ZNAK FORGE
ZNAK FORGE is an evidence-gated coding agent runtime for developers maintaining Python repositories. It separates model proposals from local repository authority: the model proposes a repair, while the local runtime controls mutation and checks the result.
What it does
orient -> evidence -> localization -> falsifiable hypothesis -> bounded patch -> independent verification -> final verdict
FORGE controls repository authorization, change budgets, targeted tests, shadow tests, regression tests, rollback, and the final ACCEPT / BLOCKED / UNKNOWN state. Verification executes repository test code; untrusted repositories require an appropriate sandbox or disposable environment.
Hackathon track
Best Apps and Agents is the honest track for the current validated build. The project uses Nebius Token Factory inference with local repository authority, verification and rollback; it does not claim Token Factory Sandboxes.
Public test build
https://znak-forge-demo.znakhumbledrummer.chatgpt.site
This no-key browser build demonstrates the accepted canonical repair and the rejected/rolled-back naive repair. It makes no model call and changes no repository. The live Nebius path remains available from the public source.
Nebius + NVIDIA
The integration uses NVIDIA Nemotron 3 Super 120B A12B (nvidia/nemotron-3-super-120b-a12b) through Nebius Token Factory and its OpenAI-compatible inference endpoint.
Nemotron receives bounded repository context and is prompted to return JSON containing evidence, localization, hypotheses, a patch plan, and verification commands. Repository access, mutation and test execution remain local. This implementation does not use Token Factory Sandboxes, AI Cloud compute, Serverless Endpoints or Serverless Jobs.
Recorded live results and their limits
The fixture contains an inventory bug where normalization-equivalent SKUs silently overwrite each other.
A saved real Token Factory dry-run (20260927T004704Z-1ce102a39e) returned DRY_RUN_NO_SOURCE_MUTATION, exit code 0, with identical inventory.py SHA-256 values before and after. The measured end-to-end run took 31.556 seconds.
A saved real Token Factory apply run (20260927T004815Z-7243136b3f) returned ACCEPTED / APPLIED_VERIFIED, exit code 0. Targeted, shadow and regression checks passed, and the receipt recorded a root-cause verdict of PASS. The measured end-to-end run took 16.670 seconds.
An earlier local deterministic negative run rejected a naive raw-string duplicate repair after the normalization shadow test failed, returning ROLLBACK / ROLLED_BACK with the source restored.
These are results on one controlled fixture, not a general reliability benchmark. Equal hashes establish byte identity for the compared file, not semantic correctness or comprehensive security. No controlled same-task comparison with other models has been completed.
Current video
The linked public video on @ZnakHumbledrummer (video ID BZf4hUlFIYs, approximately 2 minutes 48 seconds) is a continuous recording of actual Windows terminal execution from verified public main acf3a75d4836257fb0c2d997f8cd5a5556e96d43. Commands and tests execute during the recording; it is not a still-panel walkthrough or a sped-up replay.
The correct repair returns ACCEPTED / APPLIED_VERIFIED with targeted, shadow and regression checks all PASS. The repaired inventory tests are run again and both pass. A separate, deliberately incomplete repair passes the targeted test but actually fails the shadow test with AssertionError: ValueError not raised. FORGE returns ROLLBACK / ROLLED_BACK, and an independent comparison confirms that the original source bytes and SHA-256 were restored. The original public command python hackathon/run_local_demo.py also completes with exit code 0 during the recording.
This recording is explicitly OFFLINE and uses bundled deterministic repair proposals. It is not a new Nebius or Nemotron inference call. The narration explains how NVIDIA Nemotron 3 Super through Nebius Token Factory supplies proposals in live mode while the local engine retains write and acceptance authority. The recorded live API results above are separate evidence. This video demonstrates the controlled fixture path, not general agent reliability.
Fresh public-package check: 27 September 2026
A fresh unauthenticated clone of public main at commit acf3a75d4836257fb0c2d997f8cd5a5556e96d43 was tested on Windows with Python 3.11.9.
The portable-demo newline binding regression suite passed 17/17 tests, covering both LF and CRLF fixtures, rejection of unexpected content, exact post-binding identity enforcement, interrupted-run evidence handling, and the strict success gate.
The clean end-to-end portable demo then completed with exit code 0. The positive scenario returned ACCEPTED / APPLIED_VERIFIED with TARGETED_TEST, SHADOW_TEST, and REGRESSION_TESTS all PASS. The negative scenario reached the intended shadow counterexample, returned ROLLBACK / ROLLED_BACK, verified rollback, and restored the original source bytes.
This closes the previously observed LF/CRLF fixture-hash portability mismatch for the tested public commit. These results are evidence for this controlled demo path, not a general reliability benchmark.
Nebius / NVIDIA feedback
The OpenAI-compatible API let us add a small, Python-standard-library provider without granting the remote model direct repository write authority. Structured proposals supported the recorded dry-run and accepted repair. We used prompting around a strict JSON contract, not fine-tuning.
Project-specific onboarding friction was the relationship between billing activation, API-key creation, trial credit and promotional credit. A clearer end-to-end onboarding guide, explicit credit-state feedback and examples combining inference with bounded local execution would be useful.
Further improvements we would value include structured-output reliability on larger repository contexts and explicit token/cost telemetry. Exact costs, token totals and promotional-credit activation are not yet verified. No numerical satisfaction ratings or comparative superiority claims are asserted here.
Would we build with this stack again? Yes, for bounded coding-agent workflows where remote model inference can stay separate from local repository authority. The OpenAI-compatible Token Factory API kept the provider integration small, while Nemotron 3 Super produced structured repair proposals that could be independently verified before mutation was accepted.
Implementation work during September 2026
Work added a Nebius Token Factory / NVIDIA Nemotron provider, bounded secret-filtered context selection, a JSON proposal contract, changed-function budget hardening, local/live demo runners, rollback evidence, and hackathon-specific documentation. The local core build receipt is dated 23 September 2026; this writeup does not assert an unverified pre-26-August project start date.
Repository
MIT-licensed source, setup instructions and the current demo harness: https://github.com/HumbleDrummer/znak-forge-nebius-nvidia-2026
The final hackathon submission is still being prepared.
Log in or sign up for Devpost to join the conversation.