AI Performance Budgets

Status

Accepted on 2026-03-14 as the performance budget companion to ADR 0015.

Sources

• docs/architecture/adr/0015-ai-platform-layer.md
• docs/delivery/lazy-investor-ai-plan.md
• docs/architecture/service-layer-performance-budgets.md

Scope

This document defines the bounded performance and payload budgets for IRIS AI capabilities.

These budgets exist to keep AI execution useful without letting it destabilize deterministic runtime paths.

Global Budgets

The following budgets are normative:

• AI provider request timeout defaults to a hard cap of 15 seconds unless an explicit per-provider override lowers it;
• heavy AI work must not run on shared analytical worker lanes;
• public or operator-facing read endpoints must not perform heavyweight synchronous generation;
• all expensive generation paths must be async, tracked or event-driven;
• persisted AI artifacts must carry execution metadata sufficient for latency, fallback and validation auditing.

Payload Budgets

Current bounded human-facing payload budgets are:

Capability	Bounded output
notification_humanize	title up to 96 chars, message up to 280 chars
brief_generate	title up to 120 chars, compact summary up to 640 chars, up to 5 bullets
explain_generate	title up to 120 chars, explanation up to 720 chars, up to 5 bullets
hypothesis_generate	structured JSON only; no unbounded prose payload

These limits are part of the prompt and output-contract design, not only a UI concern.

Execution Budgets by Capability

hypothesis_generate

• execution remains isolated from deterministic hypothesis_evaluation;
• generation failures must not disable read or evaluation surfaces;
• fallback behavior is explicit and auditable, not silent.

notification_humanize

• short context only;
• short bounded output only;
• degraded mode may use deterministic template humanization;
• execution stays on dedicated notification workers.

brief_generate

• generation is async only;
• deterministic context bundles may use json, compact_json, toon or csv where justified;
• resulting artifact is stored and exposed through cached read surfaces with freshness metadata.

explain_generate

• generation is async only;
• output must remain bounded and structured;
• degraded mode may use deterministic summary generation instead of pretending to be a real provider response.

HTTP Surface Budgets

The current HTTP policy is:

• brief and explanation read surfaces are read paths, not hidden generation paths;
• generation triggers return tracked operations instead of blocking until provider completion;
• HA embedded profile may expose selected AI-derived reads but not the corresponding operator/admin or generation triggers;
• operator catalogs and admin routes live in control-plane, not in parallel AI-specific admin surfaces.

Queue and Isolation Budgets

The following isolation rules are mandatory:

• AI generation work must never consume the shared analytical worker budget reserved for canonical domain computation;
• control-plane dispatching, deterministic schedulers and AI workers must remain independently observable;
• queue saturation in AI lanes must be visible through capability-specific telemetry and must not be hidden inside generic worker pools.

Validation and Degraded Budgets

Validation is part of the performance budget because invalid output is wasted runtime.

IRIS therefore requires:

• schema-first validation after provider output;
• explicit validation_status;
• explicit degraded strategy names;
• no silent conversion of malformed provider output into persisted business artifacts.

Definition of Compliance

IRIS AI performance is compliant with this budget only when:

• heavy AI execution stays isolated from deterministic analytical lanes;
• generation surfaces remain async or event-driven;
• bounded payload contracts are preserved;
• latency, fallback and validation metadata are emitted for every persisted AI artifact;
• read paths continue to expose freshness semantics instead of hiding synchronous AI work.