Corpus in. Canonical data out.
CanonicAI turns unstructured knowledge — books, research papers, domain documents — into canonical, queryable datasets. At scale, with provenance.
Commission a canon →“Anyone can write a prompt. The defensible thing is running thousands of multi-step extractions reliably, idempotently, and with lineage — at a cost you control. That’s why factory is the honest metaphor: a production line with QA and inventory control, not a clever prompt.”
The production lines
Book Factory
Whole books deconstructed at chapter-respecting fidelity into tagged summaries, argument models, and factor structures — not naive chunks.
Article Factory
Peer-reviewed research distilled into instruments, constructs, citations, and effect data — feeding a living evidence engine.
Compendium Factory
Reference catalogs of validated measurement scales extracted item-by-item — validated against known ground truth at 95% recall.
Schema Authority
One canonical measurement vocabulary — constructs, items, instruments, effect sizes — defined once, conformed to by every consumer.
The line, running
Canon catalog
Book Elements
Summaries, narrative structures, constructs, relationships, and source-grounded models across more than 4,900 works.
BuiltResearch & Evidence
Research assets resolved into studies, citations, effects, constructs, and reusable evidence records.
BuiltInstrument Canon
Measurement instruments, scales, items, response options, scoring notes, and construct associations.
BuiltJob Elements
Occupations, tasks, knowledge, skills, abilities, activities, styles, and interests grounded in O*NET.
BuiltMetric Canon
Operational and scientific measures with definitions, entities, formulas, provenance, and deduplicated identities.
BuiltRegulation & Statistics
Wage, labour-regulation, and labour-statistics records normalized for comparison and downstream computation.
BuiltCompetency Elements
Competency statements and classifications transformed into addressable, joinable elements.
BuiltTaxonomy & Routing
Controlled vocabularies and routing canons that place heterogeneous source material into one navigable system.
BuiltFrames — shared coordinate systems
A canon can be more than a list. A Frame gives every record an address in a shared coordinate system, so applications and agents can join, compare, translate, and compose data without bespoke integrations.
What a record looks like
The product is not a chat transcript. It is durable, joinable records with stable IDs, aliases, provenance, and typed relationships — the kind of object an application can query a year later.
{
"canonical_id": "business_strategy_alignment",
"name": "Business/Strategy Alignment of Analytics",
"tier": "core",
"four_s": "strategy",
"aliases": ["Business Priority Alignment", "Problem Framing…"],
"relationship": {
"from": "business_strategy_alignment",
"to": "analytics_capability",
"type": "enables",
"support": 0.14,
"provenance": ["van_vulpen", "fundamentals_of_hr_analytics"]
}
}
“Organizations struggling with unclear strategies can use the Diamond Model to define desired outcomes and develop corresponding competencies and capabilities.”
{
"problem": "unclear_strategy",
"solution": "diamond_model",
"book": "rewarding_excellence",
"chapter": "ch01",
"source_kind": "problems_addressed"
}
Taxonomy as address space
Heterogeneous sources only become a catalog when every item receives a stable address. JobFrame alone carries 14,948 occupational cells; the guide taxonomy classifies thousands of works into navigable tracks.
jf_outpatient · clinic and ambulatory care
jf_long_term_care · extended residence and memory care
jf_in_home_care · home nursing and hospice
How a problem gets resolved
Problems already hide in the corpus as attributes on tools, brandscripts, deep extracts, and cases. Making them first-class is harder than it sounds: statements conflict, synonyms proliferate, and every resolution has to carry provenance or it cannot be trusted.
Scatter
Problems live as strings on tools, frameworks, brandscripts, deep extracts, and cases — never as records. Useful locally; unjoinable globally.
Harvest
Deterministic lift with full provenance. One cluster prove-one already produced 514 raw statements across 9 books — 435 from chapter problems_addressed alone — at $0.
Resolve
Cluster and dedupe into canonical Problem Cards. Measure rater agreement; withhold low-agreement merges. A synonym map without evidence is fiction.
Join
Wire each card to solutions, models, personas, jobs, and stories. The differentiator is diagnosis first, then better-fitting solutions grounded in evidence.
Surface
Browse-by-problem, wizards that diagnose before they recommend, guides whose Orient movement opens on a real Problem Card — not a marketing slogan.
Wide nets from one canon
Once the factory produces a guide, model, or occupational cell, many surfaces can publish from it. Same records. Different fronts. That is the point of a canon.
People analytics — destination guide
Full depth on the PeopleAnalyst surface: constructs, playbook, measurement joins.
Same subject — Bicycle front
The travel-guide rendering of the same corpus model: route, stops, honest disagreement.
Occupational spine — JobFrame
14,948 cells as a shared address bus for careers, compensation, and capability joins.
Domain home — compensation
CompensationProfessional consumes the same producer surface for a specialist audience.
Reference library
The shared book corpus as a navigable library — identity and status, not a second catalog.
Analytics that need a canon
A single PDF or a vector store can answer a question. A canon answers questions that cross books, instruments, jobs, and evidence — questions that are otherwise impossible without months of hand joins.
Construct coverage across a field
Which capabilities appear as core across dozens of people-analytics books, which are contested, and which edges are weakly supported? That is a graph query over stable construct IDs — not a keyword search.
Problem → solution inversion
Start from “unclear strategy / weak incentive design” and retrieve every framework that claims to address it, with chapter provenance. One cluster already yields 435 such joins before resolution.
Instrument × construct × effect
Which validated scales measure which constructs, and what effect sizes appear where? Evidence products need item-level instrument canons joined to construct space — not abstracts glued to PDFs.
Occupation × capability joins
Map a capability model onto JobFrame cells to ask which roles need which constructs, or which guides should bind to which jobs. Impossible without a shared occupational address bus.
Sellability from measured depth
Price and gate guides from counted paid bodies, not vibes. Empty rooms and free projections become operational signals the factory can act on.
Trust Scale — faith and confidence in peers and management
“If I got into difficulties at work I know my workmates would try and help me out.”
“Management at work seems to do an efficient job.”
“Our management would be quite prepared to gain advantage by deceiving the workers.” (reverse-scored)
Commission a canon
Prove one unit before you fund the line.
Bring a standards library, scientific literature set, technical-documentation corpus, historical archive, or proprietary knowledge base. We scope one representative unit, run it end to end, show the canonical output and provenance trail, and measure its real production cost before proposing scale.
Tell us about your material.
Your pilot brief.
This is the starting point for the scoping conversation — it's included in the email you send below.
- Material fit and acquisition path
- Per-unit production cost and timeline
- Delivery format for the canonical output
Who should we confirm the call with?
Your pilot brief is ready.
Send the brief below, then book the 30-minute scoping conversation.
Glass box, not black box
Every dataset CanonicAI ships traces to its source — file hashes, extraction lineage, model and prompt provenance, idempotent re-derivation. If a number is in the output, you can walk it back to the page it came from.
The engine is the producer and source-of-truth; everything downstream is a consumer. It powers the PeopleAnalyst family: