The previous 28 articles explained how to use silverpoint. This one, the last of the series, tells how it was made. There are no recollections in it: every fact comes from the repository, its specifications, its git history or its CI. The numbers are from 1 November 2026, the day of release 0.9.2.

The numbers

WhatHow much
From the first commit to 0.9.2From 12 September to 1 November 2026
Commits up to 1 November250
Versions published on npm10, from 0.1.0 to 0.9.2
Packages, all on the same version9
Requirements in the PRD369: 343 MUST, 23 SHOULD, 3 COULD
Technical design decisions60, from DD-001 to DD-060
Diagnostic codes29, from SP001 to SP029
Test files237
Files under fixtures/8,708
Lines of the core, @silverpoint/coreAbout 20,800
Lines of each adapter3,907 in React, 5,065 in Vue, 5,894 in Angular

The lines count the source files without the tests, comments included. The gap between the core and the adapters is no accident: it is the rule that weighs most in the project, and it is further down.

Specification first

The repository starts with specs/, not with the code. The context file for agents says so in its first line: before proposing or writing code, read the constitution.

The methodology is what the repository itself calls Spec-Driven Design: the specification is the standing truth, and the code implements it. There are seven documents:

DocumentWhat it settles
constitution.mdNine articles that bind everything else
prd.mdWhat the library does, as requirements with a REQ-NNN identifier
api-spec.mdThe public contract: packages, props, events, diagnostics
technical-design.mdHow it is built: the 60 DD-NNN decisions
data-model.mdThe shape of each chart's data and its invariants
implementation-plan.mdThe phases, from 0 to 10
tasks.mdThe tasks, each with the requirement it fulfils

The constitution was ratified on 12 September, the day of the first commit. Its nine articles are the ideas this series kept showing: the geometry of the data is exact (1), there is a single geometry engine (2), the four adapters draw the same thing (3), the render is deterministic (4), there is a precision mode and an accessible alternative (5), tone is made with hatch density (6), grounds are data (7), no CSS framework is imposed (8) and the specification is the source (9).

The requirements in EARS

Every requirement in the PRD is written in EARS (Easy Approach to Requirements Syntax): one template per kind of requirement. Of the 309 that are not rows of the chart catalog, 189 are ubiquitous (always in force), 48 unwanted-behaviour, 45 event-driven, 23 optional and 4 state-driven.

Article 28 showed an SP002 warning labelled REQ-008. This is the requirement, of the unwanted-behaviour kind:

IF a datum contains null, undefined or NaN, THEN THE SYSTEM SHALL omit that point from the stroke, honour the declared connectNulls policy and warn only in development builds.

And this one, ubiquitous, is REQ-124, the one that keeps hatching from being the only way to read a datum:

No chart SHALL encode information solely by hatch style.

Phases 1 to 4 each closed with a chart-by-chart review of REQ-124, kept in changes/. The Phase 2 one lists every visual channel of the fourteen charts and the datum it carries, and every finding has a test that draws the chart in precision, with no hatching, and reads the datum from another channel.

Analyze: reviewing the specification before the code

specs/analyze.md is a read-only cross-check of the seven documents. Its fourth run, on 13 September, counted 105 requirements, no identifier cited without a definition and no task without a requirement. And it writes the rule down: “Analyze presents findings; approval and correction are human decisions.”

That run also records a change of scope before any code was written. The third run had been triggered by a misstated question: the framework to add was Vue, not Vite. The specification absorbed both changes in a row, Vite as a validated integration and Vue as a third framework.

Changes: deltas and features

Once approved, the specification is not edited by hand. A change goes into changes/ and is folded into specs/ when it is approved. There are two sizes:

  • Deltas, 15, for changes to something already specified. For instance, delta-007 widened SP002 to any value a chart cannot draw as given, and delta-015 changed how the agent catalog pins the Reflex version.
  • Features, 6, for new work: the dashboard, the UI components, the second chart catalog, the composed components, agent access and decks.

Every feature goes through approval gates. The decks one, feature-006-deck, has this sequence:

GateDocumentWhat it decides
0research.mdWhat exists to build on; Marpit measured; risks
1prd-delta.mdThe what: requirements REQ-570 to REQ-628
2api-delta.mdThe contract: @silverpoint/marp, SpDeck in the three adapters, directives, fences
PoCpoc.mdHow to scale the slides in Chromium, Firefox and WebKit
3technical-design-delta.md and data-model-delta.mdThe how: DD-051 to DD-060
4plan-and-tasks.mdTasks T-252 to T-289

All of them were approved on 31 October, the same day Phase 10 was implemented and folded into specs/. That day has 51 commits. 0.9.0 shipped the next day.

The architecture decisions

Four decisions explain almost everything the series showed.

An engine without a DOM. All the geometry is computed in @silverpoint/core, which knows neither the DOM nor any framework. An adapter translates what the core gives into its framework's nodes and nothing more: it computes no scales, axes, arcs or paths. CI fails if an adapter imports d3-* or roughjs. That is why the core has about 20,800 lines and each adapter between 3,900 and 5,900. Article 2 of the constitution gives the reason: two implementations of the same drawing always end up diverging.

Grounds are data. A ground is a set of tokens: substrate, inks, inking parameters, typography and tonal mechanism. Adding one touches the code of no chart. cyanotype, the second, arrived in 0.2.0 with a tonal mechanism of its own: line weight instead of hatching.

The adapters are compared with a canonical render. Each adapter is compared with the canonical serialisation stored with each test case, not with the other adapters. Adding a framework costs one comparison, not one per pair.

Determinism. There is no Math.random() and no Date.now() on the render path. The hand-inked stroke comes from a public seed, derived from the id. Without that there would be no visual comparison at all, and no server render matching the browser's.

The CI gates

Every push to develop and every PR goes through the CI workflow. The release runs all of it again before uploading anything to npm. The main gates:

GateWhat it checksRequirements
ContrastEvery ink of every ground against every substrateREQ-126, REQ-127
SVG stringEvery adapter against the canonical render of every case, with zero toleranceREQ-180
PixelsThree image comparisons, with fixed thresholds, in a browser pinned by digestREQ-181
WeightEvery package within its size budgetREQ-164
Agent catalogIt regenerates with no diff and matches what the adapters exportREQ-535, REQ-536
Generated codeThe reference files of generate_code compile in the example appsREQ-557
TraceabilityEvery MUST requirement has at least one test that cites itREQ-183, REQ-184

The pixel gate has three thresholds, written in the constitution: against the canonical render of the same commit, against the stored golden image, and a ceiling for any single pixel. The comparisons only hold with a pinned browser, scale 1, animations off and the fonts loaded.

The traceability report is generated from 237 test files: 342 of the 343 MUST have a test, and the missing one, REQ-564, is deferred to another repository. None blocks.

Path weight has its own report: a budget of 40,000 bytes of path data per chart.

The release history

Commits per silverpoint release: 82 up to 0.1.1, 39 in 0.2.0, 15 in 0.3.0, 12 in 0.6.0, 25 in 0.7.0, 16 in 0.8.0, 53 in 0.9.0, 4 in 0.9.1 and 3 in 0.9.2

The chart was drawn by the render_chart tool from article 28, with the commits between each tag and the previous one. The call:

{"chart":"bar-chart","props":{"xKey":"release","valueKey":"commits","title":"Commits per release","footerLeft":"2026-09-12 to 2026-11-01","footerRight":"249 commits"},"data":[{"release":"0.1.1","commits":82},{"release":"0.2.0","commits":39},{"release":"0.3.0","commits":15},{"release":"0.6.0","commits":12},{"release":"0.7.0","commits":25},{"release":"0.8.0","commits":16},{"release":"0.9.0","commits":53},{"release":"0.9.1","commits":4},{"release":"0.9.2","commits":3}],"width":720,"height":280,"id":"commits","embedFonts":true}

The releases, with their npm dates. The last two came out on 15 November, after the numbers above:

VersionDateWhat it brought
0.1.025 SeptemberThe first catalog, 33 charts, in React, Vue and Angular
0.1.125 SeptemberA README per package; from here on, published from CI only
0.2.026 SeptemberThe Dashboard, the cyanotype ground and @silverpoint/tailwind
0.3.01 OctoberThe 17 UI components
0.6.010 OctoberThe second catalog: 27 more charts
0.7.020 OctoberPalettes, composed components, overlays and the timeline
0.8.021 OctoberAgent access: @silverpoint/mcp, the catalog and the skill
0.9.01 NovemberDecks: @silverpoint/marp and SpDeck
0.9.11 NovemberReact SpDeck under <StrictMode> (issues #7 and #8)
0.9.21 NovemberThe Reflex catalog with sp_deck (issue #10)
0.10.015 NovemberThe bugs this series turned up (issues #12, #14 and #17 to #22)
0.10.115 NovemberThe Reflex catalog of reflex-silverpoint-react 0.10.0

There was no 0.4.0 and no 0.5.0. The second catalog was built in three waves, and the decision was to publish them together as 0.6.0. 1.0.0 is parked on purpose: it would put the API specification in force as a stable contract. Its changeset waits in changes/release-1.0.0-changeset.md.

Since 0.1.1, releases are published by CI through npm Trusted Publishing with provenance, without a token. The nine packages share one version.

The Reflex wrapper and the documentation

There are two more repositories.

reflex-silverpoint-react started on 25 September and has eleven releases, from 0.1.0 to 0.10.0. It wraps the components of @silverpoint/react as Python builders. It has a curious history: the palettes, the composed components and the timeline were drawn there first, in its 0.7, outside the parity contract. silverpoint's constitution records it as the source of the 0.7.0 inventory, where they were rewritten on the core. Since the wrapper's 0.7.1, they are the npm components behind the builders.

The documentation site is built in Reflex with that same wrapper (silverpoint_web). It is frontend only: it is exported statically to Vercel, with no backend, and its controls are silverpoint components. Its vercel.json also publishes the MCP server from article 28: it rewrites /mcp to the Vercel project that deploys it from apps/mcp-vercel.

The AI assistants

The repository does not hide it. 244 of silverpoint's 250 commits carry Claude's Co-Authored-By line. In the wrapper it is 27 of 59, and in the documentation site, 13 of 49.

The specification explains why it is built this way. Article 9 of the constitution, the one that requires every task and every test to cite a requirement, gives this reason: “the project is implemented in large part with agents; without traceability there is no way to answer why a line of code exists.”

What the files show is a division of roles:

  • Decisions belong to the person. The context file marks each one with its date: “user decision, 2026-10-10” to publish the three waves as 0.6.0, “user decision 2026-11-01” to compare versions by release line. Analyze presents findings; approving them is a human decision. Every feature gate has its approval date.
  • The agent works in batches. tasks.md asks for it: the first batch is T-001 to T-004; then review, adjust, and only then scale. “Do not launch all 29 unsupervised.”
  • The context lives in the repository. The plan says it is a one-person project with no ceremonies. tasks.md is marked with a date as each task closes, so that any session, human or agent, can resume with no prior context. The context file has a section titled “Settled decisions — do not rediscover them”.
  • A deviation stops the work. If the code cannot meet the specification, the specification is updated or a delta is opened, and only then does the work go on.

The CI gates are the other half. An agent can write code that looks right; the SVG string, the pixels, the contrast and the traceability say whether it is.

What was hard

The repository also keeps what went wrong or cost more than expected.

  • Hatching is heavy. Hatching each shape separately came to 913 KB for a 12-card dashboard. The weight report shows it per chart: with per-shape hatching, the large heatmap takes 61,348 bytes, over the 40,000 budget. The decision was one hatch pattern per tonal level, hatchFill: 'tile', which brings that heatmap down to 3,909 bytes.
  • A field that already existed. cyanotype needed a line weight per tone. Stroke.weight existed, but it was the BulletChart target marker's, and reusing it changed 48 canonical renders. A separate field was created.
  • Publishing is code too. The trusted publisher of @silverpoint/mcp on npm expired before 0.9.0, and npm answered E404. That package's 0.8.0 had been published by hand, without provenance, because it had no publisher yet.
  • React's strict mode. SpDeck was inert under <StrictMode>: the double mount destroyed its controller. It was fixed in 0.9.1, the same day as 0.9.0.
  • What is still open. The manual NVDA and VoiceOver pass over the components needs Windows or macOS and is pending. The nightly workflow is disabled.
  • What this series turned up. Writing it raised eleven issues, #12 to #22, such as SpTable cells aligned to the right or the exported SVG without its stylesheet. The ones in code were fixed on 15 November, in silverpoint 0.10.0 and in reflex-silverpoint-react 0.10.0. The two about documentation, #15 and #16, were fixed on the site.

End of the series

This is where the series ends. It started with a Renaissance drawing technique and ended with a server that draws charts for an assistant, by way of one operations panel in four frameworks.

If you came straight here, go back to article 0. Its “Reading paths” section has a short route for each framework, one for evaluating the library and one for working with AI. If you already followed it, the documentation and the gallery keep up with every release.