4edafdbc0`). Its 19 public files and exact bundled release SDKs were audited. Two retained Linux/amd64 Harbor 0.24.0 runs exercised the TypeScript and Python players. Both produced valid artifacts, ineligible outcomes and reward0, with nine metered calls and one accepted decision despite a duplicate request. Each run's 1,170 trusted file hashes were checked. These are transport/isolation demonstrations, not successful task-solving claims or substitutes for the canonical Harbor release gate. The earlier failed TypeScript startup remains operator-only evidence.

The final source/runtime/SDK archives passed independent byte, manifest, path and known-private-material checks. All13 native qualification receipts and listed artifacts were rehashed; the last native gate finished at10:47:03 UTC on October12,2026. The canonical Docker Harbor gate passed three tests/17 assertions at10:52:31 UTC; standalone passed the unchanged `bun run release:smoke` at10:54:18 UTC. The public certificate gate independently checked698/698 instances and passed at11:27:59.794 UTC; the final handoff gate passed at11:28:01.063 UTC. The separately governed private complete check passed96/96 at11:27:39.655 UTC, with actual exit0 and no pause. Its original12-worker/4-heavy process completed; no takeover was used.

The final audit matched every ordered registered identity, with zero exclusions, and rehashed all96 private certificate files plus579 current execution files. Private certification ran from the retained f5eb checkout at the same actual9a execution; this does not relabel that checkout as058f or its older release inputs as63e. Final public distribution and qualification are bound separately to058f/63e. A supplementary clean macOS source consumer passed five documented commands with verified logs; it is not Linux certificate authority.

## Installed SDK runtime coverage

Actual installed-package contract tests passed on **Node.js22.22.0, Node.js24.13.0 and Bun1.3.14** (14 tests each), and **CPython3.11.14,3.12.12,3.13.11 and3.14.2** (19 tests each, including sync and async clients). These contract tests include mocked HTTP and retained real-response fixtures. Full live examples against the packaged server and PostgreSQL ran on **Node.js24.13.0 and CPython3.11.14 only**, with both Python sync and async flows. Those examples were valid but ineligible illustrative API workflows, not successful business strategies. They conditionally submit an unresolved categorical assessment and exercise assessment idempotency when that question is available. The evidence does not claim live lifecycle runs on every interpreter. The sanitized `SDK-RUNTIME-MATRIX.json` records this distinction and its supporting hashes.

## Limits

The environment models finite, explicitly authored business rules. Source material supports mechanisms; it does not measure every fictional coefficient, joint event probability or future customer response. The evidence does not establish optimal agent performance, universal exploit resistance, unrestricted horizon support, real-company equivalence or real-world predictive accuracy. Public playbooks alone are not held out: protected compositions, requirements, parameters and realizations are separately governed, with development exposure retained.