AirNLP
Try HOWI

Platform

Latest Advancements

A running record of the systems behind AirNLP and HOWI — what shipped, when, and why.

Contourv1.0.0Meridianv1.2.0Soundingv1.1.0Kilnv2.2.0Isochronev1.0.0Escapementv1.1.0Concordancev1.1.0Plumblinev1.3.1Fulcrumv1.1.0

SCOUT

Contourv1.0.0Intelligence & Signalsbeta

SCOUT is AirNLP's governed self-improvement engine. It continuously observes AirNLP for evidence-backed weaknesses, generates SCOUT Reports and bounded Missions with reproducible risk/priority scoring, and — for an explicit allowlist of low-risk, reversible change classes — autonomously experiments inside an isolated git worktree, validates against required evaluators (typecheck, full test suite, lint), and integrates the result to a candidate branch under policy authorization. It never edits the active branch directly, and every action is recorded to an append-only audit log with a tested kill switch and pause controls SCOUT cannot bypass itself. Currently in beta: autonomous production release, live post-release monitoring, and automatic rollback are not yet implemented — those stages remain human-gated pending dedicated release infrastructure.

Observes AirNLP signals (repo history, build quality, Systems Registry drift, documentation drift, redacted feedback) on a governed scheduleGenerates evidence-grounded SCOUT Reports and bounded Missions with reproducible risk and priority scoringRuns isolated git-worktree experiments with command allowlisting, resource budgets, and required evaluatorsIntegrates validated low-risk changes to a candidate branch under explicit policy or human authorizationGenerates SCOUT Memory from every completed Mission and records a full audit timelineEnforces a global kill switch and product/mission pauses that SCOUT cannot weaken or bypass itself
PowersAirNLP Research

July 2026

major

v1.0.0 · Jul 15, 2026

Initial registration: governed observe → diagnose → prioritize → propose → improve → validate → integrate → learn cycle for AirNLP, running in isolated git worktrees with policy-gated autonomy (L2 experiment / L3 local integration for an explicit low-risk allowlist). Production release (L4) is disabled by default policy.

HOWI Experience Engine

Meridianv1.2.0Intelligence & Signalsactive

Shared context, journey, event, and surface-presentation infrastructure underneath the guest link and kiosk — Pass 3 adds production lifecycle (retention), privacy-safe SCOUT signal aggregation, contextual-help / support contracts, and expanded Studio diagnostics.

Versioned HOWI Experience Context with Available<T> fields — never fabricated12 evidence-grounded guest journeys with client-side state persistence25-name privacy-safe Experience Event taxonomy with shadow-mode ingestionresolveSurfaceConfig() — deterministic kiosk vs guest-link presentationShared navigation registry — same destinations on every guest surfaceReal capability, announcement, and event wiring from property_config and live data180-day experience-event retention job with dry-run and batch limitsPrivacy-safe SCOUT Experience Events adapter (aggregates only, flag-gated)Contextual Help Registry + Guide Adapter contract + stable guide anchorsProduct-support namespaces, diagnostics, actions, and escalation contracts
PowersHOWI

July 2026

minor

v1.2.0 · Jul 15, 2026

Pass 3: production hardening — recoverable error boundary telemetry, connectivity recovery, experience-event retention cron, privacy-safe SCOUT guest_experience_events adapter, contextual help / guide / support foundations, and expanded Experience Studio diagnostics (health, parity, support readiness, rollout).

minor

v1.1.0 · Jul 15, 2026

Unified the kiosk and guest-link presentation behind a deterministic Surface Contract and shared navigation registry; wired real capability, announcement, and stay-event data into the Guest Experience Context; extended the Journey Registry and Experience Events into the shared navigation/overlay layer; added a mobile drawer, browser-back navigation, and safe-area/keyboard handling for the guest-link surface; added Surfaces and Flags inspectors to the Studio Experience department.

major

v1.0.0 · Jul 14, 2026

Initial release: Guest Experience Context, Journey Registry, Support Context contract, and redacted Experience Events pipeline, with an AirNLP Studio inspector.

Deep Research Pipeline

Soundingv1.1.0Research Pipelineactive

A multi-pass reasoning pipeline that researches, drafts, critiques, and polishes AirNLP's long-form research articles, then adds supporting charts and diagrams.

Outline, draft, self-critique, and polish across four sequential passesLive web research injected into the drafting passA dedicated visuals pass so pieces ship with supporting charts or diagramsRecent-post overlap avoidance via semantic corpus search
PowersAirNLP Research

July 2026

minor

v1.1.0 · Jul 13, 2026

Added a dedicated visuals pass and a public "Deep Research" flag on published posts.

minor

v1.0.0 · Jul 10, 2026

Initial release: four-pass generation (outline, draft, critique, polish) wired into Autopilot's trending mode.

Editorial Engine

Kilnv2.2.0Content Generationactive

AirNLP's core article generation system — turns a topic or idea into a fact-checked, illustrated, publish-ready research post.

Standard one-pass generation with grounding rules built into the system promptAutomatic fact-check pass against sources before publishChart and diagram generation, embedded directly into the articleAutomatic cover artwork generationIdea suggestion with duplicate-topic detectionOne-click repurposing into LinkedIn, X, and newsletter copy
PowersAirNLP Research

July 2026

minor

v2.2.0 · Jul 13, 2026

Chart and diagram generation embedded directly into articles, an in-article visuals regenerator for existing posts, and admin feed-placement controls for the public research index.

minor

v2.1.0 · Jul 11, 2026

One-click repurposing of a published post into LinkedIn, X, and newsletter copy.

major

v2.0.0 · Jul 10, 2026

Grounding rewrite: added the fact-check pass, the News category, and Deep Research mode integration.

minor

v1.1.0 · Jul 9, 2026

Automatic cover artwork generation added.

major

v1.0.0 · Jul 7, 2026

Initial release: draft/publish workflow and the research_posts store.

Command Center Intelligence

Isochronev1.0.0Intelligence & Signalsactive

A registry of independent signal sources — editorial activity, competitive moves, product activity, documentation drift, growth metrics, and guest-quality trends — merged and synthesized into one daily briefing.

Seven independent signal sources running in parallel with isolated failure handlingOne synthesized highlights-and-recommendations briefing per loadRaw per-source feeds available without synthesis
PowersAirNLP Research

July 2026

major

v1.0.0 · Jul 11, 2026

Initial release: Command Center panel with the first signal sources online.

Autopilot

Escapementv1.1.0Automationactive

Schedules and runs the Editorial Engine unattended — sourcing an idea, generating a post, and publishing it with risk-tiered review, without a human in the loop.

Cron- or on-demand scheduled runsTrending-topic sourcing from arXiv, Hacker News, and the webConfigurable Deep Research cadenceRisk-tiered auto-publish — News and Deep Research pieces always stay in draft for reviewPer-run cost estimate tracking
PowersAirNLP Research

July 2026

minor

v1.1.0 · Jul 10, 2026

Trending-topic sourcing and configurable Deep Research cadence added.

major

v1.0.0 · Jul 9, 2026

Initial release: scheduled runs with risk-tiered auto-publish.

Context Retrieval

Concordancev1.1.0Retrievalactive

AirNLP's shared vector-retrieval infrastructure — embeds and retrieves context chunks for host chat, world search, product docs, and research-corpus deduplication.

pgvector similarity search over context chunksMulti-tenant retrieval, including synthetic tenants for product docs and the research corpusPowers host chat context, world search, and duplicate-topic detection in the Editorial Engine
PowersHOWIAirNLP Research

July 2026

minor

v1.1.0 · Jul 10, 2026

Added a synthetic tenant for the research corpus, enabling duplicate-topic detection in the Editorial Engine.

December 2025

major

v1.0.0 · Dec 26, 2025

Initial release: pgvector retrieval infrastructure.

Verified Context

Plumblinev1.3.1Grounding & Safetyactive

The grounding contract behind HOWI's guest AI — classifies every answer into verified, unverified-suggestion, general-local, or unknown, and refuses to invent specifics.

Structured property and world config assembled into verified-context prompt blocksFour-tier knowledge classification enforced directly in the system promptOffline, negation-aware eval detectors for fabrication (Wi-Fi, parking, host attribution, and more)A canonical identity and story-mode layer for a consistent, first-person voiceGuest identity and returning-guest memory folded into grounded answers
PowersHOWI

July 2026

minor

v1.3.0 · Jul 7, 2026

Added the offline, negation-aware eval suite that checks for fabricated specifics (Wi-Fi, parking, host attribution, and more).

patch

v1.3.1 · Jul 7, 2026

Expanded the hard rule against explaining in-unit fixtures (blinds, thermostats, TVs) from general knowledge when they are not in verified context.

June 2026

minor

v1.2.0 · Jun 17, 2026

Verified context now folds in guest identity — guestbook entries and returning-guest memory can inform a grounded answer, still scoped and never fabricated.

May 2026

minor

v1.1.0 · May 30, 2026

Added a canonical identity and story-mode layer on top of the verified-context rules — signature beats and a consistent first-person voice, without loosening what counts as verified.

December 2025

major

v1.0.0 · Dec 26, 2025

Initial release: the verified-context prompt contract and its offline eval suite.

AI Core

Fulcrumv1.1.0AI Orchestrationactive

HOWI's guest- and host-AI orchestration engine — retrieval-first, bounded, trust-aware generation that decides how a question gets answered before any model call happens.

Multi-pass pipeline (intent, plan, answer, and an optional review/retry pass) with model routing by complexityAn internal eval harness to catch regressions before they reach guestsAsync compression of runtime facts so long conversations stay fast without losing operational detailCoordinates with Plumbline's grounding contract and Concordance's retrieval layer, without being either one itself
PowersHOWI

December 2025

minor

v1.1.0 · Dec 23, 2025

Added async compression of runtime facts (devices, stays, policies, billing) so long host-chat threads stay fast without dropping operational detail.

major

v1.0.0 · Dec 22, 2025

Initial release: multi-pass intent/plan/answer pipeline with an optional review-and-retry pass, plus an internal eval harness.