Definitions, named.

A quiet software studio that builds vertically integrated AI engineering infrastructure. The lowercase wordmark is intentional — smpl is meant to read as an imprint, not a hero. Brand philosophy: reduction without loss.

The capability smpl produces. Continuously maintained structural understanding of a codebase that survives across sessions, accumulates institutional memory, and grounds investigation and execution in actual system state rather than improvised context.

smpl's selective, manually-reviewed entry point — a written diagnostic of where legibility, context loss, and structural drag are most expensive in a specific codebase. The first step into a deeper engagement.

The codebase intelligence layer. Fuses semantic retrieval, structural code intelligence, relation-aware retrieval, and domain discovery into a single continuously maintained structural view of a codebase. Symbol-, type-, and reference-level truth that plain-text search cannot provide.

The investigation engine. Evaluates each ticket through a pipeline of bounded-epistemology stages — reconnaissance, archetype classification, problem scoping, archetype-specific investigation — and produces targeted pushback questions grounded in actual codebase analysis. Ambiguity surfaces before engineering time is spent on it.

The institutional-memory substrate. A structured, git-backed knowledge repository where investigations, architectural decisions, accumulated specialist expertise, and resolved incidents are captured as queryable findings. The 50th investigation in a service is grounded in everything the previous 49 surfaced.

The orchestration substrate. Decomposes work into a DAG with explicit dependency edges, runs each task in a bounded context, and structurally prevents context exhaustion. Where the layers above are stable, drafts changes, scaffolds migrations, and opens pull requests with tests and evidence — always carrying the context of the layers beneath.

An automated research platform that runs experiments across the seven optimization tiers, scored quantitatively against an evaluation corpus that grows with every real ticket. Treats every prompt, configuration, context strategy, and tool exposure as a tunable parameter calibrated empirically rather than by intuition.

Per-stage, per-archetype specialist agents that accumulate domain-specific expertise across many tickets in the same service. Faculty knows its domain deeply, knows the limits of its domain, and defers outside its scope. Their accumulated findings are first-class retrieval sources for future investigations.

Recon's typed, mandated, bounded sequence: Reconnaissance → Archetype Classification → Problem Scoping → Investigation (with branches per archetype). The structure is the enforcement — no stage sees the full investigation, no stage is permitted to reason outside its scope.

A workspace-scoped configuration of external data sources that investigation agents can query during analysis: code intelligence (Lux), Linear, Slack, Sentry, read-only database replicas, log groups. Credentials are resolved per workspace and injected ephemerally, scoped to a single evaluation.

The principle that an agent should know the limits of its knowledge and operate strictly within its epistemic domain. Enforced structurally — through typed interfaces, separated planes, and pipeline stages — rather than encouraged through prompts. Humility imposed by guideline disappears under load; humility imposed by ontology can't be escaped.

The taxonomy that defines an agent before it defines what the agent does. Data Layer specifies what the agent perceives, produces, and remembers. Logic Layer specifies how it reasons and what authority it holds. Action Layer specifies what it can affect and where it is forbidden. Behavior follows from being.

The semantic logic modes that map separated cognitive powers onto agents. Executive (binding decisions only), Operational (direct execution), Advisory (guidance without authority), Observational (analysis without modification). Each agent operates within a bounded cognitive scope.

The architectural embodiment of the ontology. Each plane runs in its own infrastructure with strict write boundaries that match its ontological role. The orchestrator runs in Control and writes nothing; advisory and observational agents run in Knowledge and write only to long-term memory; operational agents run in Action and write only to artifacts.

The five layers that determine the quality of any AI-driven use case, where each layer is a ceiling for the layers below. The compounding gains live in simultaneous optimization, not depth at any single layer.

Researcher's optimization ladder, from Tier 0 (prompt variants) through Tier 6 (dataset curation). Each tier consumes the signal produced by the tiers below it. Lower tiers are cheap and fast; higher tiers are expensive and slow but increasingly structural.

The persistent failure mode where an agent spends its reasoning budget navigating its own context rather than solving the problem. The architecture defends against it at four levels simultaneously — persona, ontological, epistemological, and DAG-orchestration segmentation — none sufficient on its own.

A complete, self-contained execution unit — Team Lead, Implementation, Verification, Validation — that pulls work from a backlog and ships verified artifacts. ATUs are functionally substitutable. Scale comes from running more in parallel, not from growing any individual one.