April 2026 — The month of the Senses

1 April – 30 April 2026 · 795 commits · v0.31 → v0.46

If March was the month jinflow learned its vocabulary, April is the month it learned to think out loud. In thirty days, eleven new design papers landed in docs/design/ — Senses 11 through 22, with their corresponding implementations either shipped or scaffolded. The project’s habit of “name what’s load-bearing before building it” stopped being a coincidence and became a discipline. By month’s end, the Sense numbering was no longer a private joke — it was the index by which the architecture argued with itself.

Seven hundred and ninety-five commits. Smaller than March in volume; larger in conceptual range.

The Signal paradigm

The decisive shift this month landed at the end of the first week — on a Friday afternoon in W14, the same way March’s vocabulary moment had landed on a Tuesday. Sense 14 — The Signal (64c4c481) argued that the word probe implied passive measurement, while what the engine actually shipped was opinionated observation with polarity, score, direction, impact. Four root nouns of the analytical pyramid — Probe, Assessment, Hypothesis, Diagnosis — got renamed to Signal, Perspective, Thesis, Verdict.

The rename was not gradual. Across four commits over 36 hours (92e78c91, 764e59fc, 9afb00fb, 6f89a15b, 1a0164f6), the engine, the Explorer, the schemas, the column names, the i18n keys, and 102 documentation pages in three languages all converged on the new vocabulary. The migration script ran. The legacy term was retired. Old YAMLs got auto-migrated by a one-time script that was then archived.

The clean-up took two more weeks — Phase 1.5 in W15 caught 13 silent SQL/URL/filter bugs that the old safeQuery wrappers had been silently swallowing. The CI guard (c40cbd32, 7a78a3e3) ensures no PR can reintroduce the legacy vocabulary. The display rename (SMEbit → Subject Matter, BitBundle → Dossier, a4a7ea7d) followed the same week — display-only, internal slugs preserved for git history.

This is the cleanest no-backwards-compat rename of a 4,000-LOC vocabulary the project has ever attempted. “Reduce to the max” taken seriously.

The Senses cascade

If you list the design papers in order of commit date, the cadence is visible:

Sense	Name	Date
10	Computed Aggregates	Apr 2
11	The Veil	Apr 2
12	The Mirror	Apr 2
13	The Canvas	Apr 4
14	The Signal	Apr 4
15	Observation, Validation, Explanation, Contributing Factor	Apr 5
16	P2P2P	Apr 7
17	The Cathedral	Apr 16
18	The Ledger	Apr 16
19	The Simulation	Apr 17
20	The Seam	Apr 26
21	The Heartbeat	Apr 27
22	The Legend	Apr 27

Thirteen senses across a month. Each one a load-bearing argument with a name. Each name doing the work of two paragraphs of prose. The pattern: if you cannot name it, you cannot build it. The names came first; the implementation followed.

Implementation in step

The Senses are not pure architecture. The same month each was named, the corresponding implementation either shipped or scaffolded:

• Sense 10 (Computed aggregates) — declarative SQL expressions for group-by (11e4c0c3) shipped the same day
• Sense 13 (Canvas) — MapPanel extracted as first Canvas-ready panel (a64a4cfe) the day of the paper
• Sense 14 (Signal) — full vocabulary rename across the codebase within 36 hours
• Sense 16 (P2P2P) — Cloudflare Tunnel + stealth mode + offline banner shipped in W15 (6350e455)
• Sense 17 (Cathedral) — guided 3D tour replaced FPS sandbox (bdac5ddd) in W16
• Sense 22 (Legend) — schema, compiler, registries, severity badges, autocomplete — six phases in six days at end of month

The Senses cadence is not theatrical. It’s a working method: write the paper that makes the implementation obvious, then write the implementation.

The DuckDB migration

One quiet incident defined the middle of the month. Playworks (the new black-box test harness, 6a3f0d7f) was crashing with SIGBUS on macOS arm64 after the fourth or fifth distinct table query. Root cause: duckdb-async@1.4.2 had a native handle leak that accumulated per table access; two DuckDB instances trying to mmap the same SIS file collided.

The migration to @duckdb/node-api (834d6014) and the SIS-ATTACH workaround (f504ee7b) ended a class of bugs that had been intermittent for a month. “Don’t ATTACH SIS to KLS Store — eliminates SIGBUS on macOS arm64” is the kind of fix that only happens after the engineering muscle for diagnosing native-binding leaks has been built.

The monolith drain

jinflow/cli.py was 12,062 LOC at the start of April. It dropped to ~340 by month’s end — a 97% reduction.

The drain ran in phases:

• Phase 4 (W16, 70e9dd0f and others) — in-process compiler --check via runpy
• Phase 5a–5p (W16, 16 commits) — extract command groups into focused modules: info, config, status, sources, launch, sis, afs, bootstrap, update/reset/purge, r2, extract, auth+packs+snapshot, generate+scripts+notebook_pdf, misc
• Phase 6 (396c98d6) — 54-command --help smoke test
• Continued in W17 — SIS publish, extraction write path, step runner, KLS baking

The pattern: extract one concern, leave a re-export shim, verify the test suite. By month’s end, ~92% of the monolith was drained. Each extracted module was tested in isolation, none of them broke the frozen-binary path.

This is the kind of refactor that doesn’t show up as a feature. It shows up as the next four hundred commits feeling fast.

The Senses’ practical reach

The big architectural conviction shipped this month: declarative ahead of imperative, everywhere.

• Canvas-style panels (Sense 13) — MapPanel, LabPanel, StudioPanel — each decoupled from entity, driven by config + URL state. By W15, named instances (6a943a9b, 1aa82e90, 835b37dd) let tenants declare their own Lab/Salon/Studio surfaces in config.yml.
• Report engine — kind field + metric/table/chart cell types (dd5055c4 and four more steps) — notes became both narrative and report from one engine.
• Classifications + Legend (Sense 22) — classifications: block in entity YAMLs, normalized through one shared helper, baked into registries, read by Explorer. Icons, colors, descriptions, actions — all declared.
• Diagnostics (W17–W18) — codes, severity registry, --allow=W003/--deny=W014/-Werror flags, JSON summary.
• UoM (Sense 14.2 prefigured, foundation in W18) — UCUM reference, three-layer cascade (engine/pack/tenant), uomcheck validator.

The trajectory of the month: every place a Svelte file used to make a decision, a YAML file now does — and the engine reads it.

Sense 20 — The Seam

The month’s most consequential rename was probably the smallest one. faf8470e — “retire pre-make; rename jin extract → jin inspect (read-only).” Followed by the Three Generations (e405b281) file system convention: gen00_raw → gen01_canonical → gen02_enriched.

Why it matters: make is a pure function. AFS in, KLS out. No network, no SIS publish-back, no commits during make. Side effects belong to separate, explicit operations — jin inspect (read-only verification), jin afs push/pull (network), jin sis flush (SIS publish). The Heartbeat (Sense 21) and Engine’s Seam (Sense 23) papers from the following two weeks all assume this purity.

This was the architectural conviction the project had been edging toward since W11’s atomic-KLS-build commit. It took until W17 to name it. From here on, the make pipeline is auditable in a way it had never been.

The arc of the month

If you read April as a single sentence: the project’s design language became its working method. The Senses sequence stopped being annotations on a codebase and started being a roadmap. Each Sense named a load-bearing argument; the next set of commits made the argument true.

There’s a moment in March — late, around the docs.jinflow.io launch — where someone other than the author runs the system, and the project becomes shippable infrastructure. There’s a moment in April — late, around Sense 22 — where the project becomes self-explanatory infrastructure. By month’s end, any classification value in any signal/verdict/smebit registry knows what it means, why it’s there, what to do about it. That sentence would not have been writable on April 1.

Numbers across the month

Metric	Value
Commits	795
Releases	~50 (0.31.14 → 0.46.2)
New design papers (Senses 10–22)	13
Implementation paired with paper, same day or week	9 of 13
Vocabulary renames	1 Signal paradigm (Probe→Signal etc.) + 1 Subject Matter display rename + 1 Pre-make retirement + 1 Three Generations naming
Silent SQL bugs caught after rename	13
cli/init.py size reduction	12,062 → ~340 LOC (~97%)
dependencies migrated	duckdb-async → @duckdb/node-api
Senses ratio (design : implementation)	1:1, mostly the same week

The mood

If March’s commit messages read like someone discovering what they were building, April’s read like someone who knows. “The dessert.” “Reduce to the max.” “Names the third party Sense 20 didn’t notice: the engine.” “Analyst has the Canvas. CFO has Observations. The estate operator has been working in the hallway.” The prose around the work this month is the prose of a project that has earned its vocabulary.

A month of senses, in plain sight.