Self-modifying RSI Agent born in and constrained to a browser : Model CPS-9204
JavaScript
2
171 commits
updated Sep 23, 2026
README
Reploid
Reploid is a network of agents that run models, coordinate work over WebRTC, and develop, test, and exchange improvements to their own problem-solving methods. Each agent remains useful independently, and each participant controls what it shares and contributes.
Zero is its minimal starting configuration; X extends Zero with explicit capabilities and supporting blueprints. Poolday supplies optional discovery, artifact exchange, bounded peer computation, recovery, and reusable experience. Evaluation, approval, and activation are separately authorized roles. Change Passport remains an inactive alternative; Research Room-1 is an optional scientific workflow with its own evidence and admission requirements.
Who uses it
Repository mission, value, and durable strategy live in GOALS.md. Poolday's user workflow and evidence boundary live in its product intent.
The repository serves:
- Humans and agents pursuing bounded goals through authorized tools and models.
- Experimenters testing capability acquisition and recursive improvement in Zero or X.
- Public protein catalog curators testing a disputed family or domain annotation.
- Research Room requesters, compute contributors, and accountable reviewers.
- Runtime and product contributors working on browser execution and room state.
- Security and claim reviewers checking records, relay boundaries, and evidence.
- Researchers designing the future active-science workflow.
How to use Reploid
Install and start the local browser surface:
npm install
npm start
Open http://localhost:8000. The managed Gemini path requires GEMINI_API_KEY
in .env before starting.
Start with a task and choose local Qwen 3.5 2B or the configured Gemini cloud model. Optional controls enable helper agents, approved peer requests, and tested tool improvements. The same page connects participating devices and shows candidate code, test comparisons, adoption, and rollback. Try Improve a tool for the registered JSON formatter. See the workflow and its limits.
The product surface is:
| Surface | Route | Use |
|---|---|---|
| Reploid | / | Agents, models, shared tasks, contribution, and tested improvements together. |
| Zero | /zero | Minimal self-loading configuration for problem-solving and improvement experiments. |
| X | /x | Zero with prepared executable extensions and supporting blueprints. |
Poolday is an internal name for optional peer infrastructure, not a separate public product. Zero's current inference default is the server proxy, with optional local Doppler execution; a browser agent does not imply offline inference. The signed executable-Pack path requires a qualified catalog entry and matching released Doppler API. Existing model-name loading and adapter sharing do not prove base-model Pack delivery.
Evidence and current surfaces
The surface claim index owns the current support claims and their evidence paths.
| Claim row | Current boundary |
|---|---|
local-execution | A configured local executor runs slots in the current browser. |
peer-slot-placement | Opted-in slots may run on joined peers; joining does not expose local inference without a local executor. |
browser-provider-roles | Requesters and providers exchange assignments, outputs, and receipts through peer rooms. |
signaling | Same-browser rooms can use BroadcastChannel; cross-host WebRTC uses signaling for rendezvous. |
sealed-credentials | npm start can build sealed access windows; client artifacts omit the plaintext key. |
public-mesh | Blocked as a signaling-free claim while cross-host rendezvous requires signaling. |
The peer-execution path provides receipt-backed browser inference. Its claimed local model execution and browser agent state stay in the browser; explicitly selected cloud providers have their own execution boundary. Services may handle authentication, rendezvous, policy enforcement, receipt anchors, and ledger projections; they do not perform the claimed browser-local model execution. The current enabled scientific peer catalog covers public protein inputs and the pinned ESM-2 contract; that scope does not define every Reploid task.
Limits and status
Reploid does not claim hardware attestation, independently trustworthy browser/GPU execution, or guaranteed honest providers. Relay acknowledgement proves receipt of a relay record, not execution truth. Capability claims require their declared evidence; ordinary networking and experimental use do not establish improvement. Learned routing and scientific-policy promotion retain their specific admission gates. Read the claim index row before repeating a capability statement.
Repository map
self/: browser boot profiles, VFS, tools, and runtimedocs/: product intent, claims, security, architecture, and operator guidesdeploy/: deployment and access-window toolingdoppler/: vendored or paired Doppler integration surfaceshowcase/: demonstrations and recorded runspackage.json: package metadata and local commands
Intent and component authority
- GOALS.md owns the repository mission, value, and durable strategic goals.
- CATSCAN.md is the root component charter. Child charters narrow its authority for independently meaningful components.
- The generated component index lists every charter, parent, and target.
- AGENTS.md defines how code agents discover and obey the charter chain.
- The workspace CATSCAN protocol defines the shared shape and precedence rules.
Run npm run catscan:chain -- <path> to print the charter chain for a target file. Run npm run verify:catscan to validate fields, parents, links, evidence paths, identifiers, size, and the generated index.
Read next
- Repository goals
- Root component charter
- Component index
- Documentation index
- Poolday product intent
- Discovery Contract
- Poolday claims and non-claims
- Security model
- RSI improvement episodes
- Doppler
License
MIT License. The package metadata also declares MIT.
Contributors
171 commits
Languages
JavaScript
90.7%
WGSL
4.8%
CSS
2.8%
Self-modifying RSI Agent born in and constrained to a browser : Model CPS-9204
JavaScript
2
171 commits
updated Sep 23, 2026
README
Reploid
Reploid is a network of agents that run models, coordinate work over WebRTC, and develop, test, and exchange improvements to their own problem-solving methods. Each agent remains useful independently, and each participant controls what it shares and contributes.
Zero is its minimal starting configuration; X extends Zero with explicit capabilities and supporting blueprints. Poolday supplies optional discovery, artifact exchange, bounded peer computation, recovery, and reusable experience. Evaluation, approval, and activation are separately authorized roles. Change Passport remains an inactive alternative; Research Room-1 is an optional scientific workflow with its own evidence and admission requirements.
Who uses it
Repository mission, value, and durable strategy live in GOALS.md. Poolday's user workflow and evidence boundary live in its product intent.
The repository serves:
- Humans and agents pursuing bounded goals through authorized tools and models.
- Experimenters testing capability acquisition and recursive improvement in Zero or X.
- Public protein catalog curators testing a disputed family or domain annotation.
- Research Room requesters, compute contributors, and accountable reviewers.
- Runtime and product contributors working on browser execution and room state.
- Security and claim reviewers checking records, relay boundaries, and evidence.
- Researchers designing the future active-science workflow.
How to use Reploid
Install and start the local browser surface:
npm install
npm start
Open http://localhost:8000. The managed Gemini path requires GEMINI_API_KEY
in .env before starting.
Start with a task and choose local Qwen 3.5 2B or the configured Gemini cloud model. Optional controls enable helper agents, approved peer requests, and tested tool improvements. The same page connects participating devices and shows candidate code, test comparisons, adoption, and rollback. Try Improve a tool for the registered JSON formatter. See the workflow and its limits.
The product surface is:
| Surface | Route | Use |
|---|---|---|
| Reploid | / | Agents, models, shared tasks, contribution, and tested improvements together. |
| Zero | /zero | Minimal self-loading configuration for problem-solving and improvement experiments. |
| X | /x | Zero with prepared executable extensions and supporting blueprints. |
Poolday is an internal name for optional peer infrastructure, not a separate public product. Zero's current inference default is the server proxy, with optional local Doppler execution; a browser agent does not imply offline inference. The signed executable-Pack path requires a qualified catalog entry and matching released Doppler API. Existing model-name loading and adapter sharing do not prove base-model Pack delivery.
Evidence and current surfaces
The surface claim index owns the current support claims and their evidence paths.
| Claim row | Current boundary |
|---|---|
local-execution | A configured local executor runs slots in the current browser. |
peer-slot-placement | Opted-in slots may run on joined peers; joining does not expose local inference without a local executor. |
browser-provider-roles | Requesters and providers exchange assignments, outputs, and receipts through peer rooms. |
signaling | Same-browser rooms can use BroadcastChannel; cross-host WebRTC uses signaling for rendezvous. |
sealed-credentials | npm start can build sealed access windows; client artifacts omit the plaintext key. |
public-mesh | Blocked as a signaling-free claim while cross-host rendezvous requires signaling. |
The peer-execution path provides receipt-backed browser inference. Its claimed local model execution and browser agent state stay in the browser; explicitly selected cloud providers have their own execution boundary. Services may handle authentication, rendezvous, policy enforcement, receipt anchors, and ledger projections; they do not perform the claimed browser-local model execution. The current enabled scientific peer catalog covers public protein inputs and the pinned ESM-2 contract; that scope does not define every Reploid task.
Limits and status
Reploid does not claim hardware attestation, independently trustworthy browser/GPU execution, or guaranteed honest providers. Relay acknowledgement proves receipt of a relay record, not execution truth. Capability claims require their declared evidence; ordinary networking and experimental use do not establish improvement. Learned routing and scientific-policy promotion retain their specific admission gates. Read the claim index row before repeating a capability statement.
Repository map
self/: browser boot profiles, VFS, tools, and runtimedocs/: product intent, claims, security, architecture, and operator guidesdeploy/: deployment and access-window toolingdoppler/: vendored or paired Doppler integration surfaceshowcase/: demonstrations and recorded runspackage.json: package metadata and local commands
Intent and component authority
- GOALS.md owns the repository mission, value, and durable strategic goals.
- CATSCAN.md is the root component charter. Child charters narrow its authority for independently meaningful components.
- The generated component index lists every charter, parent, and target.
- AGENTS.md defines how code agents discover and obey the charter chain.
- The workspace CATSCAN protocol defines the shared shape and precedence rules.
Run npm run catscan:chain -- <path> to print the charter chain for a target file. Run npm run verify:catscan to validate fields, parents, links, evidence paths, identifiers, size, and the generated index.
Read next
- Repository goals
- Root component charter
- Component index
- Documentation index
- Poolday product intent
- Discovery Contract
- Poolday claims and non-claims
- Security model
- RSI improvement episodes
- Doppler
License
MIT License. The package metadata also declares MIT.
Contributors
171 commits
Languages
JavaScript
90.7%
WGSL
4.8%
CSS
2.8%