AIGC — Governed Gemini Opportunity Assessment

RegenExcalibur's evidence-first Gemini workflow separates structured assessment, deterministic policy, human review, and external action.

Comment

Inspiration

Opportunity assessment often mixes useful AI analysis with risky assumptions, private data, unclear authority, and pressure to act quickly. AIGC explores a more accountable pattern: separate analysis from action, bind claims to evidence, and make deployment and publication explicit human decisions.

What it does

The repository implements a dedicated Gemini opportunity-assessment path. It validates and sanitizes an operator-supplied intake, rejects recognized secret-like material, requests a strict structured assessment, validates the model's input-evidence pointers, and applies deterministic policy. Every assessment remains an internal draft until an authenticated operator records one approve, deny, or revise attestation. Even approval does not authorize an application, message, purchase, submission, or other external business action.

The production path is designed to fail closed. Its repository implementation includes a durable emergency-stop epoch, idempotency binding, per-principal and global rate and reserved-token limits, an isolated operator service, a separate read-only judge service, non-generative dependency readiness checks, an exact container file allowlist, exact-version Secret Manager references through the Omega-X and Ultra-Secrets chain, and create-only Cloud Storage evidence with object-generation receipts.

How we built it

The application uses Python, FastAPI, Pydantic, the Google Gen AI SDK for Vertex AI, PostgreSQL for durable runtime controls, and Cloud Storage for evidence. The Gemini response contract requires a bounded recommendation, rationale, risks, assumptions, proposed scope, disclosure text, and pointers to supplied input fields. Deterministic code then evaluates AI permission, disclosure, grounding pointers, risk, and human-review requirements.

The repository also contains a default-deny Canadian staging authorization validator, an allowlisted non-root container build, a plan-first deployment helper that resolves and deploys an immutable image digest with no traffic, and a permit-bound canary helper for exactly one synthetic call with no automatic retry. These helpers do not prove that deployment or a live model call occurred.

Responsible AI and safety

AIGC treats model output as an assessment of operator-supplied material, not as verified fact. It records assumptions and evidence pointers, discloses AI use, requires accountable human review, and never grants the model authority to act externally. Shared-credential possession is not described as verified human identity. Readiness checks use only configuration, metadata, storage, and control-store probes; they do not generate model output.

If final evidence persistence or the durable completion marker fails, the result is not exposed as a completed execution. A network-unknown canary outcome is recorded as non-retriable rather than silently spending a second model call.

Challenges

The hardest part was making operational claims narrower than implementation claims. A passing unit test can show how an emergency stop, idempotency rule, or generation-bound receipt behaves in a controlled test; it cannot show that a specific cloud environment is configured correctly or that a live Gemini call occurred. The workflow therefore treats CI, staging authorization, immutable deployment, live execution, judge publication, and final submission as separate receipt-gated events.

Accomplishments

  • A strict Gemini assessment schema with deterministic, fail-closed policy.
  • Durable stop, idempotency, rate, and reserved-token control contracts.
  • Separate operator and provider-free, read-only judge surfaces.
  • Generation-bound evidence and content-bound judge-publication contracts.
  • A default-deny Canadian staging and one-call proof sequence.
  • An explicit distinction between repository evidence and runtime evidence.

What we learned

Governance is most useful when it changes runtime state transitions. It is not enough to describe a human gate or an audit trail: the system must prevent a provider call when control state is unavailable, quarantine evidence that is not bound to durable completion, and refuse public claims that exceed their receipts.

What's next

Restore GitHub-hosted runner admission and obtain a passing receipt for the exact release SHA; review and authorize the Canadian staging configuration; deploy the allowlisted image by digest; execute one synthetic, permit-bound Gemini assessment; publish only a sanitized judge receipt; capture a redacted walkthrough; and merge and submit only after those receipts pass review.

Current status disclosure

This submission text describes repository implementation. It does not, by itself, claim a live Gemini call, production availability, cloud-budget enforcement, verified human identity, public judge access, PR merge, or Devpost submission.

Built With

Share this project:

Updates