Many designs. One conviction. Feedback makes people — and A.I. — better. Data workflow is the spine.

AI coding tools' productivity claims rest on agent-side measurements — lines produced, tasks completed, time-to-PR. If the Ironies of Automation are operative — operator vigilance falling as agent reliability rises — those measurements systematically overstate net effect. There is no operator-side cockpit catching the loss.

A multi-agent kanban with a completion-block protocol that tracks per-card execution across heterogeneous AI tools. Continuous coordination-event log. Two-phase actor handoff (builder → reviewer) where the second transition requires an artifact only the reviewer can produce. Cross-tool sync via a hub SDK so an operator coordinates Cursor, Claude Code, Replit, and other agents through one board. The continuous production telemetry that runs the C1 risk-compensation field study — a pre-registered test of Bainbridge 1983 in real coding work, with hypotheses, analysis plan, and falsification criteria specified before data accumulates.

• 01Two-phase actor handoff (builder → reviewer) where the second transition requires an artifact only the reviewer can produce — enforces review without trusting it
• 02Coordination-event log as a research instrument, not just an audit trail — the apparatus for the C1 risk-compensation field study
• 03Completion blocks as a protocol — every assignment ends with structured machine-readable completion, not free-text close-out
• 04Hub-and-spoke sync between heterogeneous AI tools so an operator coordinates Cursor + Claude Code + Replit + custom agents through one board

Private. The operator-side coordination spine for the multi-app portfolio. Live, measuring, instrumented for the C1 field study.

Fantasy products are either marketplaces with shallow projections or hardcore stat tools that do not make a case. The middle — readable intelligence with a point of view — is empty.

A multi-app monorepo: a public newsroom (apps/web), a GM workflow console (gm-console), a strategy portal with cover-art game modes (apps/strategy), and a Python analytics API. Live MFL adapter, projections provider, Monte Carlo strategy engine, weekly LLM-generated stories under editorial discipline.

• 01Newsroom-as-product: weekly LLM-generated stories with editorial structure, briefs, and StoryNumbers
• 02Monte Carlo strategy engine running per-decision (fourth-down, lineup, waivers)
• 03Multi-app monorepo with formal adapter contract for swapping fantasy providers
• 04Magazine-grade cover art system for each game mode (the-pick, the-matchup, survivor, drive-duel, fourth-down-gambit)

Private. Multi-app monorepo with engine, adapters, magazine, and analytics API in active development.

Most baby-naming tools are SEO listicles. They cannot tell you whether a name is rising or fading this month, whether it sounds right next to your other kids, or what its narrative texture actually is.

A naming tool that pulls live trend data, scores phonetic and sibling fit, and surfaces narrative associations from a literary corpus. A calibration loop that learns parent preferences over the session.

• 01Live trend frequency from external data sources, not stale annual rankings
• 02Sibling-set acoustic checks (rhyme, stress, syllable count, ending overlap)
• 03Narrative association layer pulled from curated literary references
• 04Within-session preference calibration so the tool gets sharper as parents use it

Private beta. Decision tool for one of the highest-stakes irreversible decisions parents make.

Image platforms either flatten taste into engagement metrics or hide behind gatekeepers. Editorial platforms publish on calendars rather than to readers. Neither produces a reading rhythm. Neither learns from you. Vela is built on the bet that a single substrate can do both — taste-driven figurative discovery and longform editorial work — when adaptive measurement runs underneath.

385 active works pulled from museum APIs (ARTIC, Met, BnF, Smithsonian, Europeana) under full attribution and license discipline. The Reincarnation engine learns per-reader desire and pool composition across visual rhyme and emotional register. A magazine with original fiction, editorial criticism, and three load-bearing editorial axes — figurative response, emotion architecture, developmental theology — paced per-user (each reader's magazine begins when they arrive, not on a calendar). Penwright lives in /labs/penwright (F-03 Authorship Packet UI MVP shipped; F-19 Adaptive Authorship Control Kernel is the spine). Editorial Office: Writer's Desk for 1:1 with each writer; The Office for multi-writer convening with round-2 react. Stripe membership in live mode. Derivative pipeline that produces new transformative works under license.

• 01Reincarnation engine: per-user desire scoring with RID/SID adaptive measurement and visual-rhyme sequencing
• 02Per-user magazine pacing — each reader's editorial schedule begins when they arrive; positioning wedge is "your magazine begins when you do"
• 03Three editorial axes (figurative response · emotion architecture · developmental theology) coexist on a single substrate; emotion is now the core editorial axis (post the 2026-04-30 pivot)
• 04Editorial Office — Writer's Desk + multi-writer convening turns the writer roster from production tool into colleagues
• 05Adaptive-authorship substrate underneath — Vela is property #1; siblings reuse the substrate
• 06Museum-grade attribution and license discipline as a first-class feature, not a footnote

Live at vela.study. Stripe membership in live mode. Magazine publishing weekly. ~1,300 commits since project start. Penwright in early build inside the same repo; Editorial Office and per-user magazine pacing in active development.

Most AI writing tools optimize for output fluency. They make it easier to produce something faster — and that something is often shaped by the model rather than the writer. The longer-term cost (capability erosion, voice flattening, sycophancy spirals, source attribution buried) is barely measured because the field measures what is easy to measure. The result is a generation of tools that look like assistants and act like substitutes.

An authorship environment that inverts the prompt-then-edit pattern. Writers assemble Authorship Packets — intent · structure · key ideas · relevant passages · counterpositions — before the AI is invoked. Corpus selection is explicit: writers choose which sources influence the work rather than inheriting the model's training distribution. The Adaptive Authorship Control Kernel (F-19) is the spine — central registry of skill measurement, intervention, and genre-aware behavior (memoir / nonfiction / fiction never collapsed). The Penwright Measurement Framework — six skill dimensions, six derived indices, three measurement layers, five-step learning loop, and four non-negotiable failure modes — determines whether a session made the writer better. Lives inside Vela's repo (app/labs/penwright/) for now; graduates when the design stabilizes.

• 01Authorship Packet Model — replaces freeform prompting with structured input units; the structure itself is data
• 02Corpus Control Layer — writer selects sources rather than inheriting the LLM's training distribution
• 03Adaptive Authorship Control Kernel (F-19) — central registry of measurement and intervention; genre-aware behavior forks copy + schema enums + prompts + metrics rather than collapsing them
• 04Penwright Measurement Framework — first multi-dimensional measurement system for AI-augmented writing skill development; four non-negotiable failure modes (output-only optimization · over-automation · weak measurement · ignoring genre differences) act as veto
• 05Anti-invention constraint — when a structural rhetorical move requires biographical material the user has not supplied, the tool refuses to render rather than confabulating
• 06Has its own published research program at peopleanalyst.com/research/ai-human-interaction (12-paper Penwright Research Program across three tiers)

Early build inside Vela's repo (app/labs/penwright). F-03 (Authorship Packet UI MVP) shipped. F-19 (Adaptive Authorship Control Kernel) is the architectural spine; it ships first or in parallel with the first feature. 19 features (F-01..F-19) sequenced across 6 implementation waves; 79 ASNs in flight.

Most products treat performance review as a transaction record (HRIS posture) or a dashboard (BI posture). Neither matches the posture the data warrants — comparative, longitudinal, distribution-aware, small-N-honest, decision-support-shaped, behavioral. And the data carries plenty: rater bias (halo, recency, leniency, severity), distribution politics (forced ranking, ceiling compression, sandbagging-under-cap), multi-rater agreement, longitudinal career-arc shape, the motivational consequences of every rating call. Among the most poorly-instrumented data in any organization.

Early build. Two simultaneous bets. The analytical bet: a behaviorally-grounded posture — rater dynamics treated as signal not noise, multi-rater agreement (ICC, kappa) surfaced not averaged away, distribution comparison across managers and units, longitudinal career-arc shape under small-N inference (most managers rate <10 people) — produces insights neither HRIS-posture products nor BI-posture dashboards can. The architectural bet: Performix is the first vertical port of the library-core multi-property pattern outside Vela — testing whether the dual-grade research-ingest, adaptive measurement, and canonical-vocabulary discipline host a non-editorial vertical cleanly. Substance arrives as the research-ingest extraction lands in MetaFactory's packages/research-ingest/ and Performix becomes its first consumer.

• 01Behavioral posture on performance data — rater dynamics, calibration politics, and distribution shape treated as the signal, not the noise to be averaged away
• 02Multi-rater statistics surfaced where most products show only the average — ICC and rater-agreement diagnostics make compression and rater-effect bias legible
• 03Longitudinal career-arc analysis under small-N constraints — most managers rate <10 people; the stats discipline has to match
• 04First non-Vela consumer of the dual-grade research-ingest pipeline — a real port test of the library-core pattern
• 05If the substrate ports cleanly to a vertical that shares no editorial subject matter, the cross-property architecture is real; if it doesn't, the substrate is Vela-specific and the multi-property bet has to be re-thought

Early build. Repo seeded; substance gated on the research-ingest extraction landing in MetaFactory.

Cross-cutting infrastructure — book ingestion at chapter-respecting fidelity, schema-conformant extraction of behavioral constructs, job/competency/persona generation, survey factories — gets re-implemented per consumer when there's no production-factory substrate. The cost isn't only engineering: each consumer ends up with its own slightly-drifted definition of competency, persona, engagement, effect-size, and behavioral constructs that should be comparable across products become incomparable. The half-renovated state is worse than either decision: a 'Universal Information Factory' framing that tried to do too much, the resulting analytics-vs-production drift, and a multi-hundred-file cleanup-doc midden in the repo root that documents an effort that didn't land. The substrate has to be narrowed and its purpose made legible in 30 seconds.

A monorepo of production factories — book ingestion (collector / organizer / referee with statistical-quality referee gates); job-matching from extracted role-DNA; competency models extracted from organizational-psychology canon; persona factory; survey-instrument factory with item-bank-honest construct mapping; pay; business-plan; variableizer — plus core infrastructure (asset registry, storage resolver, schemas, prompts, deep-research agent). Currently undergoing the consolidation drive (ASN-1003): KEEP/DROP rule fully executed (production factories KEEP; analytics offerings DROP); the dual-grade research-ingest pipeline migrated in from Vela; the cleanup-doc midden archived; CLAUDE.md / AGENTS.md / README rewritten to actual current state.

• 01Production-factory architecture — every package outputs a behavioral artifact (competency model, persona, survey instrument, job profile, pay analysis); analytics offerings explicitly excluded
• 02Dual-grade corpus ingestion — same database holds editorially-selected curator passages and bulk research chunks, distinguished by tag
• 03Cryptographic provenance contract — SHA-256 tracked for every source file; safe-delete invariants require hash verification on backup and durable storage before any local delete
• 04~$0.13 per research-run synthesis at 30K+ passage scale — most 'AI research' tools run 10–100× more expensive because they retrieve without pattern extraction. The discipline is statistical: extract patterns once, cite the evidence, don't re-retrieve
• 05Schema-extracted measurement vocabulary (@measurement/core) shared across consumers — constructs, items, instruments, effect-sizes defined once in canonical form, so behavioral measurement compares cleanly across Performix / Principia / PA Platform / the toolbox. Cross-product comparison becomes structurally possible rather than aspirational

Private. Five-track consolidation in flight (audit · pruning · production-factory finalization · research-ingest migration · consumer rewiring). Eight sub-ASNs (ASN-1004..ASN-1011) under the parent (ASN-1003).

HR analytics products treat behavioral data as if it were transactional. Engagement gets reduced to a survey score; performance to a rating average; retention to a churn rate; compensation to a band. The richer questions — what actually drives engagement in this organization, what signal in performance distributions matters for decisions, what statistical posture handles small-N segmentations honestly, what value would more information have before another study runs — get lost. Cross-cutting concerns (anonymization, metric calculation, segmentation, survey delivery, decision support) also get re-implemented per product, brittle and fragmented. The combination — wrong analytical posture and fragmented infrastructure — is the field's default state.

Seven live spokes — reincarnation (adaptive psychometric diagnostic engine; IRT-weighted item selection; pool-based item lifecycle), preference-modeler (Likert/multi-choice/free-text plus MaxDiff, conjoint, penny allocation, paired comparison; BIBD-balanced task generation; MNL utility estimation via Newton-Raphson), data-anonymizer (PII detection, deterministic HMAC tokenization, k-anonymity min-N gate, substitution-strategy registry), segmentation-studio (HRIS canonical-field normalization with 35-field priority catalog; multi-membership cohort resolution; OneModel adapter; recipes; pack publishing), calculus (statistical enrichment, anomaly detection, time-series imputation, metric × segment × period combinatorial factory; auto-selects Wilson / t-interval / normal CI), anycomp (comp models, market band math, stateless evaluation, auditable cycle runs), forecasting (Monte Carlo simulation, EVPI, discrete EVSI on aligned-chance decision trees) — plus a reserved namespace job-family-agent whose canonical home is meta-factory-prod. Each spoke owns its own schema, contract, and audit trail. 51 MCP tools, 48 HTTP routes, hand-authored tool descriptions with whenToUse / whenNotToUse disambiguation. Per-route structured logs plus per-tool audit rows in mcp.mcp_audit. Consumer apps vendor the typed contracts; the algorithms live here.

• 01Substrate, not product. Consumer apps (Performix, vela, future analytical products) vendor only the typed Zod contracts; the algorithms stay in the toolbox. Adoption is spoke-by-spoke; no all-or-nothing migration.
• 02AI-native by construction. Every spoke is callable from AI agents over MCP (Model Context Protocol) without bespoke integration. Per-consumer auth, scope-restricted keys, fire-and-forget audit log. Performix migrated to MCP transport 2026-05-11 as the first external consumer; devplane operates the wildcard key.
• 03Behavioral science in the algorithms, not bolted on. Reincarnation runs IRT a-parameter-weighted adaptive item selection with Cronbach α tracking. Preference-modeler runs MNL utility estimation on real MaxDiff/conjoint designs. Calculus auto-selects the right confidence-interval method by data shape. These are textbook psychometrics and choice theory implemented as service APIs.
• 04Privacy is a service, not a setting. Data-anonymizer is cross-cutting — every spoke that surfaces team-level rollups calls min-N-check before responding. Anonymity-gated aggregations return blocked status below the floor; tokenization is deterministic and cache-backed; substitution strategies (mask / pseudonymize / synthetic-realistic) live in a registry.
• 05Systems × survey × behavioral-science join is first-class. Segmentation-studio normalizes HRIS canonical fields; data-anonymizer makes the join safe under k-anonymity; calculus enriches the joined records into MetricEnvelope objects. The same envelope shape carries data from a Workday extract, a survey response, or a derived rollup — consumers don't care which.
• 06Explicit contract versioning. Every spoke ships CONTRACT_VERSION; every additive change is a semver bump; every breaking change is a major bump with affected-consumer notes. Consumers vendor a copy and re-vendor on major bumps. The deploy boundary is clean.

Seven live spokes plus one reserved namespace. 51 MCP tools (3 toolbox-discovery + 48 spoke); 48 HTTP routes. All health endpoints green; mcp.mcp_audit writing real rows; database migrations auto-run on Vercel production builds (DP-134). First external consumer (Performix) migrated to MCP transport 2026-05-11. Solo build.