Changelog

All notable changes to the htmlcapsule project — spec, reference implementation, landing page, voices archive, and research record. The full-spec version (e.g. v0.3.4) is the load-bearing identifier; the Core spec version (v0.3.0) bumps only for normative rule changes; landing / notes / voices versions move independently and are tracked here only at major revs or when they coincide with spec work.

Format follows Keep a Changelog loosely. Commit IDs in parentheses for cross-reference.

[Unreleased]

Nothing pending right now. v13 landing-cleanup queue from earlier was largely satisfied by the v13.1 framing recovery + v13.2 on-site rendering work. Idea queue (spec/DOMAIN_CAPSULES.md) and voices queue (voices/README.md) hold candidates that haven't met the empirical-pressure bar.

[Spec v0.3.8] — 2026-05-22

Second documentation patch in 24 hours — no normative schema changes, no validator changes, no compat surface added. Lands the technical thesis under the format: a capsule is a document first, an app second. Apps when alive. Documents when dormant. Plus the conceptual elevation of Rule 2 from constraintive (a rule capsules follow) to definitional (the line where capsules end and a different category of artifact begins). Plus the reader-side floor declaration pattern that closes the loop on the 5-tier framework. The slogan is the compression; the five-tier framework, the capsule boundary, and the floor declaration are the engineering substance underneath it.

Landed in two commits. Part 1 (ddda6fc): document-first thesis, 5-tier framework, §2.3.3 reorganization, "Document-first artifact" glossary entry, landing-page touch, version bumps. Part 2 (this commit): capsule boundary (§1.5), floor declaration pattern (§2.3.3 sub-section), Core Rule 2 reframe, "Capsule boundary" + "Floor declaration" glossary entries. Both commits land v0.3.8.

Why now

Three pieces of pressure converged in the same conversation.

(1) Producer-side framing pressure: the four-producer family (Mintel, capsule-midi, Shasta, capsule-photo) kept arriving at the same crossroads — should this capsule be a runtime app, or a static document? — and the existing §2.3 / §2.3.1 framing answered "both" without naming the design vocabulary that lets producers stack deliberately.

(2) Consumer-side misreading pressure: external commentary on the project periodically over-rotates in either direction ("Capsule forbids interactivity" / "Capsule is a sealed mini-app"). The 5-tier framework is the vocabulary that makes the actual design space legible without either misread.

(3) Boundary-clarity pressure: a related conversation surfaced that Rule 2 (no network) had been operating as a quiet technical constraint rather than as a definitional commitment, leaving room for "but what if a capsule just fetched a little live data" misreads. Elevating Rule 2 from constraintive to definitional language — "network requests are outside the capsule boundary; an artifact that crosses it is a different category, not a degraded capsule" — removes the negotiation surface entirely. The seal is what makes the floor possible; without the seal the entire interactivity stack collapses because there's no guaranteed substance at any tier.

The slogan landed alongside two alternatives ("The capsule is a document first, an app second." / "Progressive enhancement for AI-era artifacts.") and was chosen because it compresses the temporal dimension that matters most to the format's preservation case: a capsule may live for an hour next to its producer, then sit dormant for years; the dormant case is when the document-first property earns its keep.

Added — spec (documentation only; no normative changes)

• §2.3 epigraph — adds a slogan-bearing two-line preamble to the Rendering Model section: "A capsule is a document first, an app second. Apps when alive. Documents when dormant." The epigraph names Rule 12 as the enforcement of that thesis at the rendering layer and forwards to §2.3.2 for the interactivity tiers built on top.
• §2.3.2 new — "Document-first: the five tiers of capsule interactivity." The organizing framework for the rest of §2.3.x. Five tiers tabulated with mechanism, no-JS survival, and capsule examples:
• Tier 0 — Static document (text, images, headings, tables — every preview surface)
• Tier 1 — Native HTML interaction (<details>, <audio controls>, <video controls>, <a href="#anchor">, SVG <title> / <a> — browser-owned controls)
• Tier 2 — CSS-state interaction (<input type="radio"> + :checked ~, <input type="checkbox"> + sibling selectors, :target — form-control state machines with CSS-only logic)
• Tier 3 — Precomputed interactivity (selection among prebuilt alternates shipped inline — the substance pattern that tier-2 mechanisms select among)
• Tier 4 — JavaScript runtime (live filter / sort / search / DSP / continuous controls / network — the only tier that doesn't survive no-JS)

Plus compounding logic (a well-built capsule stacks tiers; substance lives at the lowest tier that can carry it; higher tiers are progressive enhancement), a worked stack for domain.exploration_map (tier 0–4 inventory), and where this lands in the four-producer family mapping each producer's typical tier residence (photo at 0–2, song at 1, music_stems at 4-with-1-fallback, exploration_map at 4-with-0-3-substance).

• §2.3.3 (renumbered from §2.3.2) — "No-JS interactivity techniques." Same content; reorganized so each technique sits under the tier it implements. Tier 1 primitives table (<details>, <audio>/<video>, SVG natives, plus the new <noscript> reader-notice row added in part 2 pointing at the floor-declaration sub-section). Tier 2 primitives table (radio-tabs, checkbox-layers, :target nav). Tier 3 substance pattern (pre-rendered alternates). The three worked snippets and the "what still requires JS" honest list and the two producer caveats are preserved unchanged.
• §2.3.3 new sub-section "Tier-4 capsules: declare the floor" (part 2) — adds the <noscript> reader-notice pattern as a tier-1 primitive recommended specifically for tier-4 capsules. Worked snippet uses layer vocabulary ("You're viewing the document layer of this capsule") rather than mechanical framing ("JavaScript is disabled"). Framed as a floor declaration — not an apology for missing features but a courtesy to the reader naming what tier they're seeing and what the runtime tier adds. Tier-4 only (DAW transport, live mix, search, in-capsule LLM, vector pan/zoom); tier 0–3 capsules don't need the notice because there's nothing meaningful to miss. The notice is recommended, not required — the boundary (§1.5) already guarantees the floor whether or not the notice is present.
• §2.3.1 internal reference updated — the v0.3.7 cross-reference to §2.3.2 below was renumbered to §2.3.3 (techniques) and extended with one sentence pointing forward to v0.3.8's §2.3.2 (framework). Appendix E.12's reference updated correspondingly.

Added — spec (the capsule boundary, part 2)

• §1.5 new — "The capsule boundary." Elevates Rule 2 from constraintive ("capsules must not make network requests") to definitional ("network requests are outside the boundary; artifacts that cross it are a different category"). Names what's inside the boundary (5 blocks, data: URIs, runtime operating on inline data, user-dropped files processed in-runtime, computed-at-runtime derived data) and what's outside (network fetches across all browser APIs, external resources via <script src>/<link href>/<img src=http(s)://>, live API/database/LLM data, server-held auth/session state). Quotable thesis sentence: "The seal is not a restriction on what capsules can do — it's what makes capsules possible. Without the seal, there's no floor, no archive, no durability, no 'open this in ten years and it still works.' Rule 2 isn't a limitation; it's the definition." Includes the user-provided-files clarification: a capsule that accepts a CSV/image/MIDI/etc. via <input type="file"> and processes it entirely in-runtime does NOT cross the boundary (no network egress; the file lives in the user's browser session). Names the validator's scope-aware scan (v0.3.6) as the mechanism that lets prose about the forbidden APIs (e.g., this section) not trip Rule 2. Closes with the different-problems-different-formats framing: external-data artifacts solve a different problem (anti-staleness, always-current) with a different commitment; capsules solve anti-context-loss with the seal commitment. Both legitimate; trying to be both at once means being neither well.

Changed — Core (part 2; clarification, not a rule change — Core stays at v0.3.0)

• CAPSULE_CORE.md Rule 2 — wording reframed from purely constraintive ("No network. Zero fetch, XHR, CDN references…") to definitional + constraintive ("No network — the capsule boundary is definitional. … The seal is not a restriction on what capsules can do — it's what makes capsules possible. An artifact that fetches live data, calls an external service, or depends on network availability at runtime is a different category of artifact (web app, live dashboard, connected document) — not a degraded capsule. … See spec/CAPSULE_SPEC.md §1.5 for the conceptual treatment."). Not a normative change — the rule's enforcement is identical; the framing carries more conceptual weight, which is what survives into LLM prompts every time the Core is pasted. Same pattern as v0.3.6's Rule 4 extension (added derived_from to the listed optional fields without bumping Core) — clarifying-not-changing edits to Core text don't move the Core version number.

Added — glossary

• GLOSSARY.md — new entry "Document-first artifact" (part 1) in the Positioning terms section. Lives alongside Sealed handoff format, Portable context contract, Anti-context-loss, Universal AI reader, Interactive archive. Names the thesis + slogan + 5-tier stack + the "open in QuickLook and Safari: same file, two experiences, no degradation cliff" payoff. Cross-references spec §2.3.1, §2.3.2, §2.3.3 and the Interactive archive entry.
• GLOSSARY.md — new entry "Capsule boundary" (part 2) in the same section. Names the inside/outside line, the seal's enabling (not restrictive) role, the "different category not degraded capsule" distinction. Cross-references Core Rule 2 and spec §1.5.
• GLOSSARY.md — new entry "Floor declaration" (part 2) in the same section. Names the <noscript> reader-notice pattern, the layer-vocabulary tone guidance, the tier-4-only recommendation. Cross-references spec §2.3.3.
• Glossary state line bumped 0.3.0 → 0.3.1 and spec-cross-reference bumped v0.3.2 → v0.3.8 (catching up two prior revs that hadn't propagated).

Changed — landing page (v13.2.1)

• index.html "What this is not" section gains a fourth bullet: "Not an app pretending to be a document. A document that can wake up into an app." Heading bumped "Three lines of fence-posting" → "Four lines of fence-posting." Surfaces the document-first thesis to landing-page readers without expanding the hero or pushing the spec doorways. capsule_version 13.2.0 → 13.2.1 (patch — single bullet + heading word).

Changed — infrastructure

• CITATION.cff 0.3.7 → 0.3.8.
• README.md v0.3.7 → v0.3.8 (state line + body references + phase-table row).
• llms.txt v0.3.7 → v0.3.8 + full-spec description updated to summarize v0.3.8 (slogan + 5-tier framework + §2.3.3 renumber) + CHANGELOG version range bumped to v0.3.0 → v0.3.8.
• 8 markdown capsules rebuilt by compiler/build_md_capsules.py to pick up the spec + glossary + changelog content updates. All pass 26/26.

Numbering note

The user's approved plan named the framework subsection §2.3.0 (a zero-numbered subsection that would have lived before §2.3.1). The actual landed numbering is §2.3.2 (new framework) + §2.3.3 (renumbered technique inventory), which reads rule (§2.3) → principle (§2.3.1) → framework (§2.3.2) → techniques (§2.3.3). The deviation is editorial — narrative ordering benefits from reading the framework after the principle that sets up why tier separation matters, and most readers won't expect a 2.3.0 subsection. The substance (slogan + 5-tier framework + glossary entry + landing touch) is exactly as approved.

Cross-project notes

No FEEDBACK harvest required for v0.3.8 (no items in producer-side queues). The 5-tier framework is reference material that all four producer projects (Mintel, capsule-midi, Shasta, capsule-photo) can cite when documenting their own tier residence. Suggested follow-on: each producer project adds a brief "Tier residence" note to its README naming which tiers it ships at, which makes the multi-producer interop story legible across the family.

[Spec v0.3.7] — 2026-05-22

Documentation patch on top of v0.3.6 — no normative schema changes, no validator changes, no compat surface added. Lands a documented inventory of no-JS interactivity techniques + parks one new candidate manifest extension.

Added — spec (documentation only; no normative changes)

• §2.3.1 gains a second tagline: "No-JS capsule interactivity is declarative, not computational. It should expose precomputed states through native HTML and CSS controls." Pairs with the existing v0.3.6 tagline ("degrade from app → document → preview"). One half of the principle (degrade), the other half (technique).
• §2.3.2 added — "No-JS interactivity techniques." Pattern inventory documenting that no-JS capsules can support real interaction through native HTML controls + CSS state selectors, not just static documents. Seven techniques tabulated with recommended use cases:
• <details> / <summary> — collapsible drawers (manifest panels, per-stem metadata, source notes, license, etc.)
• <audio controls> / <video controls> with data: source — native media playback
• Radio buttons + CSS sibling selector — tabs (one of N views visible at a time)
• Checkboxes + CSS sibling selector — layer toggles (N independent on/off)
• :target pseudo-class + URL hash — page-like navigation
• SVG <title> / <a> / hover CSS — static-but-rich diagrams
• Pre-rendered alternate states — selection-driven alternatives

Three worked snippets: radio-button tabs, checkbox layer-toggles, :target navigation. Honest "what still requires JS" list (live MIDI synthesis, stem mixing with continuous volume control, file parsing, search/filter/sort, save state, map pan/zoom, AI interaction, network). Two real caveats for producers: (1) size budget — pre-rendered alternates count against the 20 MB cap; for media, plan ~3-4 MB per alt at reasonable quality; (2) information-architecture cost — radio-button tabs and :target navigation require the full content of every view in the static HTML, so 5 views = 5× the body content.

• spec/DOMAIN_CAPSULES.md — per-domain notes added:
• domain.music_stems — recommends short-loop / sketch scope with a small set of pre-rendered alt mixes (full + 2-5 strategic alts at Opus 96 kbps fits comfortably under 20 MB at ~3-minute lengths), each selectable via radio-button tabs. Also: when this graduates, the media. family (v0.3.6 §5.1.2) supersedes the ad-hoc audio. capability names sketched in the original entry.
• domain.photo — noted as the cleanest fit in the four-producer family for demonstrating the v0.3.7 no-JS techniques in anger. Radio-button tabs for view-switching (Photo / Metadata / EXIF / Location / Provenance); checkbox layer toggles for face-region overlays on top of the static image. Suggested as the v0.3.7 reference demonstration of the techniques once capsule-photo ships.

Added — Appendix E (parked, not adopted)

• E.12 — "No-JS interactivity declarations in the fallbacks manifest section." Three candidate sub-fields for the existing fallbacks field (§3.2.1): native_controls, css_views, precomputed_assets. Parked because (a) producers can use the techniques today without manifest declaration (the patterns are visible in rendered HTML), (b) no consumer has yet asked the question "what no-JS interactivity does this capsule provide?", (c) risk of premature canonicalization while the patterns are still evolving in producer practice. Graduation trigger named: a registry viewer's index-thumbnail wanting to deep-link a specific CSS-state view, or a downstream search tool wanting to enumerate bundled alt mixes without parsing <audio> elements. Same parking discipline that worked for derived_from[] (parked 2026-05-21, graduated 2026-05-22 — 24-hour cycle when empirical pressure arrived).

Changed — infrastructure

• CITATION.cff 0.3.6 → 0.3.7.
• README.md v0.3.6 → v0.3.7 (state line + body references).
• llms.txt v0.3.6 → v0.3.7 + full-spec description updated to summarize v0.3.6 and v0.3.7 separately + CHANGELOG version range bumped.
• 8 markdown capsules rebuilt by compiler/build_md_capsules.py to pick up the spec + DOMAIN_CAPSULES + CHANGELOG content updates. All pass 26/26.

Why a same-day patch rather than rolling into v0.3.6

v0.3.6 shipped earlier today as a normative spec release (graduated derived_from[], formalized media.* family, added export.fragment_provenance, named graceful-degradation principle). v0.3.7 is a same-day documentation pass on top: no schema changes, no validator changes, just adds the technique inventory + parks one new candidate. Keeping them separate preserves the v0.3.6 release entry's clarity (normative changes there) while letting the v0.3.7 entry stay focused on the technique documentation. Future readers can see at a glance which release introduced normative changes vs which was a docs-only refinement.

Cross-project notes

No FEEDBACK harvest required for v0.3.7 (no items in producer-side queues). The techniques in §2.3.2 are useful for capsule-photo (suggested as the first demonstration) and capsule-midi (whose Tier 2 work to bundle a rendered audio mix can adopt the radio-button-tabs alt-mix pattern). Neither producer is blocked by this release; both gain documented reference material.

[Spec v0.3.6] — 2026-05-22

First spec release driven by the upstream feedback discipline (named in F28) — every change in v0.3.6 has empirical pressure from a downstream producer project, none are speculative. The producer family that surfaced the pressure: capsule-midi (MIDI / DAW capsules, v0.2.0), Shasta (audio songs, in progress), capsule-photo (single-image artifacts with metadata, new), alongside the existing Mintel (geospatial maps, production at v0.3.4 since F20/F21). MinDev remains the registry / delivery host (separate concern).

Added — spec (normative)

• §11.2 Non-Capsule provenance (derived_from) — graduated from Appendix E.11. Empirical pressure: capsule-midi's first non-synthetic build (Mozart Lacrimosa) had a composition reference that isn't itself a Capsule. Producers were stuck either burying composition lineage in prose description text (loses machine-readability) or minting a synthetic v5 UUID and putting it in parents[] to satisfy the validator's parents[i].uuid is required check (workaround, not a clean shape). Field is optional. Required per entry: type + title. Recommended: reference (URL/URN, may be null) + role. Optional: hash, date, plus domain-specific fields. Coexists cleanly with parents[]: parents[] is strict Capsule-to-Capsule (UUID required); derived_from[] is everything else. Section 11 renumbered: §11.2 (deprecated related) → §11.3, §11.3 Index Capsule → §11.4, §11.4 Registration → §11.5. Internal cross-reference at line 713 updated.
• **§5.1.2 The media.* capability family** — multi-track media transport (play / pause / stop / seek), tempo/loop (tempo_scale / loop_bar), per-component (stems.mute / stems.solo / stems.set_patch / stems.set_volume), and per-media downloads (midi.download / audio.download_mix / audio.download_stem). 13 capabilities documented as a coherent namespace. Producer reference: capsule-midi v0.2.0 implements all 11 of the symbolic-side entries; Shasta is expected to use the audio-side entries.
• §5.1.3 The export.fragment_provenance capability — universal envelope (from_capsule + from_capsule_title + from_capsule_version + exported_at + fragment + note) with domain-specific fragment shape. The load-bearing primitive of remix-style workflows across the producer family: capsule-midi exports component+bar-range; Shasta would export stem+time-range; capsule-photo would export rectangle/region/face; Mintel would export bounding-box/feature. One capability name unblocks all downstream lineage tooling.
• §2.3.1 Graceful degradation as a design principle — promotes the implicit Rule 12 + JS-off-litmus framing in §2.3 to a named design principle. Tagline: "A capsule should never become useless when JavaScript is unavailable. It should degrade from app → document → preview." Names the three modes (runtime / document / preview) and identifies iOS Files-app QuickLook as the canonical hostile environment to design around. References the existing image-fallback worked example as one instance of the broader pattern.
• §3.2.1 The optional fallbacks manifest section — declarative metadata about what JS-off fallbacks the capsule bundles: preview_audio_present, poster_image_present, static_summary_present, requires_js_for (string array), preview_mode_description (prose). All optional within the section; section itself optional. Lets consumers discover available fallback paths without scraping rendered HTML. The actual fallback content still lives in the static HTML per Rule 12; this section is metadata.
• JS-off fallback guidance per domain (in spec/DOMAIN_CAPSULES.md, between "Initial domains" and "How to propose a new domain") — table covering all 8 domain entries (4 graduated + 4 idea-queue). For most domains the answer is "the static HTML IS the substance, no extra work needed" (made explicit). For the music / map / photo cases the recommended fallback is named: domain.midi_stem → bundled rendered audio mix; domain.song → embedded MP3 already serves as the fallback natively; domain.photo → image itself; domain.exploration_map → already-documented image-fallback.

Changed — Core (clarification, not a rule change — Core stays at v0.3.0)

• CAPSULE_CORE.md Rule 4 — added derived_from to the listed optional manifest fields with a one-line description and cross-reference to §11.2 of the full spec. Not a Core rule change (Core's discipline of allowing optional fields is unchanged); just brings the listed-fields enumeration in sync with the full spec's normative section. The "Deprecated fields" note for related updated to mention both parents and derived_from as the replacement homes for provenance.

Changed — validator

• compiler/validate.py field-format check now distinguishes parents[i].uuid is required (Capsule-to-Capsule lineage) from derived_from[i].type is required + derived_from[i].title is required (non-Capsule sources). The parents[i].uuid is required error message now includes the hint "for non-Capsule sources use derived_from[] — see spec §11.2" so producers hitting the error get pointed at the right field.
• Scope-aware external-reference scan (already shipped at f61e504; documented here for completeness) — EXTERNAL_PATTERNS split into MARKUP_PATTERNS (scanned everywhere), JS_PATTERNS (scanned only inside <script> blocks), and CSS_PATTERNS (scanned only inside <style> blocks). Code-block and <pre> content excluded from the scan via _strip_data_blocks. Eliminates false positives on capsules that render documentation about the forbidden APIs (e.g., a capsule with prose containing "no fetch (rule 2 stays intact)" no longer trips the fetch( regex). Not a Rule change; a heuristic refinement.

Added — research record

• F28 — Producers reach for Capsule-shape independently when given the idiom but not the spec. Companion to F25. Empirical case: ChatGPT-produced MIDI POC reached for the shape (5-block-ish, parents[], sha256, license_note) without Core attached, but missed the specifics (5-block-exact, integrity hash, capability convention). Validator: 5/10 against the reference. Names the upstream feedback discipline as a cross-project methodology: producer projects own the friction, spec project owns the response.
• F29 — iOS QuickLook surfaces graceful degradation as a first-class spec principle, not just a Rule 12 implication. Empirical case: capsule-midi v0.2.0's <noscript> warning naming iOS Files-app preview explicitly + external strategic-review discussion proposing three-mode taxonomy and fallbacks manifest field. Spec response landed in this v0.3.6 release as §2.3.1 + §3.2.1 + per-domain guidance. Also includes the architectural-alternatives rejection (package format / native viewer app / required hosted viewer) so the rejection is on the record.

Added — domain idea-queue entries

• domain.midi_stem (capsule-midi) — symbolic-only multi-track music capsules. Already shipped v0.2.0 working implementation. Graduation pending first publicly-validated capsule.
• domain.song (Shasta, in progress) — audio-primary song capsules. Pre-graduation; first capsule from the F26 song-capsule experiment.
• domain.photo (capsule-photo, new) — single-image family-photo artifacts with metadata. Different namespace family from media.* (image.* sibling).
• Cross-reference note on the existing domain.music_stems (parent multi-modal idea) naming midi_stem and song as specializations.

Changed — infrastructure

• CITATION.cff version bumped 0.3.5 → 0.3.6; date 2026-05-22.
• README.md state line: full spec v0.3.5 → v0.3.6; F-finding count F25 → F29.
• llms.txt full-spec line updated with v0.3.6 + summary of changes; F-finding count updated everywhere F25 → F29; CHANGELOG version range v0.3.0 through v0.3.5 → v0.3.0 through v0.3.6.

Cross-project annotations (capsule-midi/FEEDBACK.md)

Resolved upstream in this release: Q1 (domain.midi_stem parking), Q2 (media.stems.* namespace), Q3 (export.fragment_provenance capability), Q4 (derived_from[] graduation). The capsule-midi project's FEEDBACK.md should be updated in a coordinated commit to mark these as resolved and move them to its "Resolved" section. F-A was filed as F28 in this release.

[Landing decision — v13.0.0] — 2026-05-22

Not a spec change. The landing-page exploration arc that ran from v10.x through v12.0.0 (nine hero candidates, numbered Observation / Question / Answer editorial spine, devil's-advocate copy revision pass, the Claude Design design-system pass) is now closed. The maintainer has committed to the Linear/Stripe-style stripped landing direction — iterated through landing-sketch.html (v1, UUID 9bde04e1) and landing-sketch-v2.html (v2, UUID 2860450e) — as the production landing.

Changed — index.html (v12.0.0 → v13.0.0)

• Page structure completely replaced. The 1883-line research-narrative hybrid (hero gallery 01-09 + Observations 1-5 + Questions 1-3 + Answers 1a/1b/2a/2b/2c/3a/3b/3c + downstream framing) is gone from the live page. In its place, the 612-line stripped landing from landing-sketch-v2.html (v0.3.0 form, post-design-system-pass): one hero ("HTML you can keep."), one primary CTA ("Read the spec"), one secondary CTA ("View an example"), the 5-block anatomy card with the manifest row highlighted in indigo, the lifecycle pill strip with Seal as the solid-indigo project slot, the rose-tinted "What this is not" callout, the slim demo + about + footer.
• Identity preserved. UUID 7d1a1ac8-c6d9-4ed1-98a7-f5399466262a — the canonical landing's stable identity across structural restyles — is retained. parents[] records the two sketch UUIDs the design was iterated through (2860450e v0.3.0 + 9bde04e1).
• Version jump. capsule_version 12.0.0 → 13.0.0 (major structural restyle, not just a token swap).
• Title. "HTML Capsule — hero framing exploration" → "HTML Capsule" (production title, not exploration title).

Added — exploration.html

• New file. Captures the prior index.html content (the v12.0.0 form: full research-narrative hybrid + 9 hero candidates + numbered editorial spine + the Claude Design design-system pass already applied at the section-kind taxonomy level — Observations teal, Questions violet, Answers amber).
• Fresh UUID 881fed04-727e-441b-b626-c0637585043e — this is now a distinct artifact (an archive page), not the production landing it once was.
• capsule_version 1.0.0 — first version of this file as an independent capsule.
• parents[] references the canonical landing's 7d1a1ac8 (the v12.0.0 form was the immediate ancestor).
• Archive banner (amber-tinted aside, top of body) signals the role explicitly: "Archive. This page is the landing-exploration phase that ran through May 2026 ... preserved here for record. The production landing is now at htmlcapsule.org."
• Why preserved as a live page instead of git-only: the editorial content (Observations / Questions / Answers with verified-quote sourcing, three convergence findings, the maintainer's case study, the lineage / diff-tool / capsule-aware-review-surfaces argument) is meaningful research material in its own right. Future readers — and future Claude / future-maintainer — should be able to follow the argument as a live document, not have to git-archaeology it.

Unchanged — sister sketches and notes

• landing-sketch.html (v1), landing-sketch-v2.html (v2), research-sketch.html (v1), research-sketch-v2.html (v2), positioning-sketch.html — all preserved in place as records of the exploration phase. Each carries its own UUID and design-system-pass version (v0.3.0 / v0.2.0 / etc.). The v2 design-system passes are the canonical record of how the design was applied per genre commitment.
• notes.html (v1.3.0) — already the production essay page; no integration needed. Design system already applied; preferred by maintainer in the same breath as the landing.

What this validates

Three independent reads converged on this commitment over the May 2026 exploration arc: the devil's-advocate critique pass identified the genre tension on the long-form hybrid; the landing-agent's hero pick selected "HTML you can keep." as the strongest single claim; the external ChatGPT Deep Research review framed the project as research that doesn't need a sales-y landing. Each pointed at a different exit from the hybrid; the maintainer's chosen exit is the Linear/Stripe-style stripped landing (the cleanest conversion shape) with the research narrative preserved live at /exploration.html for those who want it.

Added — research record (post-ship)

• **F27 — *The landing-page genre tension for applied-research projects resolves by splitting, not merging.*** Recorded in RESEARCH.md as the formal capture of this decision. Generalizable observation: a single URL cannot serve both "convert in 30 seconds" and "walk the full argument with citations" without paying both genres' costs. The two-page split — one URL per audience, same UUID lineage — is what worked. Includes a method-observation: external review convergence on a structural decision (here, three independent reads pointing at the split) is empirical pressure even when the decision feels like cowardice.

[Design system v1] — 2026-05-22

Not a spec change. First canonical visual design system for the project's landing pages. Received as a handoff bundle from Claude Design (claude.ai/design) after the maintainer asked it to "propose visual enhancements for all of these" (the seven landing-page sketches in flight). Claude Design produced two proposals; the maintainer rejected the first (editorial-archival, italic display serif, oxide accent) and accepted the second (modern-sans, weight contrast, semantic color taxonomy). The chat transcript captures the rejection language verbatim and the project preserves both proposals for record.

Added — design folder

• design/proposal.html (canonical) — full design memo from Claude Design with mockups of all four target pages (landing, research, positioning, notes). Contains the design's CSS in context. Uses Google Fonts directly; this is a design-tool output, not a Capsule, and is preserved as-received.
• design/proposal-v1-editorial.html (rejected) — the first version Claude Design produced before the maintainer's redirect. Preserved for record.
• design/README.md — documents the system in repo-native terms: origin story, five-hue taxonomy table, paper/ink/rule token tables, typography pattern (weight contrast replaces italic), Capsule-compliance discipline (system-font fallback, no network), and a per-page application table mapping each landing page to its lead taxonomy color.
• design/tokens.css — extractable CSS variable block + utility classes for copy-paste into Capsule <style> blocks. System-font fallback only; Core rule 2 compliant when inlined.

The system in brief

• Type: Geist (300, 400, 500, 600, 700, 800) + Geist Mono, with system-sans fallback for Core rule 2 compliance when inlined into Capsules.
• Paper system (cool fog): --paper #f5f6f8 / --paper-soft #ebedf1 / --paper-deep #dee1e7 / --paper-page #ffffff.
• Ink scale: five values from --ink #0c0e13 down to --ink-quiet #b0b5be.
• Rule scale: --rule / --rule-soft / --rule-strong.
• Five semantic hues, each with base/soft/edge/deep variants:
• Indigo #4f46e5 — brand · primary CTA · headline emphasis · spec material
• Violet #7c3aed — Questions · open work · research-flagged
• Teal #0d9488 — Observations · evidence · quotes
• Amber #b45309 — Answers · maintainer voice · essay
• Rose #be123c — Pain · negation · "shouldn't die" framing
• Pattern: weight contrast (300 light + 800 punch) replaces italic emphasis in headlines.
• Radii: 8px (controls), 14px (cards), 999px (pills).

Changed — landing page (index.html v11.19.0 → v12.0.0)

First page to adopt the canonical design system. Visual-layer-only pass — editorial copy unchanged from the v11.x devil's-advocate revision pass.

• :root token block replaced with the full design system (paper / ink / rule scales + five-hue palette with base/soft/edge/deep variants + Geist system-font stack).
• Back-compat aliases preserved. --accent now maps to --indigo; --accent-soft is a brighter indigo #6366f1 (used for dark-frame hover on data-num="05"); --accent-deep maps to --indigo-deep. Existing rules pick up the new brand color without rewriting every reference.
• Section-kind backgrounds re-mapped to the taxonomy:
• .is-observation → var(--teal-soft) (was white)
• .is-question → var(--violet-soft) (was paper-soft grey)
• .is-answer → var(--amber-soft) (was cream #fcfaf3)
• Kind-tinted inline text — .cb-kicker (the "Observation 1:" / "Question 2:" / "Answer 3c:" prefix), .cb-frame-tag (small mono section labels), and .cb-aster pick up their hue's deep variant so the taxonomy carries through to inline text, not just card backgrounds.

UUID 7d1a1ac8 preserved throughout (the page is the same capsule, restyled).

Changed — sister sketches (parallel design-system pass)

Four sister sketches updated in parallel to apply the same design system, each picking up its lead taxonomy color per the application table in design/README.md:

• landing-sketch-v2.html — lead color: indigo. Hero "HTML you can keep." picks up weight contrast (300 light "HTML you can" + 800 indigo "keep."). Anatomy card with manifest row highlighted in indigo.
• research-sketch-v2.html — lead color: teal. Paper title "HTML Capsule" picks up weight contrast (300 + 800 teal). Findings O1-O5 converted to a 2×3 grid of taxonomy-colored cards (O1 teal, O2 indigo, O3 amber, O4 violet, O5 rose).
• positioning-sketch.html — lead colors: rose → indigo. Two-color headline ("die" in rose 800, "seal" in indigo 800). Lifecycle becomes a weighted timeline with the Seal node 30% larger, indigo-filled, on an indigo halo.
• notes.html — lead color: amber. Amber category chip, margin numerals in amber 700, amber pull quote at the "file over app" pivot moment.

All sister sketches retain their existing UUIDs (capsules are restyled, not replaced); capsule_version bumps minor for each.

Provenance

• Bundle origin: https://api.anthropic.com/v1/design/h/1h9LEWoDBWZgTL3Fqnnz6w (received 2026-05-22)
• Bundle path: html-capsule-improvements/
• Design tool: Claude Design (claude.ai/design)
• Decision chat: preserved in the bundle at html-capsule-improvements/chats/chat1.md (the design-tool conversation where direction was decided)

[Spec v0.3.5] — 2026-05-21

Doc-only patch. No schema, validator, or behavior change. Records external-review-driven candidate manifest fields in spec Appendix E without adopting them, preserving the project's "no addition without empirical pressure" discipline while making the considered-but-not-adopted ideas discoverable to future readers (including future Claude / future-maintainer who might re-derive them from scratch otherwise).

Added — spec

• Appendix E.11 "Extended manifest fields raised by external review (parked)" in spec/CAPSULE_SPEC.md. Records three candidate fields surfaced by two converging external-LLM reviews of the v11.x landing page's HTML-version-control narrative (Obs 3 → Q3 → Answers 3a/3b/3c):
• supersedes[] — UUIDs of capsules this one explicitly replaces (distinct from parents[] which records what this descended from). Semantic difference: "I came from X" vs. "use me instead of X."
• derived_from[] — non-Capsule sources this artifact was built from (chat sessions, screenshots, external documents, datasets). Different shape from parents[] (Capsule-to-Capsule lineage).
• change_summary — short human-readable note about what changed vs. the previous version; like a commit message scoped to the capsule_version bump. Would feed a future capsule-diff tool.

Each is justified individually for why it's parked, not adopted (no real producer/consumer has hit the limitation yet for supersedes[] and change_summary; the Mintel producer has some real pressure for derived_from[] but no consumer has needed it programmatically yet).

• Methodological observation in the same Appendix E entry: external-LLM review is now a recurring source of spec-design candidates, alongside producer pressure, consumer pressure, and maintainer reflection. Worth tracking as a pattern in RESEARCH.md if it keeps happening.

Changed — infrastructure

• CITATION.cff version bumped 0.3.4 → 0.3.5.
• README.md updated in three places: intro state, full spec reference, status table.
• llms.txt updated to reflect v0.3.5 in the full-spec entry and the CHANGELOG range.
• index.html landing badge updated to v0.3.5, About panel reference updated to "full spec v0.3.5", data.spec.full field updated to v0.3.5. Landing capsule_version bumped 11.15.0 → 11.15.1.

Changed — infrastructure (freshness — surfaced by external strategic review)

• README.md F-finding lag fixed — three places that referenced "F1–F22" or "F22" updated to "F1–F25" / "F25". Added one new sentence acknowledging F24 (host-vs-registry, the missing commitment layer) and F25 (ChatGPT producer-population reads of Core supplementary guidance) as recent findings worth surfacing in the intro paragraph. External ChatGPT Deep Research review explicitly flagged that the summary surface (README, llms.txt) lagged the underlying corpus (RESEARCH.md) — this fix closes that gap.
• llms.txt F-finding lag fixed — Research-log line and CHANGELOG-description line both updated from "F1-F21" / "F1–F22" to "F1-F25". Research-log line gained a one-sentence note pointing at F24 and F25 by name.
• GLOSSARY.md "The four layers" section gained a bridge paragraph explicitly mapping the verb-shaped four-layer model (Think / Shape / Seal / Harden) to the noun-shaped ecosystem-stack framing in PRECEDENTS.md (substrate / live editing / seal / host / discover). Both name "seal" as the slot Capsule occupies; the bridge note prevents the two vocabularies from drifting apart as they get used across landing-page sketches and external reviews.

Added — research record (post-ship additions)

• F25 — ChatGPT producer-population reads Core supplementary guidance reliably; aesthetic adapts to content domain; legacy "Artifact Capsule" wording persists in user-side prompt templates. Recorded in RESEARCH.md. Drawn from review of a batch of 7+ ChatGPT (GPT-5.5 Thinking) conversation-summary capsules across varied domains (coding, geology, legal, fire pits, vehicles, mining permits, identity-rights, design-award fit). Five distinct sub-findings: (1) all five required blocks present; (2) Rule 4 supplementary QR-code guidance followed faithfully — placement, sizing, URN encoding all respected, which means Core's supplementary sections (not just the twelve numbered rules) are load-bearing in practice when written as recipes; (3) where Core leaves implementation open (integrity, Rule 7 verification convention) producers diverge; (4) producer aesthetic adapts to content domain (geological capsule → earth tones; fire-pit capsule → warm cream/orange; legal letter → clean neutral; etc.); (5) legacy "Artifact Capsule" terminology persists in user-side prompt templates because Core had no explicit "use the canonical name" reminder. Initial framing of QR-code presence as "emergent producer behavior" was corrected (caught by maintainer) — it's actually Core spec being followed correctly, a stronger positive signal.

• F26 — Core spec accommodates 10 MB domain-specific media capsules without rule changes. Recorded in RESEARCH.md. Drawn from a one-off domain-specific song capsule experiment (Paul McCartney & Wings, "Nineteen Hundred and Eighty-Five", 1973) embedding a 7.6 MB MP3 plus Wikipedia-sourced metadata plus a transcribed lyric sheet into a single 10.16 MB self-contained HTML capsule (UUID e26b58da, currently at capsule_version 1.1.0). Shipped 25/25 against the reference validator on first build with zero spec changes. Seven sub-findings: (1) domain via type field (type: "song") — Core accepts arbitrary domain values; (2) CSP delta minimal (media-src data: is the only addition required); (3) capability vocabulary extends naturally (media.play_audio and media.download_audio using <domain>.<action>; Rule 7 markers heuristically validated); (4) size limits hold under-cap (10.16 MB under both 15 MB soft-warn and 20 MB hard cap; validator graceful with binary-heavy under-cap); (5) round-trip extraction works (the embedded MP3 is recovered byte-identically via dataUriToBlob runtime helper — genuinely portable, not just embedded for display); (6) aesthetic adapts to content domain (warm 1970s earth tones — extending F25's same-pattern-finding from ChatGPT-text-capsules to hybrid-produced media); (7) version semantics in practice (v1.0.0 → v1.1.0 same UUID, capsule_version bump alone — the parked Appendix E.11 fields stayed parked correctly since nothing was distributed between versions, but the empirical-pressure point for supersedes[] is now recorded for the next time someone hits it).

Changed — Core spec (post-ship clarification, not a rule change)

• CAPSULE_CORE.md — added a "use the canonical name when you write your prompt" reminder immediately above the "How to ask an LLM to produce a capsule" section. Cross-references F25. Closes the loop on the legacy-prompt-template leak by making the canonical name unmissable to anyone templating their own prompts. No rule change — Core stays at v0.3.0; this is a clarification, not a normative edit. Full spec stays at v0.3.5.

Changed — landing page (devil's-advocate copy revision pass, v11.16.0 → v11.19.0)

Four-version sequential pass through the landing page's editorial spine — Observation 1, Question 1, Answers 1a/1b, Question 2, Answers 2a/2b/2c, Observation 2, Observation 3, Question 3, Answers 3a/3b/3c, Observation 4, Observation 5. All twelve sections tightened against a single set of devil's-advocate critiques: footnote-apologies → word swaps, redundant cross-references removed, walls-of-bold converted to real <ul>s, jargon ("empirical pressure", "anyone technical", "bring your own bucket") replaced with plain English, forward-pointer parentheticals stripped, weak hedges committed-or-cut, punchlines promoted from buried paragraphs to titles where they earned the slot.

• v11.16.0 — Observation 1 retitled "HTML has emerged as a key substrate" → "HTML is becoming the default output format." Self-apologetic footnote about "substrate" removed (the fix is the word swap, not the apology). Karpathy quote demoted from parallel blockquote to one-line attributed mention in body prose; Thariq carries the structural claim alone.
• v11.17.0 — Q1/A1a/A1b. Q1 depersonalized ("How do I get HTML out of my LLM" → "How do you get HTML out of an LLM"); A1a re-quote redundancy of Karpathy + Thariq dropped; A1b title hedge "(according to the Spec)" dropped; wall-of-bold contract-elements paragraph converted to real <ul>; "first-class spec-aware participant" promoted from body prose to .cb-frame-punchline.
• v11.18.0 — Q2/A2a/A2b. Q2 retitled with framing sentence; A2a tag retitled "Viewing strategy" → "View it"; A2b's two-paragraph 270-word architectural description collapsed to one punchline + one short body (~95 words), load-bearing "Format and host are different jobs" finding now visually dominates instead of buried.
• v11.19.0 — A2c (was held), Obs 2, Obs 3, Q3, A3a, A3b, A3c, Obs 4, Obs 5. A2c title "Maybe let's take it a step further" → "Build your own producer"; A3c punchline "The HTML is the artifact. The manifest is the review surface" promoted to title; Obs 5 title sharpened; Obs 4 hedge replaced with committed scope; cross-reference web removed.

Net effect across the four versions: visible prose 19,113 → 15,675 chars (~18% reduction), with section-level cuts in A1b / A2b / Obs 3 / A3c closer to 40%. UUID 7d1a1ac8 preserved throughout (the page is the same capsule, content-tightened — not a new derived capsule). All four versions validated 25/25 against the reference validator at ship.

Added — comparison sketches (parallel direction exploration)

Three sibling capsules built alongside index.html to test alternative genre commitments before deciding the page's direction. Each one is itself a valid Capsule (5 required blocks, 25/25 against the reference validator) and declares its position relative to index.html via the genre signal in its data block.

• landing-sketch.html (350 lines, UUID 9bde04e1) — Linear / Stripe / Vercel commitment. Hero: "HTML you can keep." One primary CTA, one secondary CTA, one code-block functional demo of the 5 required blocks, slim footer. Single accent color, dark/light theme. Zero numbered observations, zero cross-references, zero research jargon. Roughly 9% of index.html's prose volume.
• research-sketch.html (490 lines, UUID 766faed9) — NeRF / gaussian-splatting / academic project page commitment. Structure: Title + author block + abstract + numbered findings O1-O5 (deliberately avoiding collision with RESEARCH.md F1-F25) + Methods + Related Work + Spec & Artifacts + Cite this work + slim footer naming the page as a Capsule. Restrained academic aesthetic (serif body, single muted accent, generous margins). Zero CTA buttons; pointers do the conversion job.
• landing-sketch-v2.html (397 lines, UUID 2860450e, parents → 9bde04e1) — landing-sketch.html iteration absorbing external strategic review signals (ChatGPT Deep Research + maintainer devil's-advocate critique). Three additions kept from absorbing v1's conversion shape: (1) "Anti-context-loss" accent eyebrow above the hero (project's own native term from GLOSSARY.md); (2) sharpened lede itemizing failure modes; (3) CSS-only lifecycle pill strip ("Substrate · Live edit · Seal · Host · Discover") naming Capsule's slot in the landscape; (4) a "What this is not" callout pulling verbatim from GLOSSARY.md.

These sketches are exploratory artifacts, not landing-page replacements; the production landing remains index.html at v11.19.0 pending a maintainer commitment to one of the genres. (research-sketch-v2.html and positioning-sketch.html still in flight at time of writing.)

[Spec v0.3.4] — 2026-05-20

Interactive-archive category named; lifecycle decomposition (live-editing / format / hosting / discovery) made explicit across the project; voices archive grown from 0 to 2 archived + 2 queued; second convergence finding (F22) documented.

Added — spec

• §2.3 "Interactive archive" subsection — formal name for the category; JavaScript-off litmus test as the operational distinction between archive and app; permitted vs. forbidden examples (Mintel measure-tool as the canonical permitted example) (56b5b7b).
• §2.3 "Carve-out for visualization geometry" — image-fallback pattern documented as the principled resolution to Rule 12 in cases where geometry can't reasonably be pre-rendered as static markup (9bdb2db).
• spec/HOSTING.md — descriptive host-contract pattern observed across MinDev + htmlbin (short URL + /raw endpoint + minimal chrome + honest attribution + optional integrity attestation in response headers) (ef11394). Adoption section also recommends hosts publish /llms.txt indexing their hosted capsules (db49dad).
• domain.exploration_map schema: image-fallback as a required convention with HTML template (9bdb2db).
• spec/DOMAIN_CAPSULES.md "Idea queue" section + first entry domain.music_stems — pre-proposal home for domains noodled on but no working example yet (29957f8).
• GLOSSARY entry for "Interactive archive" (56b5b7b).

Added — research record

• F20 — First publicly-fetchable Mintel production capsule (Copper Dome, 13.7 MB) validates spec at scale; five empirical findings around it (9bdb2db).
• F21 — Independent convergence on the host-contract pattern (MinDev + htmlbin) (2a21d76).
• F22 — Independent convergence on the live-editing-layer pattern (html-docs.com + workplane.co) (1f298cf).
• F23 — URN-not-URL QR encoding empirically validated by a host-side visibility-tier removal (MinDev dropped public, breaking Mintel-encoded URL QRs for anonymous scanners overnight). Also added "Visibility tiers as host-side policy" section to spec/HOSTING.md naming this case.
• F24 — Refinement of F23. Introduces the host vs. registry distinction: a host serves capsules; a registry is a host that commits, publicly, to keeping its behavior stable (URL stability, visibility honor, deprecation discipline, attestation headers, immutability). F23's URN-as-default is correct for producers without signal about host commitments; producers with signal (target host has declared registry compliance) can encode URLs as a calibrated bet against a published contract. Capsule Registry Compliance v1 sketched in spec/HOSTING.md as the optional commitment layer hosts can opt into — including a proposed well-known declaration shape at <host>/.well-known/capsule-compliance.json. Not yet adopted by any host; MinDev / htmlbin / the maintainer's planned personal host are natural first candidates. No Core spec change; format stays agnostic. Synthesis came from maintainer pushback on F23's first framing — a research-method observation worth tracking.

Added — infrastructure & citability

• CITATION.cff at repo root — makes the project formally citable via GitHub's "Cite this repository" button; consumed directly by Zotero / citation managers (29fbc9e).
• /llms.txt at site root — Chrome Lighthouse-style site-discoverability index per the Jeremy Howard / Answer.AI proposal (db49dad).
• README "How to cite" section pointing at CITATION.cff (29fbc9e).
• Validator URL mode — python3 compiler/validate.py <https://host/.../raw> fetches the body via the host's raw endpoint, cross-checks any x-capsule-content-hash / x-capsule-uuid response headers against the manifest, then runs the standard 26 checks (ef11394). Real-world tested against the live MinDev Copper Dome capsule.

Added — voices archive

• /voices/ directory established as a first-class archive for primary-source voices in the conversation Capsule is part of.
• voices/karpathy-html-progression-2026.html v1.0.0 (UUID e8a4c1d2) — Karpathy's "structure your response as HTML" X post + LLM-wiki gist + four-step progression (02530e9).
• voices/utsengar-htmlbin-2026.html v1.0.0 (UUID b3d8f2a1) — Utkarsh Sengar's htmlbin.dev, the hosting-layer voice (ef11394).
• voices/README.md — directory index + queue-tracking pattern (parallel to the Idea queue in DOMAIN_CAPSULES); graduation rule for queued → archived (5ea0296).

Changed — maintainer attribution

• CITATION.cff authors: block — B. F. Garden placeholder replaced with the project's actual maintainer: Luke Schuss (Vancouver, CA) with email [email protected] and website https://lukeschuss.com. Citation form is what Zotero / GitHub's "Cite this repository" button consume, so this fixes the citability surface first.
• README.md — new "Maintainer" section at the end with name, location, website, and email; invites independent-implementation reports and producer/consumer feedback.
• Notes essay byline updated in both the data block and rendered byline (linked to lukeschuss.com); essay bumped to v1.2.0.
• Landing-page colophon — adds "maintained by Luke Schuss · Vancouver" line; also fixes a stale static colophon version string (v10.11.0 → v10.13.0) carried over from v10.11→v10.12 ship. Landing bumped to v10.13.0.
• GitHub repo description updated to lead with "HTML Capsule" and include "Maintained by Luke Schuss" (visible on the repo's main page).

Changed — landing page (v10.2 through v11.0)

The landing went into hero framing exploration mode at v11.0.0 — the conventional hero/four-questions/voices/footer structure (v10.x) was replaced with eight numbered candidate hero variants stacked vertically, each with its own framing (discipline / outcome / substrate question / contract / max-short / PDF comparison / mark-first / problem question), its own layout (left-aligned, centered, dark inverted, mark-led), and its own background color. Numbering is for reference; the winning framing will become the page and the rest will be removed. The v10.15 landing structure is preserved in git at commit 94c3802.

Twelve incremental v10.x iterations preceded the v11.0 reset. Highlights:

• v10.2 — "File over app, taken further" aside citing Steph Ango's principle (5457f5b).
• v10.3 — Hero pivoted from narrow single-column stack to horizontal two-column split (title left, sub/CTA/stats right) (5457f5b).
• v10.4 — Adjacent Voices panel added with 4 entries: Thariq, Karpathy, Blake, Simon Willison (6fe2d56).
• v10.5 — Karpathy archive link from the panel (02530e9).
• v10.6 — Two-beat bridge: "The durable unit of knowledge is shifting from the note to the artifact." + "So HTML is the substrate. Now what?"; "Tools, not apps." aside naming the interactive-archive category (56b5b7b).
• v10.7 — Utkarsh / htmlbin added (hosting-layer voice) (2a21d76).
• v10.8 — Jeremy Howard / llms.txt added (discovery-layer voice) (db49dad).
• v10.9 — Raunaq Bhutoria / html-docs.com added (first live-editing-layer voice) (40f7238).
• v10.10 — Adjacent Voices panel reorganized into two layer-tagged sub-sections: "Format & substrate observers" and "Layers Capsule composes with" (5ea0296).
• v10.11 — Matan / Workplane added (second live-editing-layer voice); convergence with html-docs documented as F22 (1f298cf).
• v10.12 — Top bar tightened: brand reads "HTML Capsule" (full project name) instead of "Capsule"; small linked version badge (v0.3.4) replaces the inline "open spec v0.3.0" subtitle; multi-link primary nav collapses to a single GitHub icon-link with accessible label. Page <title> and meta description updated to match. New mobile breakpoint at 600px hides the "GitHub" label (icon-only on phones), tightens topbar/stat spacing, and shrinks the badge; a 380px breakpoint stacks hero CTAs full-width. Stale data.spec.full value in the landing's data block bumped from v0.3.2 → v0.3.4 to match the current full-spec version. Same logical landing capsule (UUID 7d1a1ac8 preserved).
• v10.13 — Maintainer attribution: footer colophon now reads "maintained by Luke Schuss · Vancouver" (linked to lukeschuss.com); also fixes a stale static colophon version string (v10.11.0 → v10.13.0) per Rule 12 pre-rendered-content discipline.
• v10.14 — First real brand mark: the placeholder "C" glyph in the top bar is replaced with an inline-SVG horizontal capsule outline containing the </> code symbol — a literal rendering of "HTML in a capsule" (concept from ChatGPT image generation; redrawn as SVG for crisp scaling and currentColor integration). Monochrome, picks up --ink automatically. 36×20 on desktop, 30×17 on phones via the 600px breakpoint.
• v10.15 — Brand-mark geometry locked in: capsule outline expanded to fill the full viewBox (rect 30×16, rx 8), stroke weight tightened 2.2 → 1.8 for more refined linework matching the source design more closely, brackets pushed further out (< apex at x=6, > apex at x=26) so the code symbol breathes inside the capsule. Same SVG shape, same currentColor integration, dialed-in proportions.
• v11.0 — Page restructured to hero framing exploration mode (major bump). Topbar + small explorer intro + 8 numbered candidate hero variants stacked vertically + slim footer. Each variant has its own framing tag, layout treatment (left / centered / centered+dark / mark-first), and background color (paper / cream / cool gray / pale blue / ink / sand / soft blue / paper-deep). Variant 05 is dark-inverted to break the visual rhythm. Big mark variant (07) uses a 220px-wide inline SVG render of the brand mark. Data block restructured from the multi-section hero/bridge/questions/voices/journey shape to a flat intro + variants[] shape; runtime buildMarkdown regenerated. #capsule-root CSS loosened (no longer max-width-constrained) so frame backgrounds can bleed full-width; child blocks use new .cb-shell class for max-width centering where needed. Same UUID 7d1a1ac8 preserved; validator passes 25/25.
• v11.1 — Variant 09 added (context-led · plain-English on-ramp) — three-paragraph hero with explicit examples of what AI tools produce (reports, maps, demos, dashboards, visual notes, interactive documents), the spec definition stated plainly, and the "It is not a new file format" disclaimer made the explicit second-paragraph beat. Closes with ONE FILE. NO NETWORK. BUILT TO LAST. as a mono-uppercase accent tagline before the CTAs. Drawn from an external review's recommendation, which observed that the v11.0 variants spoke too much to readers already inside the format's vocabulary (discipline, substrate, artifact). The near-white warm background (#fbfaf5) sets it apart from the other eight without going to ink contrast. Also: variant 06 swap Like a PDF, but alive. → Like a PDF, but inspectable. (more precise per the same review). Frame labels updated 0X / 08 → 0X / 09 and intro text updated Eight → Nine. New .cb-frame-tagline class for the punctuation-line treatment.
• v11.2 — Variant 00 added at the TOP of the gallery — an observation-led preamble designed to sit above whichever hero gets picked, rather than replace one. Opens with Observation: HTML has emerged as a key substrate* for AI-assisted work. (with a self-aware italic footnote about "substrate" being overused), then renders two voice blockquotes as the evidence — Thariq Shihipar on markdown becoming restricting under more-powerful agents, and Andrej Karpathy on telling LLMs to "structure your response as HTML" — then a pivot leadin: So the question becomes: what do you do with all that HTML?. Editorial/research-paper voice rather than product-pitch voice; quotes the existing Karpathy archive (voices/karpathy-html-progression-2026.html) directly. Pure-white background (#ffffff) for editorial weight. Label format 00 · preamble (not 0X / NN) signals it's a different kind of candidate from the nine heroes. CSS: new .cb-frame.is-observation modifier, .cb-quote figure/blockquote treatment with accent left-border, .cb-frame-footnote / .cb-frame-lead / .cb-frame-leadin for the editorial bones. Explorer intro updated to acknowledge the preamble+heroes split.
• v11.3 — Removes the meta Hero framing exploration · Picking the right framing. explainer section that used to sit between the topbar and variant 00. The page now opens directly into editorial-voice variant 00 instead of a process-page meta-intro. Adds variant 00b (Lived practice · maintainer's own use) — a paired case-study preamble that opens with Thariq's FAQ answer about viewing HTML files (open it locally, or upload to S3 for sharing), then carries the maintainer's first-person account of building map-making tooling that emits HTML conforming to the Capsule spec for clients/colleagues. Explicitly acknowledges the file-size tradeoff: "Does the file size creep up? Yes — but frankly, I don't care." Signoff: — Luke Schuss · maintainer · Vancouver · lukeschuss.com. Soft-cream background (#fcfaf3) immediately below 00's editorial white — keeps the pair visually connected as a unified opening before the hero gallery begins at 01. Variant 00's CTA changed from Scroll the framings ↓ to Take it further ↓ (now scrolls to 00b instead of 01). CSS: new .cb-frame-bridge / .cb-frame-prose / .cb-frame-kicker-line / .cb-frame-signoff classes for the case-study composition.
• v11.4 — Tightens the opening pair (00 + 00b) and the topbar. Drops the meta framing-tag rows that read Observation-led preamble · designed to sit above any hero and Lived practice · maintainer's own use from variants 00 and 00b — the corner labels 00 · preamble / 00b · case study carry that signal sufficiently. Frame padding for the preamble pair tightened to clamp(40px, 5vh, 64px) (down from the heroes' clamp(72px, 11vh, 128px)) with min-height: 0 so they're sized to their content rather than viewport-height-filling. Variant 00 title font shrunk to clamp(24px, 3vw, 36px) (down from clamp(28px, 3.6vw, 42px)). Variant 00 CTA changed from a heavy primary button to a quieter mono text-link with arrow (.cb-frame-jumplink). Variant 00b bridge font shrunk to clamp(22px, 2.8vw, 30px) (down from clamp(26px, 3.4vw, 38px)). Blockquote/prose/signoff sizes and spacings all dialed down. Topbar padding-top (in .cb-shell-top) shrunk 36px → 20px, padding-bottom (in .cb-topbar) shrunk 24px → 14px. Voice quotes also fixed: Thariq's quote now uses proper … Unicode ellipses at the two elision points (after communicate with us and after restricting format) rather than periods that masked the elisions; Karpathy's pull-quote uses curly typographic “” around the inner structure your response as HTML phrase and is bracketed with leading + trailing … to signal it's a mid-sentence extract from the canonical archive at voices/karpathy-html-progression-2026.html (which has the full Karpathy passage). Result: opening pair now fits in roughly one viewport-height instead of two; topbar reads as a compact header.
• v11.5 — Restructures the opening into a clearer narrative arc: Observation → Question → Case study, then heroes. Variant 00 (observation) drops the So the question becomes… leadin and the Take it further ↓ jumplink — the section now ends after the second voice quote. NEW variant 00b (question, same white background) titled Question: What do you do with all that HTML? carries two existing answers in the wild: Answer A is Thariq's FAQ quote (open it in a browser locally, or upload to S3 for a shareable link), moved up from the prior case study; Answer B is Utkarsh Sengar's htmlbin.dev (short URLs + /raw endpoint for agent-emitted HTML; independent convergence with the MinDev hosting pattern, format itself stays hosting-agnostic; links to the existing voices/utsengar-htmlbin-2026.html archive). Existing case study renumbered 00b → 00c (soft cream background unchanged), reframed with kicker Or — the maintainer's answer: so it reads as a third answer after the two above (rather than as an arbitrary "lived practice" beat). Corner labels updated for parallel naming: 00 · observation / 00b · question / 00c · case study. New .cb-answer / .cb-answer-tag / .cb-answer-h / .cb-answer-body CSS classes for the answer-block treatment with a thin --rule-soft separator between answers. Reading flow: editorial-voice white pair (00 + 00b) → cream shift (00c, maintainer voice) → hero gallery begins at 01.
• v11.6 — Expands Answer A and Answer B in the Question section so they actually do landscape analysis instead of just naming the answers. Answer A keeps Thariq's FAQ blockquote and gains a body paragraph naming what the view-locally-or-S3 approach offers (minimum-viable, portable substrate) and what it leaves open (no default shareable URL — exactly the friction Answer B is designed for) — bridges to B. Answer B now leads with htmlbin.dev's own positioning verbatim (API for agents to share HTML. Agent-native, end to end. Drop HTML — get a public URL.) attributed to htmlbin.dev positioning · Utkarsh Sengar (@utsengar). Two body paragraphs follow: (1) what htmlbin does — purpose-built public hosting layer, May 2026, OAuth device-code first-publish flow, /p/{slug} URLs with byte-identical /raw endpoint, minimal host chrome + honest authorship attribution, no format discipline imposed; (2) why htmlbin matters for the project's landscape analysis — htmlbin and MinDev independently converged on the same hosting pattern, which is empirical evidence that the format/host split is real, not theoretical (hosting is one job, format discipline is another, they compose; valid Capsules can be served by htmlbin, MinDev, or self-hosted). Cross-refs to RESEARCH.md F21, spec/HOSTING.md, and the full voices/utsengar-htmlbin-2026.html archive. Answer tag labels gained slot suffixes: Answer A · viewing strategy / Answer B · hosting layer. Section lead changed from A couple of answers in the wild: to Two existing answers — different shapes, same question:. CSS fix: .cb-answer-body now has margin: 0 0 14px with :last-child { margin-bottom: 0 } so multi-paragraph answers breathe properly between paragraphs (Answer B's two-paragraph analysis no longer runs flush).
• v11.7 — Stronger visual separation between the opening sections, and answer blocks become cards. Variant 00b (Question) background changes white → var(--paper-soft) (#eceef2 cool grey), so the opening trio now reads as three distinct visual zones (white observation → grey question → cream case study). The Answer A and Answer B blocks inside 00b are now white cards lifted out of the grey backdrop — 8px border-radius, 1px var(--rule) border, 28px/32px padding. The thin-rule-separator pattern (between answers) is replaced with 22px gap between cards. Answer typography bumps: heading 19px → 22px, body 15.5px → 16.5px, blockquote-inside-card 15.5px → 17px. Blockquotes inside answer cards lose their own background tint (which would have read as nested boxes inside cards inside a section); just the accent left-border now marks them as quotes. Mobile breakpoint adds tighter card padding (22/20) and proportional type-size reductions so cards don't blow out at 375px. Result: the Question section reads as a clearly-different kind of content from the surrounding sections — analytical / inventory-of-answers — rather than just more editorial prose.
• v11.8 — Research-paper numbering taxonomy applied across the opening sections. Corner labels switch from NN · kind to bare taxonomy IDs: Observation 1 (was 00 · observation), Question 1 (was 00b · question), Answer 1c (was 00c · case study). Section H2 kickers gain the number: Observation 1: / Question 1:. Answer-card tags inside variant 00b update: Answer A · viewing strategy → Answer 1a · viewing strategy; Answer B · hosting layer → Answer 1b · hosting layer. Variant 00c (the maintainer's case study) reframes as Answer 1c rather than a separate "case study" beat — same prose content and cream background, but the kicker line is now a parallel .cb-answer-tag reading Answer 1c · custom producer + registry · maintainer's own answer, matching the mono-accent uppercase treatment of 1a and 1b inside the cards above. Question 1 lead now mentions all three answers and links forward to 1c via in-page anchor: "Three answers, different shapes — two in the wild, one from the maintainer (below in [Answer 1c]):" The numbering scheme is extensible: future Observation 2 / Question 2 / Answer 2a/2b/2c slot in naturally as the landscape analysis grows. data-num HTML attributes stay (00 / 00b / 00c) for CSS targeting; only the visible labels changed.
• v11.9 — Adds Observation 2 (variant 00d) at the end of the opening narrative — after Answer 1c, before the hero gallery — on the same white background as Observation 1 (signals "back to observation mode"). Title: "Observation 2: File over app is as relevant today as it ever was." Quotes Steph Ango (Kepano)'s "File over app" essay verbatim: "If you want to create digital artifacts that last, they must be files you can control, in formats that are easy to retrieve and read." Attribution links to https://x.com/kepano and https://stephango.com/file-over-app. Followed by maintainer's commentary on Obsidian's plugin route: well-intentioned but undermines the principle — when presentation requires plugin installs, you're back at app over file (italic emphasis on the reversal). Numbering scheme extends naturally: each observation now gets its full arc (Observation 1 + Q1 + Answers 1a/1b/1c, then Observation 2) before the next one. No Question 2 yet; would slot in naturally if/when one becomes useful.
• v11.10 — Observations / Questions / Answers each become their own KIND with a distinct color and parallel section formatting. Backgrounds now driven by KIND class on the <article> element instead of per-data-num CSS: .is-observation → white (#ffffff), .is-question → grey (var(--paper-soft) #eceef2), .is-answer → cream (#fcfaf3). CSS specificity is equal across the three KINDs, so later rules win the cascade — a section with both .is-observation and .is-question resolves to grey; a section with both .is-observation and .is-answer resolves to cream; plain observation sections stay white. Restructure: Answer 1a (Thariq · viewing strategy) and Answer 1b (htmlbin.dev · hosting layer) — which used to be white CARDS nested inside the Question 1 section on grey — are now extracted into their own sibling sections (data-num="00b1" and data-num="00b2"). Question 1 becomes lean: just the question + a TOC line linking to the three sibling Answer sections (1a / 1b / 1c). Answer 1c (the maintainer's case study) gains the .is-answer KIND class alongside its existing .is-case-study sub-modifier, so it sits on the same cream as 1a/1b but keeps the italic-emphasis treatment on its title phrase ("Maybe let's take it a step further."). All three Answer sections now share the same shape: corner label (e.g., Answer 1a), slot-tag eyebrow (e.g., — VIEWING STRATEGY in mono accent), H2 with Answer 1x: kicker + the heading sentence, then content (blockquote and/or prose). Dropped: the .cb-answer card CSS (cards no longer used) and the .cb-frame-bridge rule (Answer 1c now uses the standard title pattern with <em> for italic). Visual rhythm of the opening: Obs1 white → Q1 grey → A1a/1b/1c cream-cream-cream → Obs2 white → hero gallery (01 paper).
• v11.11 — Extends the analysis with two new arcs (six new sections total) inserted between Observation 2 and the hero gallery. (A) HTML version-control arc — Observation 3 (00e, white) quotes Thariq's FAQ note that "HTML diffs are noisy and hard to review compared to Markdown" and reframes a related, more general problem: artifact identity (the PDF v1 / v2 / v3-final / v3-real-final-USE-THIS problem people already know). Argues identity and lineage are addressable even when diffs aren't. Question 2 (00f, grey) — "How do you keep track of which version of an artifact you're working with?" — with TOC to two sibling Answers. Answer 2a (00f1, cream) — filename versioning (what people actually do; works because the version lives outside the file; breaks when the filename is dropped). Answer 2b (00f2, cream) — the spec's uuid + capsule_version + parents[] identity/lineage fields recorded inside the inline manifest, so the artifact knows what it is and what it descended from regardless of filename. Explicitly acknowledges HTML-diffs-proper are still hard — a semantic diff over manifests + data sections would be tractable but isn't built. (B) Editor-layer observation — Observation 4 (00g, white) names three projects in different shapes of the live-editing layer emerging around AI-emitted HTML: Claude Design (Anthropic Labs, launched April 17, 2026 — chat-and-preview HTML/CSS/JS generation), html-docs.com (Raunaq Bhutoria — agents publish, humans comment, agents revise async loop), Workplane (Matan / matanrak — MCP-first, multi-agent, same shape as html-docs). Maintainer commentary: Capsule sits downstream as the seal step; format is editor-agnostic; whether the project ever grows its own collab layer is open, and if it does it would be a registry concern (lightweight commenting on hosted Capsules), not a format concern — same posture as the format/host split. All six new sections use the KIND-class color system (.is-observation white, .is-question grey, .is-answer cream) from v11.10. No new CSS needed.
• v11.11.1 — Attribution cleanup: Answer 1b no longer names MinDev (MinDev belongs to Answer 1c — the maintainer's own custom producer + registry approach — and was previewing too early in 1b). 1b's convergence-evidence paragraph reframed with vague language ("Multiple independent hosting layers have now converged on the same minimal shape...") + the existing F21 / spec/HOSTING.md cross-refs for readers who want the named specifics. Answer 1c's slot tag updated Custom producer + registry · maintainer's own → Custom producer + MinDev registry · maintainer's own so MinDev is explicitly named where it belongs taxonomically.
• v11.12 — Two additions: a new Question 1 upstream of the existing arcs, plus a new Answer 3c on spec-evolution for diffs. New Question 1 (00a, grey) — "How do I get HTML out of my LLM?" — inserted between Observation 1 and the previous Question 1 (which becomes Question 2). Two answers: Answer 1a (00a1, cream) — "Just ask the LLM for HTML" — quotes Karpathy ("structure your response as HTML") and Thariq ("I've started preferring HTML as an output format") as two voices both pointing at the same vague-but-functional default; commentary notes you get unstructured HTML (no manifest, no identity, no integrity, no contract) — fine for one-off viewing, lossy otherwise. Answer 1b (00a2, cream) — "Paste the Capsule Core spec into your prompt" — the contract-based approach: ask for a Capsule (not just HTML); the Core spec is one page (twelve numbered rules) designed to paste into an LLM prompt; the LLM returns a sealed .html with manifest, integrity hash, UUID + parents[], declared capabilities (Rule 7), pre-rendered content (Rule 12), no network deps (Rule 2). Includes the "producer becomes a first-class spec-aware participant" reframe and pointers to CAPSULE_CORE.md, full spec, examples. Renumbering follows: existing Question 1 → Question 2 (what do you do with HTML); existing A1a/1b/1c → A2a/2b/2c; existing Question 2 → Question 3 (how do you track versions); existing A2a/2b → A3a/3b. New Answer 3c (00f3, cream) — "A Capsule-aware diff over manifest + data + content, layered on the spec" — addresses the user's note that the prior treatment didn't talk about how the spec could evolve to help with diffs. Sketches what a capsule-diff tool would do: semantic diff over each of the five required blocks (manifest field-level JSON delta, data JSON delta, visible-content text delta ignoring presentation, bounded style/runtime diffs). Includes example output. Frames the tool as a layer on the spec (analogous to git log --oneline on git's commit graph), not a spec change — the spec's existing structural discipline is what makes a semantic diff tractable. Not built yet; parked in spec Appendix E. Existing Answer 3b's closing paragraph updated to forward-point at 3c instead of just acknowledging diffs are hard. data-num values stay stable for renumbered sections (00, 00b, 00b1, 00b2, 00c, 00f, 00f1, 00f2); new data-num values are 00a / 00a1 / 00a2 (new Q1 + answers) and 00f3 (new A3c). All anchor links in TOC lines updated.
• v11.13 — Strips the inline TOC .cb-frame-lead paragraph from all three Question sections (Q1, Q2, Q3). The answers naturally follow in their own sibling sections below, so the explicit "Answer Na — slot, Answer Nb — slot..." pointer was redundant chrome. Question sections are now lean: just the corner label, the H2 with kicker + the question itself, nothing else. Each Question reads as a single beat that opens its arc; the Answer sections below do the rest.
• v11.14 — Three changes. (1) Question 3 title tightened: "How do you keep track of which version of an artifact you're working with?" → "How do you keep track of which version of an HTML file you're working on?" — more concrete subject (HTML file, not generic artifact), more direct phrasing. (2) Observation 3 gains a middle paragraph about chat-hosted artifacts: Claude's artifacts, ChatGPT's canvas — they don't have a natural versioning system, the artifact mutates in place as you iterate, forking into a new conversation doesn't inherit a clean version label or a "this is v3" anchor; the chat keeps going and the artifact loses its history. Sits between the existing PDF-v1/v2/final analogy paragraph and the "identity and lineage are addressable" closing paragraph. (3) NEW Observation 5 (00h, white) added after Observation 4 and before the hero gallery: "There's no easy way to get chat-hosted content out of an LLM platform." Three prose paragraphs — share links exist on every platform but are platform-locked viewing artifacts, not portable; cross-platform interop is poor (a ChatGPT share link handed to Claude doesn't compose: the receiving LLM might fetch and read the URL but doesn't get the structured conversation); the file-as-artifact discipline (Obs 1, Obs 2, Answer 1b cross-linked) is the project's response — a Capsule opens in any browser, can be handed to any LLM, the host doesn't need to be in the loop. "That's the practical difference between 'I shared a link' and 'I shared a file.'"
• v11.14.1 — Answer 1b retitled: Paste the Capsule Core spec into your prompt. → Ask the LLM for a Capsule (according to the Spec). Sharper phrasing — focuses on the action the producer takes (asking for a Capsule) rather than the implementation detail (pasting the spec text). The body of Answer 1b already opened with "Instead of asking for HTML, ask for a Capsule…", so the new title aligns with the body's framing. Slot-tag eyebrow updated to match: Contract-based approach · paste the Core spec → Contract-based approach · ask for a Capsule. Link target unchanged (still CAPSULE_CORE.md; link text shortened "Core spec" → "Spec").
• v11.15 — Sharpens the HTML version-control arc (Observation 3 → Question 3 → Answers 3a/3b/3c) per two converging external reviews. Observation 3 retitled "HTML diffs are noisy. Lineage is the bigger problem." and rewritten around the explicit three-layer framing (raw diff / semantic diff / lineage), rendered as an ordered list. Opens with the runtime-vs-writing-format explanation ("Raw HTML diffs are noisy because HTML is closer to a runtime artifact than a writing format") — explains why diffs are noisy, not just that they are. PDF v1/v2/final analogy and chat-hosted-artifacts paragraph preserved. Closes on "Capsule doesn't claim raw HTML diffs become pleasant. It claims artifacts need identity and lineage first." Answer 3b gains a closing punchline: "Git tracks files. Capsule tracks artifacts." (mono accent on the subjects). Answer 3c rewritten with a new opener punchline "The HTML is the artifact. The manifest is the review surface." followed by an 8-row review-surfaces table (Raw HTML / Manifest / Content / Data / Style / Asset / Render / Runtime diff) replacing the prior single-paragraph block description. Adds Asset diff (images, fonts, media by hash) and Render diff (screenshot or DOM snapshot before/after) — two surfaces that weren't in the prior treatment. Example-output code block preserved. Closing "layer on the spec, not a spec change" framing kept. New CSS: .cb-frame-list (ordered list for the three-layer breakdown), .cb-frame-punchline (the two slogan lines, accent-bold on the contrasting noun), .cb-diff-table (8-row review-surfaces table with mono accent column headers and tight mobile breakpoint). Three new manifest field ideas raised by the external review (supersedes, derived_from, change_summary) are NOT added to the spec; parked in spec Appendix E pending real-producer empirical pressure.

Changed — PRECEDENTS.md

• "Current voices in HTML-for-AI" section grown from 2 voices (Thariq, Blake) to 7 voices spanning format, hosting, discovery, and live-editing layers — Karpathy, Steph Ango, Utkarsh / htmlbin, Jeremy Howard / llms.txt, Raunaq / html-docs.com, Matan / Workplane added.
• Position picture: original three-position table → nine-row layered table organized by lifecycle slot (format / live editing / hosting / discovery).

Changed — Notes essay

• Notes v1.1.0 (2026-05-20) — added Obsidian + "File over app" paragraphs after the share-link discontent paragraph (4efb95a).
• Notes v1.2.0 (2026-05-21) — byline updated from "B. F. Garden" placeholder to "Luke Schuss · lukeschuss.com" (linked); data block byline updated to match; description records the attribution change.

Fixed

• Stale §11 hosting discussion cross-reference in Appendix E.7 corrected to point at Appendix B (3068a3b).

Resolved (moved out of parked-items)

• Appendix E.5 — Rule 12 vs. legacy compiler templates: resolved via the image-fallback carve-out in §2.3 (9bdb2db).

[Spec v0.3.3] — 2026-05-20

Size cap raised; image-fallback carve-out for visualization geometry shipped; first major external feedback signal (F19 Claude Design experiment, plus F20 Mintel validation) folded in.

Changed — spec

• §6.3 Size limits: hard cap 15 MB → 20 MB; new 15-20 MB soft-warn tier for email-attachment compatibility (9bdb2db).
• §14 Validation list: item 11 (file size) updated to reflect new cap + soft warn (9bdb2db).
• §16.2 Out of scope: boundary mention updated (9bdb2db).
• CAPSULE_CORE.md cap mention updated to point at the new constraint (9bdb2db).

Changed — validator

• compiler/validate.py: MAX_FILE_SIZE raised from 15 to 20 MB; SOFT_WARN_SIZE constant added at 15 MB; file-size check emits a contextual note when the body is between 15-20 MB (9bdb2db).

Added — RESEARCH

• F20 writeup: first publicly-fetchable Mintel production capsule; integrity hash verifies; MinDev hosting-pattern headers documented; image-fallback carve-out motivated.
• F5 conclusion appended with "Updated by F20" addendum tying the size-cap change to the empirical finding.

[Spec v0.3.2] and earlier — 2026-05-19

Initial public release of bigfancygarden/htmlcapsule to the world. Project work had been ongoing privately before this; the May-19 commits are the cleanup-and-publish phase.

Initial release contents

• Core spec v0.3.0 (CAPSULE_CORE.md): twelve numbered rules, designed to paste into an LLM prompt.
• Full spec v0.3.2 (spec/CAPSULE_SPEC.md): implementer-grade definition; integrity-hash recipe with normative test vector; security model; response protocol; Appendix E with parked v0.4+ candidates.
• Domain schemas (spec/DOMAIN_CAPSULES.md): domain.implementation_notes, domain.design_system, domain.exploration_map, domain.briefing.
• Reference implementation: compiler/compile.py (compiles JSON + template directories → capsule HTML); compiler/validate.py with 25 conformance checks initially.
• Templates: templates/decision_board, templates/news_capsule.
• Examples: spec/examples/briefing_example.html, spec/examples/implementation_notes_example.html.
• Research log: RESEARCH.md with findings F1-F18, methodology, open questions, recurring failure modes.
• Companion docs: GLOSSARY.md, PRECEDENTS.md.
• Initial landing page at htmlcapsule.org (multiple visual iterations between May 19 and May 20 from editorial to terminal to bright to essay to layered).

Post-release v0.3.2-era patches (still pre-v0.3.3)

• F19 added (2026-05-19/20): Claude Design integration experiment + design-tool theme files cataloged (b70f58a, 6f0148b).
• Validator 26th check added (4696367): recognizes the cleaner data-capsule-action="<cap>" + <cap>: function Rule 7 verification convention surfaced by external producers; no spec change.
• README expansion clarifying capsules ≠ PKM after independent reader confusion (8c83a99).
• Landing v7.0.0 → v7.1.0 repositioning + "PDF of the interactive web" tagline.
• Appendix E.8 parked for validator refinement of non-resource-loading <link> tags (566b046).

Notes essay

Parallel versioning. Tracked here only at major beats.

• v1.0.0 — 2026-05-20 — split from landing v10.0 (essay-as-landing experiment) into its own capsule (UUID c5f6a890); parents[] records lineage to landing UUID 7d1a1ac8.
• v1.1.0 — 2026-05-20 — added Obsidian + "File over app" paragraphs.

How this changelog is maintained

Updated when each spec version ships (v0.3.x patches; future v0.4+ minor releases). Landing version bumps tracked at major revs (v8, v9, v10); intermediate landing changes (e.g. v10.4 → v10.5 → v10.6) appear under the corresponding spec-version section. Notes essay versions tracked when they ship. Voices archive entries tracked when added. Idea-queue movements not tracked here unless something graduates to "Initial domains" or gets formally dismissed — the queue files (voices/README.md, spec/DOMAIN_CAPSULES.md) carry their own state.

When in doubt: a change earns a CHANGELOG entry if a future reader (including future-you) would want to find it without scrubbing git log. Internal copy-tweaks, small cleanups, and prose iterations don't necessarily.