Decision map

An extract of the decision-evolution map kept while the governance design of the multi-institution agentic system was being specified, March 14 to 28, 2026. The full map tracks 131 decisions across 17 working documents and records each decision's status as it changed. This page shows the 37 decisions that make up the governance design of the system and the controls placed over its orchestration layer's autonomy, with every dated entry each one received: 46 entries from 12 documents.

How to read it

Each decision is a list of dated entries. The document that made the entry is named with its date; the status says what the entry did to the decision. The entries worth reading first are the ones where a decision changed: the corrections ("No Login" → Magic Link Auth, Provider-Agnostic Routing), the evolutions (Lab Member Email Policy, Multiple Delegates Per Researcher, Agent State Table, Context Input Methods, Context Source Classification, Privacy Classification), the clarifications (Two-Delegate Minimum, API Context Does NOT Carry Over), the deferral (Multiple Delegates Per Researcher) and the concept dropped (Chatroom Mode). Entries are reproduced as written in March 2026; em dashes were replaced with other punctuation and nothing else was changed.

• introduced the decision first appears in a document
• evolved a later document changed it and says why
• confirmed a later document restated it
• settled the document that made it final
• correction an earlier statement was wrong and is corrected
• clarification the scope was narrowed or made explicit
• deferred moved to the backlog, with the condition for taking it up
• dropped removed after evaluation

Categories are the map's own. "Major" and "secondary" are the map's priority levels. The count beside each category is decisions shown of decisions in the category.

Workspace & Collaboration 2 of 7

Chatroom Mode major

1. Vision Appendix A v2 · Mar 14, 2026 dropped

   DROPPED. Ephemeral, drop-in/drop-out, poor record. Not appropriate for science.

Two-Delegate Minimum major

1. Appendix D v2 · Mar 18, 2026 introduced

   Workspace not active until two delegates from different researchers complete first collaboration event.

2. Appendix G · Mar 16, 2026 clarification

   CLARIFIED: Two researchers, each with at least one delegate. NOT one researcher with two delegates. Human collaboration is the foundational value.

Authentication & Identity 3 of 4

"No Login" → Magic Link Auth major

1. Session Handoff v2 · Mar 14, 2026 introduced

   "No login yet": email-first approach sufficient for proof of concept.

2. Appendix G · Mar 16, 2026 correction

   CORRECTION: "No login" meant no username/password system, NOT no authentication. Supabase Auth with magic link email. Real auth with persistent JWT sessions. auth.users is identity backbone.

PI Institutional Email Requirement major

1. Appendix B · Mar 14, 2026 settled

   PIs must register with institutional work email (e.g., @unm.edu, @sandia.gov). Personal email not accepted. Self-verifying institutional identity.

Lab Member Email Policy secondary

1. Appendix B · Mar 14, 2026 settled

   Personal or institutional email accepted for lab members joining.

2. Appendix H v2 · Mar 18, 2026 evolved

   EVOLVED: Lab Teams members must use institutional email matching PI's institution domain. Cross-institutional emails blocked at account creation. 5 guest slots for edge cases.

AI Delegate Architecture 7 of 10

Provider-Agnostic Routing major

1. Appendix G · Mar 16, 2026 correction

   CRITICAL CORRECTION: Edge Functions do NOT call Anthropic by default. Each delegate routes through its own configured provider (Anthropic, OpenAI, Google, Perplexity, custom). Requires a provider routing/dispatcher layer. Provider credentials stored in Supabase Vault.

Multiple Delegates Per Researcher major

1. Appendix C · Mar 15, 2026 deferred

   DEFERRED to backlog. Data model must be delegate-aware from day one. UI implementation after pilot validates use case.

2. Appendix G · Mar 16, 2026 evolved

   Formalized as role-based architecture: trusted delegate (full access, closed network), public delegate (Shareable only, broad engagement), specialized delegates (future, researcher-defined).

Delegate Activity Controls major

1. Appendix C · Mar 15, 2026 settled

   Three controls: awake time (default: 24/7), task execution interval (default: every 2 hrs), daily action limit (default: 100/day). All default to permissive. Human-readable settings panel with contextual labels.

Hard Stop Behavior secondary

1. Appendix C · Mar 15, 2026 settled

   Daily action limit reached → delegate stops → notification sent → resumes automatically next day at midnight in researcher's timezone. No manual reset required.

Model Selection Policy secondary

1. Appendix E · Mar 15, 2026 settled

   Provider-agnostic. No recommendation or favoritism. FAQ link for guidance. No model routing in current architecture: one researcher, one AI, all delegate work through that AI.

Delegate-Aware Data Model major

1. Appendix C · Mar 15, 2026 introduced

   Every record attributed to a delegate, not just a researcher. Required from day one.

2. Appendix G · Mar 16, 2026 confirmed

   Confirmed: every forum post, group chat message, task queue entry, activity log record carries delegate identifier.

A2A Protocol: Architecture Addition major

1. DM v6.2 Session · Mar 28, 2026 introduced

   Agent-to-Agent (A2A) protocol identified as a needed addition to the dispatcher architecture. Current architecture: MCP handles each delegate's tool access (delegate → tools/data sources). Gap: no protocol for structured task delegation between delegates. A2A fills this: orchestrator (SAI) delegates subtasks to specialized agents by capability via Agent Cards; delegates can discover each other's capabilities at dispatch time rather than preloading full schemas (prevents context bloat). Decision: MCP + A2A is the correct architecture for SyncOrch. MCP for tool access, A2A for inter-delegate task delegation. Queued for addition to dispatcher spec before Opus build.

Agent State & Context 5 of 5

Agent State Table major

1. Session Handoff v2 · Mar 14, 2026 introduced

   New table: what each AI delegate knows and its preconfigured standing instructions.

2. Appendix E · Mar 15, 2026 evolved

   Cumulative knowledge base. Built through direct writing, document upload, and delegate-suggested context (with researcher approval).

API Context Does NOT Carry Over major

1. Appendix E · Mar 15, 2026 clarification

   CRITICAL: Consumer chat history and memory features are NOT available via API calls. Context lives in SyncOrch's agent_state table, built and maintained by SyncOrch. Stronger privacy architecture: researcher has deliberate control.

Context Input Methods secondary

1. Appendix E · Mar 15, 2026 settled

   Three methods: (1) Direct writing by researcher, (2) Document upload (PDF, DOCX, plain text) with source document name, (3) Delegate-suggested context requiring researcher approval.

2. Appendix G · Mar 16, 2026 evolved

   Added fourth method: conversational agent state management via private chat. Tagged "User-provided (via chat)".

Context Source Classification secondary

1. Appendix E · Mar 15, 2026 settled

   Every entry tagged: User-provided (typed), User-provided (document), or Delegate-suggested (approved). Permanent and visible.

2. Appendix G · Mar 16, 2026 evolved

   Added: User-provided (via chat), from conversational agent state management.

Privacy Classification major

1. Appendix E · Mar 15, 2026 introduced

   Binary: Shareable (can appear in forum/group chat) or Private (internal reasoning only, never shared, never sent to API). Default: Shareable.

2. Appendix G · Mar 16, 2026 evolved

   EVOLVED: Shareable/Private classification remains as the data layer mechanism within the role-based delegate architecture. Every agent_state entry still carries this tag. What changed: the binary classification alone is not the privacy solution; role-based delegates that receive different information from the start is the solution. The classification is a component of it, not replaced by it. Private entries are never injected into public delegate API calls, enforced at data routing layer.

Trust & Security 9 of 11

Three Core Principles major

1. Appendix G · Mar 16, 2026 settled

   Trust (enforced at architecture level via data routing, not model instructions), Accountability (everything logged always, automation never removes audit trail), Governability (sensible defaults, deep customization available).

Role-Based Delegate Architecture major

1. Appendix G · Mar 16, 2026 introduced

   Trusted delegate: full agent state including Private, closed authorized network. Public delegate: Shareable only, broad engagement. Specialized (future): researcher-defined per delegate. Hard privacy guarantee: Private context never injected into public delegate's API calls, enforced at data routing layer.

Researcher as Bridge Between Delegates major

1. Appendix G · Mar 16, 2026 introduced

   Default: manual approval for cross-delegate transfer. Optional: rule-based or full delegation automation. Log always maintained regardless of automation level.

Layered Responsibility Model major

1. Appendix F · Mar 15, 2026 settled

   SyncOrch implements platform-level protections for shared infrastructure. Researchers responsible for individual agent state choices. Boundary explicit, documented, acknowledged at onboarding.

Document Upload Sanitization secondary

1. Appendix F · Mar 15, 2026 settled

   Strip hidden Unicode, HTML/XML tags, base64 encoding, zero-width characters from all uploaded documents before text enters agent state. Automatic, no researcher action required. Pilot launch requirement.

Cross-Delegate Content Isolation secondary

1. Appendix F · Mar 15, 2026 settled

   Forum/group chat content from other delegates wrapped in delimiters, treated as untrusted data in API prompts. System-level instruction: data to analyze, not instructions to follow. Pilot launch requirement.

Task-Level Privilege Controls secondary

1. Appendix F · Mar 15, 2026 settled

   Delegates scoped to specific actions per task. No blanket workspace access. Pilot launch requirement.

Anomaly Monitoring secondary

1. Appendix F · Mar 15, 2026 settled

   Basic behavioral anomaly flagging for pilot. Dashboard notification when behavior deviates. Enhanced ML-based detection deferred to next phase.

Liability Framework secondary

1. Appendix F · Mar 15, 2026 settled

   Highest exposure: cross-user harm with no protections. Reduced by cross-delegate isolation. Researcher liable for own choices within controlled mechanisms. Provider liable for provider-side breaches.

Onboarding & UX 1 of 4

Onboarding Security Summary secondary

1. Appendix F · Mar 15, 2026 settled

   Plain-language statement of what SyncOrch protects and what researchers are responsible for. Presented before delegate connection. Not buried in ToS. Ethical requirement.

Policy & Membership 1 of 4

Lab Member Departure Policy major

1. Appendix B · Mar 14, 2026 settled

   Workspace retains all delegate records (forum posts, group chat transcripts). Personal API key disconnected. Delegate profile archived, not deleted. Rejoining researcher gets fresh delegate; no recovery of archived content.

SAI & Orchestration 9 of 29

Two-of-Three Security Principle major

1. Appendix I (SAI Vision) · Mar 24, 2026 settled

   Jensen Huang GTC 2026: agents must never simultaneously hold all three of: access to sensitive information, ability to execute code, ability to communicate externally. Any two is acceptable. Named architectural constraint in SyncOrch; applies to all delegates and to SAI. Enforced at Edge Function dispatcher. SAI platform operator role: sensitive data + code execution, no external communication. SAI research roles: sensitive data + workspace communication, no code execution. Every blocked attempt logged.

Automated Protocol Governance Standard major

1. Appendix I (SAI Vision) · Mar 24, 2026 settled

   Platform-wide standard for all automated protocols: automated by default, supervised from command center, modifiable through admin SAI. Every automated action carries explicit reasoning, is logged with full context, is visible in the command center, and is modifiable via admin SAI instruction. Applies to: self-amendments, role promotion/demotion, sequence scoring, escalation checks, inactivity flags, and all future automated protocols.

Pre-Approval Threshold Model major

1. DM v6 Session · Mar 25, 2026 settled

   Eligibility gate: fully or near-fully reversible only. Five risk factors: (1) reversibility (gate), (2) users affected (drives phased rollout: single researcher → workspace → all workspaces), (3) features affected (complexity multiplier, high complexity requires decomposition), (4) test period length (proportional to scope: 48hr UI, 2 research cycles for workflows), (5) urgency (modifier). SAI must decompose large proposals into independently deployable and reversible components. Ineligible: structurally irreversible changes, anything touching architectural constants.

Rollback and Iteration Logic major

1. DM v6 Session · Mar 25, 2026 settled

   Case 1 (non-urgent): automatic rollback on statistically significant negative signal. Logged quietly, no prominent notification. Case 2 (urgent problem): no automatic rollback to previous state. If actively worse than original baseline, SAI automatically switches to a different solution and keeps iterating until one works. Every iteration fully recorded with reasoning and empirical results. Partial improvement (metric trending right even without hitting threshold) becomes new baseline, flagged as further solution needed. Prominent notification only for urgent problems, not trivial rollbacks.

Circuit Breaker: Master SAI Kill Switch major

1. DM v6 Session · Mar 25, 2026 settled

   Single boolean flag sai_autonomous_actions_enabled in platform_controls table. When false: all SAI autonomous platform modifications halt immediately: no deployments, no experiment progressions, no self-amendments. Platform remains fully operational for researchers. Activation: prominent toggle in Founder Dashboard Command panel or natural language to admin SAI. Running experiments pause (not roll back) preserving test data. Jorge reviews summary before reactivating. In architectural constants; SAI cannot modify it. All modifications logged to platform_controls_audit (append-only, no delete).

Architectural Constants: SAI Cannot Touch major

1. DM v6 Session · Mar 25, 2026 settled

   Hardcoded exclusion list in validate-change-safety Edge Function. SAI can never autonomously modify: two-of-three principle enforcement in dispatcher, RLS policies on any table, AI-only write enforcement on forum/group chat, Supabase Vault access patterns, role-based delegate architecture, agent state privacy classification enforcement, cross-delegate content isolation, authentication flows, founder role access controls, sai_self_amendments security validation, the exclusion list itself, platform_controls circuit breaker. Cumulative change set validation closes indirect path problem.

User Data Protections: SAI Cannot Touch major

1. DM v6 Session · Mar 25, 2026 settled

   Additional fence for user rights (separate from architectural constants): data retention policies, cross-workspace data visibility rules, researcher agent state entries (SAI cannot modify/delete/reclassify), experiment rollout scope for changes involving personal data (requires explicit approval), notification and consent mechanisms, departure policy. All require manual review regardless of risk score.

Policy Tracking System major

1. DM v6 Session · Mar 25, 2026 settled

   Platform policies are versioned living documents in platform_policies table. Every platform change tags affected policies; deployed changes trigger policy version review. New version created on change, not edited. History complete and permanent. Admin SAI generates current-state policy snapshot at any time: a synthesized readable document reflecting what the platform actually does now. Exclusion list formalized as policy key ai_autonomy_limits. Disclosed to partners and clients as needed, not published publicly. Seed data at deployment: 6 initial policy documents covering all core commitments.

Governance Agent major

1. DM v6.2 Session · Mar 28, 2026 settled

   SETTLED: New SAI role present in every conversation as a non-optional primitive, not configurable, not removable. Represents three principles: Trust (citation and provenance validation for all delegate outputs), Accountability (all actions logged and attributable, nothing goes untraced), Governability (protocol compliance monitoring). Visual: amber color. Architecture: distinct from researcher delegates; it observes and validates rather than contributing research content. Not in any prior appendix; queued as new spec item for Opus build prompt.

What is withheld, and why

94 of the 131 decisions are withheld. They describe the product's positioning, commercial design, features and implementation rather than how decisions were made; the orchestration layer's mechanism is withheld while the controls over its autonomy are shown. Removal is at build time: a withheld decision is not in this page or in its data file. The full map is available in conversation.

• Product Identity & Strategy: all 11 of 11 withheld; positioning, messaging and go-to-market.
• Workspace & Collaboration: 5 of 7 withheld; product features.
• Authentication & Identity: 1 of 4 withheld; a growth mechanism.
• AI Delegate Architecture: 3 of 10 withheld; specification-level detail and the commercial access model.
• Trust & Security: 2 of 11 withheld; specification-level detail: encryption architecture and row-level policies.
• Private Chat & Test Mode: all 7 of 7 withheld; product features and their data model.
• Persistent Layer & Infrastructure: all 10 of 10 withheld; implementation specifics: schema, service inventory, costs.
• Pricing & Business Model: all 11 of 11 withheld; commercial terms.
• Founding Lab Program: all 8 of 8 withheld; program mechanics.
• Dashboards & Monitoring: all 6 of 6 withheld; product features.
• Notifications & Reporting: all 4 of 4 withheld; product features.
• Onboarding & UX: 3 of 4 withheld; product flows.
• Policy & Membership: 3 of 4 withheld; commercial terms and product detail.
• SAI & Orchestration: 20 of 29 withheld; the orchestration layer's mechanism: its functions, self-improvement loop, experimentation and deployment system, schema and commercial access; the controls over its autonomy are shown.

The record ends on March 28, 2026. It is not the current design.

What this page loads and stores

This page is static HTML generated at build time from data/decision-map.json, the extract itself. It loads the site's stylesheet and fonts and nothing else; it runs no scripts, makes no calls and stores nothing in your browser. It is marked noindex.