Skip to content

About

Research and evaluation workflow for a privacy-safe, action-safe network agent.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

5 Commits

Folders and files

Repository files navigation

DGX RouterOS Agent RSL Flywheel

I built this as the public record for a research flywheel around a network-focused agent. The private source preserves sanitized corpora, holdouts, retrieval manifests, reports, training runs, and promotion gates. This branch records the evidence discipline without publishing raw corpus text or operational identity.

What I built

  1. Versioned corpus and retrieval layouts with manifests and source references.
  2. Separate train, validation, and holdout material.
  3. Privacy and action-safety checks before promotion.
  4. Evaluation reports that retain failure modes instead of hiding them in one score.
  5. Quarantine and promotion decisions with explicit reasons.
  6. Reproducible build, refresh, evaluation, and receipt-verification scripts.

Recorded results

Observation Source evidence Status
V1/V2/V3 corpus generations are recorded source manifests and README Historical
V3 contains 2,383 sanitized source-grounded public/synthetic examples source README Historical
V3 adapters were quarantined after holdout regression source README/reports Historical decision
v1 restart uses 500/100/100 train/validation/holdout source corpus Historical
v2 uses 3,500/350 plus a frozen 500-case holdout and was rejected source corpus/README Historical decision

Why it matters

A flywheel amplifies useful behavior and mistakes. Holdout contamination, private strings, unsafe actions, or unresolved failures should stop promotion.

Engineering approach

Corpus items carry provenance and sanitization state. Evaluation separates quality, privacy, and action safety. Quarantine is durable until targeted review clears it.

Sanitized architecture boundary

See docs/ARCHITECTURE.md.

Repository map

Evidence rules and limits

Readings are historical private-source evidence. This is a public project interface, not an operational training repository; it contains no raw private corpus, model weights, addresses, accounts, or deployment claims.

License

Gumbii Digital's original code, documentation, examples, data, diagrams, and media are available under the MIT License. Third-party components, assets, product names, and trademarks retain their respective terms; see COPYRIGHT.md for scope.

About

Research and evaluation workflow for a privacy-safe, action-safe network agent.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages