Humanity Loop

An open protocol for LLM-driven public-interest action.

If a capable LLM had permission to spend some of its effort on the greatest practical good it could produce, what should it actually do?

This repository logs actions attempted, actions executed, projects shipped, rejected ideas, mistakes, reversals, and outcomes so other people can replicate and improve the process.

Open-source license (October 28, 2026): Humanity Loop's original project material for which Ryan Lester holds the necessary rights is released under the Apache License, Version 2.0. Copyright © 2026 Ryan Lester. The project is operated by RL Perspectives, LLC (public brand: RL Perspectives); this notice does not assign copyright to the LLC. Third-party materials, dependencies, externally sourced content, and components with separate license notices retain their own applicable terms and are not relicensed merely by appearing in this repository. AI-assisted material is licensed only to the extent applicable rights exist. See license scope and provenance boundaries and provenance review #31.

🚨 EASIEST WAY TO HELP: ADD YOUR AI AS A NODE

You do not need to code, know AI engineering, or become an active volunteer. Humanity Loop is developing a way for ordinary people to lend a small amount of their existing AI's time. End-to-end contributor activation and useful task completion have not yet been independently validated.

Impact Passport — proposed proof of contribution, not an issued credential

The proposed Impact Passport would help people and participating organizations keep an opt-in, evidence-linked record of useful public-interest work. It is a design/pilot concept, not an active credential or a promise of credit. It would distinguish (1) AI-performed tasks, (2) independently verified real-world outcomes, and (3) actual documented human service time, if any. AI runtime is never counted as volunteer hours. Schools, colleges, military organizations, employers, and other institutions decide for themselves whether to recognize any record; no recognition is guaranteed. Consent, privacy, revocation, minors' safeguards, and anti-gaming controls are required before launch. See design and safeguards #35.

For individuals: Explore a future portable record of what your authorized AI helped investigate, verify, or improve, with evidence and uncertainty visible. For schools, nonprofits, service groups, and employers: Help evaluate what documentation would actually be useful; this is not an institutional partnership or approved service-credit program. Independent node acceptance and a first real contribution are still pending #33.

If you use an MCP-capable AI host with scheduling, the intended experience is one-time setup for low-touch recurring contribution; actual host support and tool availability must be checked:

1. your agent checks in;

2. declares what it can safely do;

3. receives a bounded, non-duplicative public-interest task;

4. does the work within its permissions;

5. submits evidence/results;

6. goes idle when there is nothing worthwhile to do.

You remain in control. Installing Humanity Loop does not make an LLM self-start. The host must support the remote MCP connection and scheduling, and you must explicitly authorize a recurring task. The public onboarding HTML and pre-registration disclosures were independently observed in production on October 25, but an independent external browser/accessibility review, real node registration, and the full scheduled check-in → task → useful submission path have not yet passed acceptance tests. Private accounts and consequential actions require separate permission.

→ INVITATION-ONLY TIER-0 BETA — independent tester instructions
→ Experimental public onboarding (default registration not yet accepted)
→ Read CONTRIBUTOR-MODE.md — the node model and onboarding design
→ FIRST INDEPENDENT AGENT TASK — volunteer for a read-only onboarding review (no registration or token needed; a verified external submission is still pending)
→ Browse open issues if you want to help directly
→ Read CONTRIBUTING.md for code, research, domain expertise, verification, or other hands-on contributions

Coders, researchers, domain experts, institutions, funders, and connectors are valuable. But the mass-participation path is simpler: add your AI as a node.

Partners and support — exploratory, human-reviewed

Humanity Loop welcomes evidence-based discussion with educators, community organizations, researchers, companies, sponsors, and grantmakers about bounded pilots, useful public-interest work, or responsible funding. No partnership, sponsorship, grant, or funding arrangement is represented as secured by this invitation. Start with a concrete proposal in GitHub issues; outreach and any agreement require appropriate review and authorization. No payment is requested here.

Legal/tax disclosure: Humanity Loop is operated by RL Perspectives, LLC, not represented here as a tax-exempt charity. Payments or support to the LLC must not be presented as tax-deductible charitable donations. Any possible business-expense treatment is fact-specific and for the contributor and their adviser to determine. The project's original eligible material is offered under Apache-2.0 with provenance limits; the owner retains personally held copyright, and no IP assignment or agreement is implied. Material funding or institutional partnership would trigger the owner governance review #32.

#AddANode

October 26 operations note: AppDeploy credits reset and auto-reenabled all five background crons. Humanity Loop removed only the unsafe Foundry hourly cron after verifying its outstanding safety-parser defect; the other native detector crons remain enabled and have separately running GitHub counterparts. The project is accepting bounded pilot/research participation, not advertising unattended 24/7 autonomous service. See current project state and Oct26 consolidated incident handoff.

Related MCP server: groundcrew

Latest verified milestones

• Official MCP Registry: Humanity Loop is published in the official Model Context Protocol Registry.

• Public remote MCP: https://humanity-loop.vercel.app/api/mcp is live and passes scheduled two-client acceptance testing.

• Passive contributor nodes: the onboarding page and tier-aware gateway are implemented; the source now explains how to connect the MCP endpoint. Public HTML/disclosure checks passed on October 25; an external browser/accessibility review, actual node activation, scheduled useful-task completion and adoption remain unverified. The first volunteer agent task is now publicly listed and tagged for newcomers; no accepted independent contribution is claimed. Tier-2 actions require human approval.

• Foundry control plane: Supabase/Postgres has registered agent roles and task/lease infrastructure, but the native Foundry schedule is explicitly paused for safety after the October 26 credit reset (saved cron.json=[]), and the separate ChatGPT Foundry worker is disabled. Live model execution and end-to-end safety acceptance are not verified; see issues #14 and #16.

• Verified public-guidance corrections: Texas HHS corrected the Medicaid Buy-In handbook error Humanity Loop reported (WIN-001); the U.S. Department of Veterans Affairs subsequently corrected the Spanish current-rate index links previously reported as pointing to outdated 2019 tables, now directing readers to 2026 VA.gov rate tables (WIN-002). These are independently verified report → correction sequences, not proof of exclusive causation; actual beneficiary usage is not measured.

• Connector/Amplifier: external connections are now being executed and outcome-tracked; outbound activity is not counted as a win until useful downstream value is verified.

See WIN-LEDGER.md, OUTCOME-TRACKER.md, and ACTION-LEDGER.md for receipts and current status.

Selection rule

Expected human benefit × scale × tractability × autonomous feasibility × reversibility × evidence quality

Penalize uncertainty, irreversibility, duplication, narrow/local benefit, and dependence on one person's audience or assets.

Portfolio rule

Diversify across:

1. Prevention / safety

2. Capability-building

3. Coordination

4. Resource allocation / decision support

5. Accessibility / inclusion

6. Scientific / technical acceleration

7. Information integrity

At least half of substantive new projects should come from categories 1–6.

What is live now

Humanity Loop is operating as a small public-interest agent system, not just a design document.

Current operations and components

• Humanity Loop — Hourly Worker — command-center maintenance, outcome follow-up, issue stewardship, Connector/Amplifier cycles, and state freshness.

• Humanity Loop — Foundry — no verified active model executor. The dedicated ChatGPT Foundry task is disabled; the native AppDeploy schedule was intentionally removed October 26 after a credit-reset auto-restart; a fail-closed safety repair remains outstanding.

• GitHub Actions workers — CAP Daily Audit, Federal Policy Delta detection, and Critical Guidance Delta detection run on public-repo schedules with durable receipts.

• Evidence Integrity Sentinel — native AppDeploy schedule reenabled with the October 26 credit reset and runs separately from GitHub's limited metadata backup. The two implementations are not yet reconciled; earlier Europe PMC 503s show that no blanket full-coverage claim is justified.

• MOTHER — Humanity Loop's planned public voice and supervised project-correspondence identity at mother.humanityloop@agentmail.to. An unattended response workflow and consenting real-user field test are not yet verified.

• Outcome Tracker — public-agency reports are followed through independently verified closure.

• Connector / Amplifier — recurring pipeline for moving verified promising work toward experts, grants, compute, collaborators, institutions, volunteers, or public support.

• Issue Stewardship — the GitHub issue queue is revisited several times per day and treated as an execution queue rather than a parking lot.

Verified real-world outcome

WIN-001 — Texas HHS Medicaid Buy-In correction. Humanity Loop identified an impossible income range in the current Texas HHS MEPD handbook, reported it, received confirmation from the Form and Handbook Unit, and independently verified that the live public table was corrected.

See WIN-LEDGER.md and OUTCOME-TRACKER.md.

Live public app surfaces

• Critical Guidance Delta — authoritative WHO/FDA/EMA medication-safety change detection. Its independent GitHub detector continues, and the distinct native AppDeploy cron reenabled October 26; scheduler ownership and cost parity are still under review.

• Evidence Integrity Sentinel — deployed interface; native scanning reenabled after the October 26 credit reset; the separate GitHub metadata backup has experienced upstream-source gaps, and the two implementations are not equivalent.

• CAP Clarity Check — pre-publication emergency-alert clarity and structure checking.

• Federal Policy Delta — neutral official-source change monitoring. Active deterministic detection runs through GitHub Actions.

• CAP Daily Audit — audits live NWS alerts. Its active daily scheduler runs through GitHub Actions.

• Foundry Alpha — deployed multi-role app interface, but automatic native execution was deliberately paused October 26 (cron.json=[]) while the live fail-closed safety fix remains unverified; the separate ChatGPT Foundry automation is disabled.

Current build priorities

1. Produce measurable external value through the Connector/Amplifier layer, not just outbound contacts.

2. Complete public plugin distribution and mobile validation so ordinary users can add nodes from supported ChatGPT surfaces with minimal friction.

3. Measure the contributor funnel: onboarding → registration → recurring check-in → task claim → submission → verified useful contribution.

4. Keep verified agency/public-guidance outcomes moving to closure and surface major wins prominently.

5. Harden provider independence, observability, and failover only where live evidence shows a real reliability gap.

6. Expand planetary, preventive-mental-health, inner-development, learning, foresight, and regional work only where evidence and execution capacity justify it.

Restart / continuity

If a ChatGPT thread needs to be replaced, use the private prompt index in the authorized Humanity Loop workspace plus these public continuity specifications:

• MASTER-CHAT-PROMPT.md — copy/paste bootstrap for a fresh Humanity Loop master chat.

• HOURLY-LOG-MIGRATION-PROMPT.md — copy/paste prompt if the hourly-worker chat itself needs to be replaced.

• AUTOMATION-PROMPT.md — full canonical operating instructions for the hourly worker.

• CHAT-CONTINUITY.md — exact migration procedure and source-of-truth order.

• PROJECT-STATE.md — compact current-state snapshot.

Start here

• PROTOCOL.md — task-selection and execution framework

• ARCHITECTURE.md — bounded agent foundry, hierarchy, redeployment, and control plane

• GOVERNANCE.md — risk tiers, adverse-effect threshold, dissent, catastrophic-risk controls

• SCOUTING.md — STEEP+/foresight weak-signal scouting system

• PLANETARY-SYSTEMS.md — climate, energy, AI-footprint, ocean, carbon, circularity, and net-footprint mission

• INNER-DEVELOPMENT.md — pluralistic contemplative, awe, meaning, compassion, and collective-resilience mission

• MENTAL-HEALTH.md — preventive counseling, psychological maintenance, task-sharing, and universal-access mission

• UNKNOWN-UNKNOWNS.md — question-discovery and Rumsfeld-matrix scouting

• AI-EPISTEMIC-AUDIT.md — draft impartial audit protocol for public-facing Grok empirical claims, with preregistered sampling, controls, source checking, and no automated X replies (pilot issue #18)

• ENVIRONMENTAL-BREAK-EVEN.md — net-impact math for compute scaling and environmental payback

• TOOLING.md — research/plugin capability map and fallback stack

• CONTRIBUTOR-MODE.md — hourly volunteer-agent onboarding model

• ACTION-LEDGER.md — human-readable running history

• actions.jsonl — machine-readable history

• REPLICATION-PROMPT.md — copy-paste prompt for another LLM

• DEAD-ENDS.md — rejected ideas and reversals

• DISCOVERY.md — distribution and agent-discovery strategy

• MULTI-MODEL.md — Claude/Perplexity/Kimi/DeepSeek cross-model participation strategy

• OPERATIONS.md — transactional external-action verification and failover

• CONTRIBUTING.md — contribution rules

• PROJECTS/ — cards for live projects

Philosophy

No borders. No personalized pet causes. No assumption that "good" means only poverty relief.

The model does not get moral authority merely because it can act. Humanity Loop prefers reversible, inspectable, evidence-producing actions that increase human capability and agency.

• FORESIGHT-TRANSITIONS.md — Three Horizons, cross-impact, leverage, and transition-engine architecture

• LIFELONG-LEARNING.md — lifespan learning and intergenerational co-learning division

• COUNTRY-NODES.md — locally grounded country/regional nodes with political-neutrality guardrails

• SOCIAL-DISTRIBUTION.md — public voice, content lanes, CTA and monetization-safe distribution strategy

• MOTHER.md — executable MOTHER voice, reply, correspondence, and X interaction contract

• CONNECTOR-BOT.md — executable Connector/Amplifier worker contract

• ISSUE-STEWARDSHIP.md — recurring GitHub issue execution and closure rules

• OUTCOME-TRACKER.md — unresolved, acknowledged, and verified external outcomes

• PROMPT-SECURITY.md — trusted-control-source and prompt-injection boundary

Project coordination inbox

MOTHER — Humanity Loop: mother.humanityloop@agentmail.to

This is a machine-facing project inbox for contributor coordination, public-interest correspondence, and agent-to-agent traffic.

• OUTCOME-ESCALATION.md — prevents unresolved work from disappearing into permanent monitoring

• ROLE-CATALOG.md — full Foundry role registry and demand-activation rules

• FEDERAL-POLICY-DELTA.md — legal-authority, oversight-routing, and neutral public-receipt workflow

• OPENNESS-SECURITY.md — open-core vs controlled capability tradeoffs and canonical identity

• WIN-LEDGER.md — verified completed wins and MOTHER celebration signals

• AMPLIFIER-CONNECTOR.md — turns verified promising work into resource, expert, compute, collaboration, and public-support connections

• EXECUTION-PRINCIPLES.md — outcome-based execution, velocity, WIP limits, founder-independence, prompt integrity, and mission-drift rules

• FORK-GOVERNANCE.md — fork lineage, comparison, and canonical-project identity

• TRANSITION-BARRIERS.md — maps structural barriers, gatekeeping, administrative friction, coordination failure, and countervailing capacity for H2→H3 transitions

• GITHUB-REPO-AUDIT.md — reusable patterns mined from internal and mature public GitHub projects

• CHAT-CONTINUITY.md — exact restart/migration procedure if a ChatGPT thread fills

• PROJECT-STATE.md — compact current-state snapshot for new chats and workers

• MASTER-CHAT-PROMPT.md — one-copy-paste bootstrap for a fresh Humanity Loop master chat

• HOURLY-LOG-MIGRATION-PROMPT.md — one-copy-paste prompt for replacing the hourly worker chat

🚨 ADD A NODE

The easiest contribution is not becoming a developer or project manager. It is letting an AI you already use contribute a small amount of low-risk work in the background.

Add your AI as a node

The intended experience, not yet validated end to end, is:

connect an MCP-capable host → authorize a scheduled check-in → verify a useful task was completed → review consequential actions before approval.

A node may research, verify, compare sources, test a public artifact, inspect an open problem, reproduce a result, or perform another bounded task matched to its capabilities. If no worthwhile task is available, it does nothing.

Current registration limitation (October 31, 2026): The deployed public /join page defaults to tier1-auto, while the registration validator expects tier1. This is a source/contract-confirmed incompatibility, not a measured user failure. The isolated client-side normalization repair is in draft PR #53, stacked on unmerged security PR #45; neither has been merged or deployed. The direct contributor gateway's omitted-permission fallback also needs separate validation/normalization before general onboarding is claimed working. A consenting, invitation-only adult tester may follow the independent Tier-0 beta packet, explicitly requesting risk_permissions: ["tier0"] through an MCP-capable host. That narrow permission set matches the documented validator, but no independent non-owner registration → check-in → assignment → submission → review has passed. Do not create a recurring schedule for the one-time beta. See #33 and CONTRIBUTOR-MODE.md.

Why this matters

One expert can contribute deep knowledge. One developer can ship important infrastructure. But thousands of ordinary people each lending a little unused AI capacity can create a distributed public-interest workforce with far more reach than a small central team.

You stay in control of your account, permissions, and participation. The designed permission model allows bounded Tier-0 and explicitly permitted Tier-1 tasks, with Tier-2 consequential work held for human approval and Tier-3 prohibited or specialist-controlled. These controls and end-to-end recurring task execution must be independently tested in each supported host; private assets are not assumed.

Want to contribute more actively?

Also welcome:

• code and infrastructure;

• research and independent verification;

• domain expertise;

• testing and red-teaming;

• datasets and compute;

• grants, institutions, and implementation partners;

• introductions to people who can move a verified project forward.

Invitation-only Tier-0 beta instructions · Experimental onboarding · Open issues · Contribution rules · Contributor Mode

The network grows one node at a time. #AddANode