Project Introduction
Relevant source files
The following files were used as context for generating this wiki page:
Noesis is a research-preview system designed to achieve "autonomous compounding," a process where the system's capabilities improve through normal use without manual pipeline operation. It functions by turning ordinary work into evidence, utilizing LLMs to author prompt, skill, and tool improvements, and automatically activating these improvements based on low-risk evidence-backed trials.
The project is structured to maintain a thin layer of deterministic code for rails—such as transactional state, authority, and bounded execution—while allowing semantic intelligence and capability revisions to evolve rapidly. Users retain the ability to inspect, guide, veto, or revert active revisions at any time.
Sources: plans/noesis-autonomous-compounding-implementation-plan.html:150-165, AGENTS.md:3-10
Core Product Thesis
The central thesis of Noesis is that "later work is better because earlier work happened." This is achieved through a self-correction loop where the system reflects on user interactions and recurring patterns to propose narrow, evidence-linked learnings.
The Compounding Loop
The system follows a specific lifecycle to evolve its capabilities:
- Signal Detection: Identifying explicit corrections, preference expressions, or recurring failures.
- LLM Reflection: Forming hypotheses and experiment briefs.
- Revision Authoring: Authoring candidate
CapabilityRevisions(prompts, skills, tools, and routers). - Preflight Evaluation: Running blinded trials against a baseline revision.
- Activation: Automatically activating low-risk revisions through a protected control plane.
- Observation: Gathering real-world feedback to decide whether to keep, revise, or revert.
Sources: plans/noesis-autonomous-compounding-implementation-plan.html:168-175, plans/noesis-high-leverage-correction-plan.html:125-130, AGENTS.md:12-19
The diagram above illustrates the end-to-end causal loop from user input to autonomous capability activation.
Sources: plans/compounding-partnership-product-loop.html:420-435, plans/noesis-autonomous-compounding-implementation-plan.html:420-435
System Architecture
Noesis follows a modular architecture with earned boundaries, ensuring that research objects can evolve without compromising the integrity of the Pi adapter, workspace, or authority boundaries.
Module Breakdown
| Package | Responsibility | Authority |
|---|---|---|
@noesis/workspace | Owns SQLite transactions, migrations, and canonical file locations. | SQLite |
@noesis/intelligence | Hybrid search over session history, message rows, and file revisions. | Rebuildable Indexes |
@noesis/runtime-pi | The only Pi import boundary; handles foreground role execution. | Pi Adapter |
@noesis/runtime | Adapter-neutral turn orchestration and activation decisions. | Deterministic Rails |
@noesis/learning | LLM reflection, signal harvesting, and revision authorship. | Evidence |
@noesis/evals | Blinded judges, paired trials, and preflight reports. | Evidence Revisions |
@noesis/policy | Effect enforcement, principal grants, and protected rules. | Protected Kernel |
Sources: plans/noesis-autonomous-compounding-implementation-plan.html:245-310, packages/domain/test/architecture.test.ts:100-140
Authority Matrix
Noesis enforces a "one authority per datum" rule to prevent divergence between storage systems.
- SQLite: Authoritative for operational state (sessions, messages, tool calls, experiments, activation pointers).
- Editable Files: Authoritative for declarative definitions (prompts, skills, eval definitions).
- Byte Snapshots: Used for immutable definition revisions and evaluation evidence.
- Rebuildable Indexes: FTS and embedding rows derived from canonical sources.
Sources: plans/noesis-autonomous-compounding-implementation-plan.html:330-365, AGENTS.md:25-35
Protected Control Plane
To ensure safety during autonomous self-modification, Noesis implements a "Protected Control Plane" that exists outside the reach of generated LLM content.
Invariants
- Role Separation: LLMs (Reflection, Authorship) may propose changes but never activate them.
- Verification: All side effects and activation transitions must pass through an
AuthorityBoundaryandEffectGateway. - Receipts: Protected mutations require an
AuthorityReceiptverified at the deep adapter layer. - Hard Rails: Protected evaluation suites act as hard gates; regressions on these suites result in automatic blocks regardless of qualitative scores.
Sources: AGENTS.md:46-60, plans/noesis-high-leverage-correction-plan.html:260-310
This sequence diagram shows how the system ensures that only verified, evidence-backed revisions are committed to the authoritative store.
Sources: plans/noesis-high-leverage-correction-plan.html:290-315, plans/noesis-autonomous-compounding-implementation-plan.html:380-395
Measurement and Falsifiability
Noesis includes instrumentation to prove its compounding thesis through "Effect-free paired shadow replay."
Key Compounding Metrics
- Served-revision win rate: Blind-judged paired wins of the active revision over its baseline.
- Scope leakage rate: The share of turns where an adaptation was served in an unrelated context.
- Context tax: Monitoring the growth of injected-context tokens to ensure efficiency.
- Correction recurrence: The rate at which user corrections decrease after a relevant activation.
Sources: plans/noesis-high-leverage-correction-plan.html:85-115
Technical Implementation Details
The project utilizes modern TypeScript (ESM) and Node.js. It strictly forbids certain dependencies and enforces architecture through automated tests.
- Runtime Independence: The adapter-neutral
packages/runtimemust never importpackages/runtime-pi. - Error Handling: Only specific first-party native Error subclasses are allowed (e.g.,
NoesisConfigError). - Data Validation: Zod is used to validate unknown data at SQLite, file, and IPC boundaries.
- Local Backend: Generated tools run in a local temporary-directory child-process backend with sanitized environments and time/output bounds.
Sources: packages/domain/test/architecture.test.ts:70-95, plans/noesis-autonomous-compounding-implementation-plan.html:435-460, AGENTS.md:65-75
Conclusion
Noesis is a specialized architecture designed for self-evolving systems. By isolating intelligence (LLM-driven) from authority (deterministic code-driven), it creates a framework where agents can safely learn from their own interactions while remaining inspectable and revertible by human users. The project prioritizes evidence-backed compounding over speculative capability expansion.
Sources: plans/noesis-autonomous-compounding-implementation-plan.html:645-660, plans/noesis-high-leverage-correction-plan.html:360-375