Agent objective
- Interact with the supplied stateful environment.
- Produce verifier-checkable actions or artifacts.
- Maximise scalar reward under the package contract.
A stateful coding-agent environment for engineering and operating a Forth virtual machine. Its twelve stages progress from arithmetic, stacks, compilation, and control flow through persistent images, migration, forensic recovery, multi-session services, portability, adversarial robustness, namespace transfer, and optimizer discovery, with private procedural families and independent replay discouraging fixture-specific patches.
A stateful coding-agent environment for engineering and operating a Forth virtual machine. Its twelve stages progress from arithmetic, stacks, compilation, and control flow through persistent images, migration, forensic recovery, multi-session services, portability, adversarial robustness, namespace transfer, and optimizer discovery, with private procedural families and independent replay discouraging fixture-specific patches.
Enough detail to understand the intellectual terrain; generated instances, hidden mechanisms, and solution paths remain inside the private package.
| Environment | Mathematical or technical frontier | Adaptive research problem |
|---|---|---|
| Stages 0–3 | VM, definitions, control flow, memory/VFS | Implement strict stack semantics, compilation, loops, allocation, virtual files, quotas, confinement, and structured failures. |
| Stages 4–6 | Persistent images, migration, forensics | Serialize complete state canonically, migrate older schemas idempotently, and salvage damaged images without false claims. |
| Stages 7–8 | Stateful service and architecture change | Preserve session isolation, cloning, crash replay, cell-width/endianness portability, segments, and old-image compatibility. |
| Stages 9–10 | Adversarial robustness and family transfer | Survive exhaustion and parser attacks, then generalize wordlists and search order across unseen namespace structures. |
| Stage 11 | Optimizer discovery | Improve constant folding and superinstructions while preserving traces, errors, semantics, image compatibility, and hidden-case gains. |
A fresh micro-tier calibration compared the intentionally incomplete starter with the trusted reference across all 12 stages. The starter averaged 0.1717 and passed none; the reference averaged 0.9998 and passed all. The final optimizer stage starts from a working incumbent and still fails the threshold until it delivers a revalidated speed–quality gain, making it genuine algorithm discovery rather than stub completion. Validation matched 332 checksums and passed 36/36 tests.
We publish aggregate behavior and task structure, while withholding generated instances, hidden labels, exact successful probes, private checks, and solution trajectories.
Shown with its provenance and limitations; it is not a performance guarantee.
Reference evaluation across all 12 stages. The untouched starter averaged 0.1717 and passed 0/12.
As identified by the supplied artifact.
12 reported runs.
notes.md
Machine-readable provenance and the exact displayed metric are available in results.json.
The paid ZIP will live in a private R2 bucket. Vercel authorizes the buyer and issues a 2–5 minute object URL; R2 serves the bytes directly.
Authenticated buyer + entitlement check
+ private R2 object + 2–5 minute signed URL
= direct, auditable download
Package SHA-256
e3c3f8997eb880be50b25c5a2f24eac59c01ce7e1a574535b631dc9b1abed560One purchase licenses this identified item to one legal organisation for worldwide, perpetual commercial model training, evaluation, research and development. Redistribution and resale of the package are not permitted.