Week 08 / 2026 — The first vocabulary

16 February – 22 February 2026 · 101 commits

The first full week of the repo. The bootstrap weekend (W07) left a medallion architecture and a dbt project. This week named the things that live on top of it: probe, assessment, hypothesis, lineage, calibration, Explorer. Five nouns the project will use for the next eight weeks — and then deliberately retire when Sense 14 reshapes the entire taxonomy. But for now, this week, these are the words.

By Sunday afternoon the project has its own SvelteKit app, a YAML-driven diagnostic engine, a hypothesis layer above it, a calibration harness below it, three taxonomies (ATC, MIGEL, BAG), and a Tauri desktop build. The 72-hour sprint that produces the original sprint report starts the same Sunday. None of this exists yet on Monday morning.

Tuesday: the probe DSL

53ec472a on Tuesday Feb 17: “Add probe DSL framework v1: YAML-defined diagnostics compiling to dbt SQL.” This is the load-bearing commit of the week, possibly of the first month. The pattern — declare diagnostics in YAML, compile to dbt, ship with a registry — is the pattern that becomes Signals (Sense 14), then Theses (W09’s Diagnosis), then Verdicts (W09’s diagnosis), then the four-layer analytical pyramid that defines the product.

Today there is one type (balance) and ten declared probes. The compiler is forty lines. The registry is a flat table.

By Sunday the DSL has six types (4fdd0c0c, b7bfb19b), data-layer awareness (86d72ae7), assessment as a type (50722f2f), Silver-layer support (41be30bb), severity rules with money-at-risk awareness (e35c438b), and 11 V4.0 probes covering the long tail (60e85e28). One commit per discovery; the type system grows ahead of the implementation.

What this commit makes possible: every later signal/perspective/thesis/verdict is a descendant of the V1 probe DSL. The taxonomy renames; the shape doesn’t.

Thursday: the Explorer is born

6001f8b8 on Thursday Feb 19: “Add nuMetrix Explorer: customer-facing SvelteKit app.” The Evidence dashboard from the bootstrap weekend is good for static reports; what’s wanted is something navigable — click an entity, see its facts, see the probes that fired against it, see the hypotheses they support.

Within 48 hours:

• b6aec430 — hero header, year selector, billing stats
• 99818cdc — “Replace misleading total money-at-risk with per-probe breakdown”. The earliest version of the no false billions discipline that will dominate April.
• d508c859, 4da5c191 — fix date display, fix [object Object] rendering. (Every new SvelteKit app passes through these two bugs.)
• 925bb78f — “Fix explorer DuckDB not actually opening read-only”. The first version of the read-only-DuckDB guarantee that the explorer leans on for the next 90 days.
• 79485b9d — “Replace static DIMENSIONS array with data-driven schema discovery”. The Explorer stops knowing about specific entities; it discovers them. The same discipline that becomes the entity registry in W09.

The Explorer that ships to docs.jinflow.io’s gallery today is a direct lineal descendant of this Thursday commit. The page shapes, the navigation idioms, the metadata patterns — they all started this week.

Saturday: the hypothesis arrives

ad0524bd on Saturday Feb 21: “Add hypothesis system — business questions evaluated against probe evidence.” The second layer of what would become the analytics pyramid lands. A hypothesis is a YAML-declared business question; an evidence chain references one or more probes; a verdict (confirmed / plausible / not_observed / insufficient) falls out of the math.

Eight commits the same day flesh out the layer: 45fcbb8a (weighted assessments + graduated scoring), 4fb18d94 (phantom-cases hypothesis as the first business question, verdict badges), 35309cac (about page + explain-like documentation).

The naming will change in eight weeks — hypothesis becomes thesis, diagnosis becomes verdict, probe becomes signal. But the shape stays. YAML declares the question; the compiler emits the SQL; the platform aggregates the answer.

The calibration harness

df8aae0a on Saturday: “Add calibration harness — defect manifest, calibrate.py, Explorer integration.” The discipline of teaching the synthetic data what failure shapes to inject, then verifying the probes catch them. The calibration system is what makes the diagnostic pyramid testable — without it, “the probe found a billing duplicate” is just a claim; with it, “the probe found 14 of 14 injected duplicates” is a measurement.

This week-end calibration commit doesn’t get a Sense paper. But the discipline it encodes — the system can be audited against itself — becomes the lineage panels, the Ledger (Sense 18), and the Simulation (Sense 19). All of them are answers to “can we trust what this thing tells us?”

Taxonomies become first-class

Wednesday-Thursday: three reference taxonomies arrive.

• 99b28a3e — ATC (drug classifications). Loaded as dbt seed, queryable.
• f242fcbf — MIGEL with per-tenant BAG reference. The Swiss medical devices catalog.
• b4e64827 — taxonomy tree/table view toggle, search filtering, green color tokens. The UI for navigating them.

Plus de7b856d — “taxonomy cardinality cross-check — dbt model + Explorer page” — and eb4fcff7 (force-directed graph view), c3c0267d (member-level drill-down), 6c2ee90f (clickable fanout bars as filters). The taxonomy page goes from “list of codes” to “interactive analytical surface” in three days.

Tauri desktop

9256af58 on Saturday: “Package Explorer as desktop app with Tauri v2.” The Explorer as a downloadable .app/.exe/.deb starts here. Three follow-ups fix the startup flow: ba6eb9da (DuckDB binding patch), 26a8dcf5 (track app-icon for fresh clones), ddace99e (prerequisite checks for sidecar and icons).

The full-binary distribution story that consumes most of W11 — embedded Python, embedded dbt, cross-platform builds — starts here as just “can we package the SvelteKit app as a desktop app?” with the answer “yes, plus three startup bugs.”

The lineage layer

f345348f on Saturday: “Add self-documentation lineage layer — YAML definitions, compiler, runtime metrics.” The first version of what becomes the four-level lineage explainer in W12 (probe detail → source records → CSV provenance → AI-enhanced narrative). The framing — every artifact knows where it came from, and that knowledge is queryable — is committed before the artifact registry exists to formalize it.

The 72-hour sprint begins

Sunday Feb 22 is day one of the 72-hour sprint that produces the Feb sprint report. The Sunday commits read as a sprint kickoff:

• 8a3df111 — update todo with probe-config and PDF-report items
• cb14ce6d — “Documentation overhaul — update narrative docs, add index and looping presentation”
• 8e7ed3d6 — “assessment design improvements — multi-field/compound severity rules, health score, composability”

The sprint extends into the following week (Mon-Wed Feb 23-25, the start of W09), and produces the Diagnosis layer that W09’s “The roof goes on” lede sentence celebrates. Today, Sunday Feb 22, the project has probe → assessment → hypothesis. By Wednesday it will have probe → assessment → hypothesis → diagnosis. The four-layer analytical pyramid is one sprint away from being whole.

Other notable

• 3da4da01 — “Remove source_system as required CLI parameter — auto-resolved from tenant.yml”. The first move away from explicit CLI flags toward declarative configuration.
• a6198e21 — “DuckDB → Mermaid ER diagram generator”. The dynamic ER diagram that becomes a docs-site asset in W12.
• 49cb67b5 — Evidence dashboard extension + cloud deployment concept. The first cloud paper.
• bfe53b65 — “Add what_this_is.md — high-level vision and approach document”. The earliest version of what becomes the What is this? tab in the Explorer.
• 4db4f426 — “Document all 68 dbt models with descriptions and column definitions”. Discipline week.
• c8f0d593 — Dagster orchestration extended for hospital_delta and multi-tenant. The orchestration story that the operating-modes paper retrofits in April.

Numbers

• 101 commits, 7 days
• 6 probe types declared in the DSL (balance, mandatory_item, distribution_outlier, duplicate, ratio, assessment)
• 11 probes shipping by Sunday
• 1 hypothesis (phantom cases) confirmed
• 1 Explorer app born
• 1 Tauri desktop build
• 3 reference taxonomies (ATC, MIGEL, BAG)
• 1 calibration harness with defect manifest
• 68 dbt models, all documented
• 0 Senses (the pattern is 8 weeks away)
• 0 acronyms (AFS, KLS, SIS — 4 weeks out)

What this week made possible

The probe-DSL → hypothesis → verdict shape that all later vocabulary renames preserve. The Explorer as a customer-facing analytical surface. The Tauri distribution story. The calibration discipline. The data-driven schema discovery in the Explorer. The taxonomy navigation idioms. The cloud and operating-modes thinking. The auditable lineage.

If W07 was “what’s the directory tree?” — W08 is “what are the nouns?” — the answer is probe, assessment, hypothesis, lineage, calibration, taxonomy, Explorer. Three of these get renamed in W14. None of them get retired.

By Sunday evening, the Diagnosis layer is one work-night away. “The roof goes on” will be its lede sentence.

The roof goes on next week.