Active-Local

Atlas and Sentinel: the codebase intelligence system.

A system that maps the WHY of its own code, and watches itself act.

The graph maps every code unit, decision, doctrine, contract, and runtime to the reason it exists, and injects that map into the agent every turn. Sentinel flags crashes and errors as they surface. Together they bound the reasoning with evidence, the way the operating loop bounds execution with receipts.

This is not a separate user product. It is wired into the per-turn packet that drives the agent that builds and operates Aria. The cockpit that briefly appeared on this page belonged to a different product and has been removed.

01 / Problem and system

Why reasoning needed a substrate.

Code reasoning and self-watching are scattered. The agent assumes. The runtime drifts. Decisions land without a traceable cause.

Problem

Reasoning without provenance

An agent that reads code without a map of why the code exists reasons from habit, from training, or from the last file it touched. Errors are observed after they ship. Decisions are made without a record of which doctrine or contract they satisfy.

Solution

A live provenance graph plus a per-turn watch

A graph that resolves every code unit, decision, doctrine, contract, and runtime to the reason it exists, injected into the agent before it acts. A Sentinel loop that flags high-severity runtime failures as they surface. Real proof sits in the proof room under NDA.

02 / Self-intelligence surface

The graph that tells the agent WHY before it acts.

The frame below is the live self-intelligence surface. It is filled from the system itself, not from a static document. Atlas reads the codebase, builds the map, and ships the slice the agent needs. Sentinel reads the runtime and reports the count.

codebase intelligence . self-map, self-watch
SELF-INTELLIGENCE
The graph that tells the agent WHY before it acts
Every code unit, decision, doctrine, contract, and runtime resolved to its reason. Watched by Sentinel. Injected per turn, never narrating from memory.
03 / What the graph contains

Six kinds of nodes, all cited.

The map is not a flat index. It is a typed graph. Each kind of node is counted and resolvable to a source.

Code units
67,115
Source files, functions, classes, and modules resolved to their declared purpose.
Decisions
4,597
Architectural and product decisions recorded with the reason and the moment they were made.
Doctrines
1,122
First-principles rules and constraints the system holds itself to, versioned and traceable.
Contracts
972
API, schema, and interface contracts that units and runtimes agree to satisfy.
Runtimes
791
Services, workers, queues, and surfaces that execute code in production.
Owner directives
105
Explicit instructions from the owner, recorded as constraints the graph respects.
04 / Edge relations, cited

300,563 edges across 8 named relations.

Nodes alone are a taxonomy. The graph is the edges. Every relation has a type, a count, and a note about what it carries.

IMPLEMENTS
Carries a contract into a unit, so a function is bound to the interface it claims to satisfy.
DELEGATES_TO
Hands work between units, so a call chain is traceable from entry to execution.
EVOLVED_FROM
Tracks learning and iteration, so a refactor points back to the version it replaced.
MENTIONS
Links a unit to a doctrine, so a code reference surfaces the rule it touches.
BEHAVIORALLY_COUPLED
Links that move together under change, so a blast radius is observable from the graph.
DEPLOYED_AS
Names the surface a unit runs on, so a runtime claim is provable from the source.
BOUND_BY
Names the doctrine that constrains a unit, so a constraint violation is detectable.
GOVERNS
Names the decision or doctrine that controls a unit, so authority is read from the graph.
05 / Per-turn injection

The actual contract with the agent.

Atlas is injected at the pre-cognition stage of the per-turn packet, before the agent selects a tool or emits a thought. Sentinel flags at the post-tool stage, after the runtime reports back. The two loops run async and in parallel, so self-map and self-watch do not block each other. The per-turn stages are pre-cognition, pre-tool, intra-cognition, intra-tool, post-cognition, and post-tool.

Stage 01
Atlas inject

Pre-cognition. The graph slice for the active context is loaded into the agent before it reasons.

Stage 02
Agent acts on provenance

Intra-cognition and intra-tool. The agent reasons from the cited edges and acts on the cited units.

Stage 03
Sentinel flags

Post-tool. Runtime errors and crashes are reported back, severity is counted, and the per-turn high count is exposed in the proof room.

The per-turn Sentinel high count is exposed in the proof room and updated each turn. Currently zero unresolved.

06 / Substrate and snapshots

9 substrates, 46 snapshots, time-diffable.

The graph is not frozen. It is built from 9 substrates: behavioral, cognitive, conversational, doctrinal, historical, hive-state, repo-state, runtime, and structural. Each substrate contributes a layer of nodes and edges. The combined state is captured in 46 snapshots, so a change to the code is diffable against the substrate it came from. A refactor traces to its cause before it lands, not after.

Substrates
9
Behavioral, cognitive, conversational, doctrinal, historical, hive-state, repo-state, runtime, structural.
Snapshots
46
Point-in-time captures of the combined graph state, diffable across time.
Total nodes
74,702
The current size of the live map, summed across all six node kinds.
07 / What it proves, and where it stops

Honest boundaries.

The system is real. The proof is real. The limits are also real, and they are named here on purpose.

Proves
  • A live doctrine-bound self-map of the codebase, with 74,702 nodes and 300,563 cited edges.
  • A per-turn provenance injection before the agent acts.
  • A Sentinel watch on runtime failures, with severity counts exposed in the proof room.
  • A cited-edge basis for blast-radius reads on a refactor.
Does not prove
  • That every architectural decision is correct. The graph records decisions. It does not validate them.
  • That every relationship is right. Edges can be mislabeled. The graph is reviewed, not self-certifying.
  • That the system is autonomous. Self-watching is not self-correction. Sentinel flags. A human or a higher gate acts.
08 / Business outcome and next proof

What the substrate buys, and what it still owes.

The boundary between a system that remembers code and a system that knows its code is the boundary Atlas closes. The next mile of proof is more demanding than the last.

Outcome

The agent reasons from WHY per turn

The agent that builds and acts on Aria's own code reasons from the cited WHY of every unit, every turn, with a watch on what breaks. That is the governance layer frontier labs still treat as research.

Next proof required

Incident-tied counts and refactor diffs

Per-turn Sentinel counts tied to specific incidents, and a diff of the graph across a substantive refactor showing the substrate actually caught the blast radius. Both are queued for the next review window.

All case studies