In regulated programs, decisions fail after they are made — not because teams lacked expertise, but because decision logic could not be reconstructed under scrutiny.
RGDS is the reference implementation for decision-layer governance in phase-gated regulated environments. It treats the decision itself as the primary artifact — structured, schema-validated, and written before memory decay and handoff loss make reconstruction impossible.
FDA Complete Response Letters cite "insufficient information" in 50% of first-cycle submissions. In most of those cases, the underlying science is defensible. What organizations cannot produce is a coherent account of why specific decisions were made 6–18 months earlier.
Traditional documentation model: RGDS model:
Documents → Analysis → Meeting Decision Question
↓ ↓
Implicit decision Options Considered (≥2)
↓ ↓
Memory + email threads Evidence Base + Completeness
↓ ↓
Reconstruction attempt Risk Posture + Residual Risk
(2–3 weeks, FDA pressure) ↓
Named Human Accountability
↓
Schema Validation → Git
↓
2-minute retrieval
The decision log is the record. Everything else — analyses, documents, source reports — serves the decision.
┌─────────────────────────────────────────────────────────────────┐
│ RGDS OPERATING MODEL │
├─────────────────────────────────────────────────────────────────┤
│ │
│ Phase Gate Event │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────┐ │
│ │ Decision Log Record │ ← Primary Artifact │
│ │ │ │
│ │ decision_question │ │
│ │ options_considered (≥2 required) │ │
│ │ evidence_base (completeness) │ │
│ │ risk_posture (explicit) │ │
│ │ residual_risk (structured) │ │
│ │ outcome (5 types) │ │
│ │ accountability (named humans) │ │
│ │ ai_assistance (if used) │ │
│ └─────────────────┬───────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────┐ │
│ │ Schema Validation (CI/CD) │ ← Enforcement │
│ │ decision-log.schema.json │ │
│ │ Semantic invariant checks │ │
│ │ Required fields enforced │ │
│ └─────────────────┬───────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────┐ │
│ │ Git (Version-Controlled Log) │ ← Audit Trail │
│ │ Immutable timestamps │ │
│ │ 2-minute retrieval under FDA query │ │
│ │ Reconstruction without interviews │ │
│ └─────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘
Every RGDS decision resolves to one of five governed outcomes. Each carries equal structural rigor — stopping is governed the same way proceeding is.
┌──────────────────┐
│ Phase Gate │
│ Decision Point │
└────────┬─────────┘
│
┌──────────────────┼──────────────────┐
│ │ │
▼ ▼ ▼
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ GO │ │ CONDITIONAL│ │ DEFER │
│ │ │ GO │ │ │
│ Full conf. │ │ │ │ Insufficient│
│ Complete │ │ Explicit │ │ evidence. │
│ evidence. │ │ conditions. │ │ Re-entry │
│ Proceed. │ │ Named owner.│ │ criteria │
│ │ │ Deadline. │ │ defined. │
└─────────────┘ └─────────────┘ └─────────────┘
┌──────────────────────────────────┐
│ │
▼ ▼
┌─────────────┐ ┌─────────────┐
│ NO-GO │ │ ESCALATE │
│ │ │ │
│ Rejected. │ │ Exceeds │
│ Rationale │ │ authority │
│ documented. │ │ scope. │
│ Re-entry │ │ Senior human│
│ logic set. │ │ required. │
└─────────────┘ └─────────────┘
Defer and No-Go are first-class outcomes. A governed No-Go is indistinguishable from a governed Go in audit quality. An ungoverned stop is indistinguishable from a failure.
The schema enforces governance-completeness at the artifact level. A decision that cannot satisfy the schema cannot be finalized.
decision_log/
├── decision-log.schema.json ← Machine-enforced
├── decision-log.schema.yaml ← Human-readable companion
└── decision-log.template.yaml ← Starting point for new decisions
| Field | Type | Governance Purpose |
|---|---|---|
decision_question |
string | Defines what choice was required |
decision_deadline |
date | Forces temporal accountability |
options_considered |
array (≥2) | Eliminates single-option rationalization |
evidence_base |
object | Links claims to sources |
evidence.completeness_state |
enum | complete / partial / placeholder |
risk_posture |
string | Forces explicit risk acceptance |
risk_assessment.residual_risk_items |
array | Captures what remains true after proceeding |
outcome |
enum | One of five governed decision types |
accountability.decision_owner |
string | Named human, not a role or team |
accountability.approvers |
array | Named humans with authority scope |
ai_assistance.used |
boolean | Disclosure trigger |
Evidence Item
│
├── complete → Final validated data. Decision proceeds on confirmed evidence.
│
├── partial → Preliminary data. Final report pending. Explicit condition
│ required if proceeding. Owner and deadline named.
│
└── placeholder → Estimated or assumed. Must be flagged as governance gap.
False confidence from undocumented placeholders is the
failure mode this field prevents.
Start here. Six examples demonstrate the full operating model across every decision outcome.
| Record | Outcome | Scenario | What It Demonstrates |
|---|---|---|---|
rgds-dec-0001 |
conditional_go | Data readiness gate | Explicit conditions, owned follow-ups, named approvers |
rgds-dec-0002 |
no_go | Risk threshold exceeded | Defensible refusal with re-entry logic |
rgds-dec-0003 |
defer | Missing required evidence | Structured pause with re-review criteria |
rgds-dec-0004 |
escalate | Pre-IND FDA interaction | Agency-facing decision framing and strategy |
rgds-dec-0005 |
conditional_go | IND authoring gate | Author-at-risk drafting, reviewer triage, lock points |
rgds-dec-0006 |
conditional_go | AI-assisted decision | Bounded AI disclosure, preserved human authority |
All six records are schema-validated in CI/CD on every commit. Reading one takes 5 minutes. Reading all six takes 30.
RGDS is valid with no AI at all. When AI is used, it operates as bounded assistance only.
┌──────────────────────────────────────────────────────────────┐
│ AI GOVERNANCE BOUNDARY │
├─────────────────────────┬────────────────────────────────────┤
│ PERMITTED (bounded) │ PROHIBITED │
├─────────────────────────┼────────────────────────────────────┤
│ Summarization │ Gate outcome decisions │
│ Field extraction │ Evidence of record by default │
│ Cross-document diffing │ Silent scope or risk acceptance │
│ Structured drafting │ Publishing or submission actions │
│ Schema completeness │ Fabricated citations or rationale │
│ checks │ Autonomous execution of any kind │
└─────────────────────────┴────────────────────────────────────┘
│
▼ When AI is used, these fields are required:
┌──────────────────────────────────────────────────────────────┐
│ ai_assistance.used → true │
│ ai_assistance.tool_name → which system │
│ ai_assistance.tool_purpose → what task │
│ ai_assistance.human_review[] → review tier + findings │
│ ai_assistance.human_override_log[] → corrections made │
│ ai_assistance.ai_risk_assessment → confidence band+cautions │
└──────────────────────────────────────────────────────────────┘
│
▼ Human decision owner remains fully responsible.
AI disclosure transfers no authority, approval
rights, or risk ownership.
Evidence rule: AI output is never treated as primary evidence. If an AI output influences a decision, the human owner must link to the underlying source and record the AI output as a drafting aid. Every decision must remain defensible without the AI output present.
Authoritative AI governance covenants: rgds-ai-governance
RGDS formalizes failure modes observed during IND preparation. Each mechanism addresses a specific, named execution pattern.
| Execution Reality | Failure Mode Prevented | RGDS Mechanism |
|---|---|---|
| Placeholders proceed without governance | False confidence, FDA gap finding | evidence.completeness_state + author-at-risk constraints |
| Scope changes emerge late without a trail | Silent ripple effects across modules | scope_change_events[] + downstream propagation |
| Reviewer routing is informal | Unclear accountability under audit | review_plan + named triage owner |
| Risk posture is implied, not stated | Cannot defend tolerance decisions to FDA | risk_posture + residual_risk_items |
| Cross-module dependencies are mentally tracked | Late-discovered misalignment after gate closes | dependency_map[] |
| Phase-gate tolerance is assumed shared | Silent misalignment between functions | Explicit risk posture field, cross-functional sign-off |
| Regulatory interaction strategy is informal | Weak pre-IND positioning | decision_category: regulatory_interaction |
| AI assistance is undisclosed | Provenance contamination, audit exposure | ai_assistance disclosure schema (mandatory when used) |
rgds/
│
├── decision-log/ ← Schema and templates
│ ├── decision-log.schema.json ← Machine-enforced schema
│ ├── decision-log.schema.yaml ← Human-readable version
│ └── decision-log.template.yaml ← Starting template
│
├── examples/ ← Start here
│ ├── README.md ← How to read examples
│ ├── rgds-dec-0001.json ← Conditional go (canonical)
│ ├── rgds-dec-0002-no-go.json ← No-go (canonical)
│ ├── rgds-dec-0003-defer-*.json ← Defer (canonical)
│ ├── rgds-dec-0004-regulatory-*.json ← Escalation (canonical)
│ ├── rgds-dec-0005-ind-*.json ← IND conditional go
│ └── rgds-dec-0006-ai-*.json ← AI-assisted (only AI example)
│
├── evaluation/ ← Decision quality assessment
│ ├── evaluation-plan.md ← Assessment methodology
│ ├── evidence-quality-rubric.md ← Evidence scoring criteria
│ ├── requirements-traceability-matrix.md ← 100% RTM coverage
│ └── scorecard-template.csv ← Structured review scorecard
│
├── docs/ ← Governance documentation
│ ├── why-rgds-exists.md ← Evidence-to-design rationale
│ ├── decision-log.md ← How to read decision logs
│ ├── governance.md ← Rules and enforcement intent
│ ├── ai-assistance-policy.md ← AI governance policy
│ ├── role-decision-artifact-matrix.md ← Cross-role ownership
│ └── change-control-log.md ← Schema change history
│
├── scripts/ ← Validation tooling
│ ├── validate_decision_log.py ← Single-record validator
│ └── validate_all_examples.py ← Batch validator (CI)
│
├── .github/workflows/
│ └── validate.yml ← CI/CD schema + semantic validation
│
├── Makefile ← Local validation commands
└── requirements.txt
Different readers have different entry points.
Are you a...
│
├── Executive / Approver
│ └── README.md → rgds-dec-0001 or rgds-dec-0005
│ Goal: understand what a governed decision looks like
│
├── Quality / Governance Reviewer
│ └── docs/governance.md → docs/decision-log.md → evaluation/
│ Goal: understand review criteria and audit artifacts
│
├── AI Governance Reviewer
│ └── docs/ai-assistance-policy.md → rgds-dec-0006
│ → github.com/mj3b/rgds-ai-governance
│ Goal: understand AI boundaries and disclosure requirements
│
├── Regulatory / FDA Auditor
│ └── examples/ + evaluation/requirements-traceability-matrix.md
│ Goal: reconstruct decision context from governed records
│
└── Technical Implementer
└── decision-log/decision-log.schema.json → scripts/
→ .github/workflows/validate.yml
Goal: understand schema enforcement and CI integration
v2.0.0 tightens decision defensibility. It does not add automation or autonomy.
| Change | What it enforces | Failure mode prevented |
|---|---|---|
| Options enumeration (≥2 required) | At least two options must be considered | Single-option rationalization passing as governance |
| Evidence completeness per item | complete / partial / placeholder on every evidence item |
False confidence from undocumented placeholders |
| Structured residual risk | residual_risk_items[] array |
Risk acceptance without recording what remains true |
| Named human accountability | Decision owner + approvers as individuals, not roles | "Who approved this?" questions with no traceable answer |
| AI assistance disclosure | Required schema fields when ai_assistance.used=true |
AI-assisted drafting without disclosure contaminating provenance |
Decision quality is assessed across four dimensions.
| Dimension | What is evaluated | Instrument |
|---|---|---|
| Decision readiness | Evidence completeness, option coverage, risk explicitness | Evidence quality rubric |
| Governance execution | Accountability chain, approval separation, escalation logic | Reviewer audit checklist |
| AI assistance safety | Disclosure completeness, human override documentation | AI governance policy + dec-0006 |
| Requirements coverage | End-to-end traceability from program objectives to decisions | Requirements traceability matrix (100% coverage) |
Evaluation focuses on decision quality and governance execution. It does not benchmark model performance in isolation.
RGDS is the biopharma reference implementation. GDI (Governed Decision Intelligence) generalizes the decision-layer architecture to domain-agnostic deployment.
RGDS (this repository) GDI
Biopharma / IND / BLA context → Domain-agnostic open specification
Phase-gate decision logs → Governed Decision Records (GDR)
IND-specific field vocabulary → Universal schema
FDA reconstructability focus → NIST AI RMF / ISO 42001 / EU AI Act
170+ commits, 6 canonical → Reference implementation +
examples, CI enforcement IETF conformance driver
| Repository | Purpose | DOI |
|---|---|---|
| mj3b/rgds | Biopharma reference implementation (this repo) | 10.5281/zenodo.20242004 |
| mj3b/rgds-independent-study | Ten-question independent study | 10.5281/zenodo.20242004 |
| mj3b/governed-decision-intelligence | GDI v3.0 open specification | 10.5281/zenodo.20244601 |
| mj3b/rgds-ai-governance | AI governance covenants | — |
v2.0.0 — Biopharma reference implementation of the GDI v3.0 open specification.
RGDS implements the decision-layer governance architecture defined in GDI v3: The Decision Architecture for Governed AI (DOI: 10.5281/zenodo.20244601) for the biopharma/IND context specifically.
- Schema-enforced decision logs with mandatory options analysis, evidence completeness, and residual risk
- Six canonical decision records spanning all five outcomes
- Bounded, disclosed AI assistance (non-agentic by design)
- CI/CD validation of schema and semantic invariants on every commit
- 100% requirements traceability matrix coverage
- Independent case study — not a production system, not regulatory advice
@software{banasihan2026rgds,
author = {Banasihan, Mark Julius},
title = {{RGDS}: Regulated Gate Decision Support},
year = {2026},
version = {2.0.0},
doi = {10.5281/zenodo.20242004},
url = {https://doi.org/10.5281/zenodo.20242004},
license = {Apache-2.0}
}Mark Julius Banasihan Decision governance systems for regulated, high-stakes environments.