User guide — from a category to a scored matrix

Everything lives in a workspace

A workspace is one comparison: its vocabulary, its metric columns, its items (the rows), its sources, and its scoring rules. The public examples are readable by everyone; every workspace you create is private to you. Three ideas do most of the work: schema before data (you approve what gets measured before anything is fetched), source-aware values (every number carries where it came from, a confidence level, and a verified date), and review-gated AI (AI proposes values with citations; a human approves them).

The Workspaces directory with example workspaces and the New comparison button — Workspaces — the directory. New comparison starts the wizard; the examples open read-only workbenches.

Create a comparison from a sentence

Sign in, then open Workspaces → New comparison.
Describe the category: a name, the subject and boundaries (what's in, what's out), and the decision it must answer. Concrete boundaries produce sharper metrics.
Press Suggest metrics — one AI research credit, and no data is fetched: it returns an editable schema, nothing more.
Review the checklist: each metric shows unit, type, direction, importance, the weight it would carry, and how a value should be verified. Weights renormalize live over your selection.
Press Create workspace. Only now does anything persist — a private workspace with your columns and zero rows.

Wizard step 1 with name, subject and decision fields filled — Step 1 — describe the category. The example used throughout: 2026 Hybrid SUVs.

The metric checklist with weights and verification hints — Step 2 — the metric checklist. EPA fuel economy at importance 10/10 carries 20% of the score; MSRP is marked lower-is-better; every metric names its source of truth.

Unchecked metrics, suggested segments, and the nothing-is-fetched notice — Unchecked metrics stay out; segments get the same treatment; the amber panel states the contract — nothing is fetched yet.

Populate the matrix

A new workspace lands on its Metrics tab: columns exist, coverage is zero. Population is two moves — get items in, then fill each column's values. Items come from the connector catalog (connectors marked mapped declare exactly which fields they import), from any JSON API URL, or from a CSV. Values fill through a waterfall: your API source first, then AI web research over the gaps (every value cited, and applied only after you approve it), then manual entry for whatever remains.

Connector catalog with the parameterized NHTSA vPIC connector and plan usage meters — The connector catalog on the Metrics tab, plus the plan meter — the demo account has used 1 of 5 AI passes.

Metrics tab showing 24 records, per-column coverage, and the gaps banner — After a vPIC pull: 24 records, six columns at 0/24 coverage, and the banner routing each column to the right fill method.

The Fill gaps dialog with the verification hint and optional API URL — Fill gaps — restates how this metric should be verified, takes an optional API URL, and runs the waterfall on demand.

Read and work the matrix

· Coverage dots in each metric header: green = every visible row has a value, amber = partial, red = none. Hover for the exact count.
· Click any row to expand it: the score's composition plus every metric value — the same data as the cells, by construction.
· Columns picks which metrics show; sorting a hidden metric un-hides it automatically.
· Weights runs what-if scoring privately — ranks reorder for you alone.
· Everything — search, filters, sort, columns, weights — lives in the URL. Copy the address bar to share your exact view.

The workbench matrix with coverage dots, filters and side panels — The workbench (AI Model Rankings example) — coverage dots beside Input Price, Output Price, and Context Window.

A row expanded into score composition and the full metric list — Row expanded — why this model scores 70.9: normalized per-metric bars on the left, every raw value on the right.

The Columns picker with metric checkboxes — Columns — all metrics stay available; the default view shows the six heaviest-weighted.

Charts pick themselves

The chart view opens in Auto: the chart type follows the selected metric's own type. A number across items is a bar — item names aren't an ordered axis, so Auto never draws a line over them — and a yes/no or text metric becomes a count bar. Sixteen manual chart types remain one click away.

Chart toolbar with Auto mode active — Auto · Bar — the toolbar shows what Auto chose; the metric picker drives it.

Bar chart of input price across models with a tooltip — Charts draw the same filtered rows as the table (capped at 40, with a visible note when capped).

Bring a CSV or any JSON API

Your data takes a .csv/.tsv file — parsed entirely in your browser, numeric columns become scored metrics, nothing uploads until you save it as a private workspace — or a JSON API URL, importing the first 500 rows with any embedded key stripped before the source is recorded.

The Your data page with API import and CSV drop zone — Your data — the fastest path from a spreadsheet to a ranked, chartable workspace.

What the score means

Each item's score is a weighted average of its metrics, normalized 0–100 against the current cohort and adjusted for data confidence and freshness. Teal ≥ 70, amber ≥ 40, rose below — bands within this cohort, not an external standard. Two cell states matter: — means never fetched; n/a means the source was checked and had nothing. The full math, a live reproduce-any-score widget, and an honest limitations section live on the methodology page.

Questions, answered straight

QDo I need an invitation to get my own account?

Yes — public signup is off; accounts are created by invitation.

QWhat are the usage limits?

On the free plan: 5 AI research passes and 30 URL pulls per rolling 30 days. The wizard's Suggest metrics costs one pass; each Fill-gaps AI stage costs one more. The plan meter on the Metrics tab shows where you stand.

QCan the AI publish values on its own?

No. AI research returns proposals with a source URL, a supporting quote, and a confidence grade; you tick the rows you accept. A value the AI can't find stays not found — never estimated.

QWho can see my workspace?

Only you. Workspaces you create are private to your account. The public examples are readable by everyone but writable only through their review pipeline.

QA pull created columns I didn't ask for. Why?

Mapped catalog connectors import only their declared fields. An arbitrary URL mirrors whatever fields its rows carry — but every auto-created column starts at weight 0, so it can't affect any score until you promote it.

QIs the API key in my pull URL stored?

No — fetched once and discarded; the recorded source drops the query string. Scheduled pulls store the URL encrypted because they must re-run without you present.

QWhy did Auto give me a bar chart instead of a line?

A line implies an ordered axis; item names aren't one. The manual Line button still exists — Auto just won't lie for you.

QCan I share exactly what I'm looking at?

Yes — copy the URL. Search, filters, sort, visible columns, and what-if weights are all encoded in it.

QMy spreadsheet is .xlsx. Can I upload it?

Export as CSV first (File → Save As → CSV). Native .xlsx is on the roadmap; CSV and TSV parse entirely in your browser today.

QHow do I turn the floating help button off (or back on)?

Open it and choose Hide help. It stays off on this browser until you turn it back on from the footer's "Show help button" link.

From a category in your head to a scored comparison matrix