SGA 14001 — Environmental Management

Environmental management, connected: planning, legal review, operations, evidence, human decisions, and bounded GPT-5.6 drafts.

Comment

Inspiration

Environmental management is often fragmented across planning, legal review, operations, waste programs, training, indicators, evidence, and corrective action. Teams lose traceability when those activities live in disconnected spreadsheets and documents.

What it does

SGA 14001 is a guided environmental management platform that connects the working system from context and planning through execution, measurement, and human review. It includes:

  • environmental aspects and impacts;
  • obligations and legal-review workflows;
  • operational controls and PGIRASA waste management;
  • objectives, programs, training, indicators, evidence, and findings;
  • dashboards that surface gaps without claiming certification or legal compliance.

The current product interface is Spanish because it is designed for real Latin American environmental teams. The judge video uses English narration, captions, and explanatory overlays.

Build Week scope

The SGA/PGIRASA foundation existed before Build Week. During Build Week, Codex helped us inspect and strengthen the full workflow, and we built TraceGuard as the new GPT-5.6 extension. TraceGuard turns an authorized environmental incident note into cited, typed draft proposals for a human reviewer.

GPT-5.6 is constrained by Structured Outputs. A deterministic compiler then verifies numbered-source citations, allowed record types, domain fields, unsupported inferences, contradictions, and human-only decisions. The model cannot approve, implement, close, certify, or determine compliance.

How we built it

The platform uses FastAPI, Pydantic, SQLite for the local demonstration, HTML/CSS/JavaScript, and pytest. TraceGuard uses the OpenAI Responses API with GPT-5.6 and a strict typed extraction contract. The live model path fails closed unless it is explicitly enabled and an API key is present. Each request requires fresh confirmation that the selected source may be sent to OpenAI and may incur cost.

How we used Codex

Codex helped map the pre-existing architecture, challenge unsafe assumptions, implement the bounded GPT-5.6 workflow, expand the modular SGA experience, build deterministic validation, generate synthetic judge data, run the test suite, and prepare the evidence package and demonstration.

Challenges

  • Connecting a broad environmental management system without overstating ISO conformity.
  • Preserving uncertainty and contradictions instead of forcing false certainty.
  • Keeping AI output draft-only and reviewable.
  • Separating pre-existing platform capability from the new Build Week extension.
  • Demonstrating a Spanish-language product clearly to English-speaking judges.

Accomplishments

  • A connected Plan–Do–Check–Act environmental workflow.
  • Line-level provenance and visible TraceGuard diagnostics.
  • Fail-closed model access and no silent fallback.
  • Reuse of a typed SGA domain contract instead of model-invented fields.
  • A synthetic, reproducible judge path with no real environmental or personal data.
  • Frozen local validation: 648 tests passed, 9 PostgreSQL integration tests skipped because PGIRASA_TEST_DATABASE_URL was not configured, and 0 failures.

What we learned

Useful AI in evidence-heavy environmental workflows needs four separate layers: typed extraction, deterministic verification, visible uncertainty, and accountable human decision. Structured output enables automation; provenance and authority limits make it trustworthy enough to review.

What's next

  • Validate PostgreSQL/RLS with an explicit test database.
  • Evaluate a larger labeled synthetic corpus.
  • Add reviewer feedback capture without treating corrections as automatic truth.
  • Design an authenticated deployment separately; the current build remains loopback-only.
  • Complete privacy, threat-model, and professional environmental review before any real-data pilot.

Scope boundary

SGA 14001 does not certify ISO 14001 conformity, determine legal compliance, or replace professional environmental judgment. Human review remains authoritative.

Built With

Share this project:

Updates