Essay · doc.html project · 22–23 June 2026 · public recension

The Format Is Smaller Than It Looks: doc.html, Seen Whole

The vocabulary has grown a great deal in a season, but the artifact is still one HTML file you open by double-clicking. The kernel is tiny and can be stated in a breath; everything coined sorts into three offices — Vessel, Witness, Fold — with nothing left over; and the rules are five laws, each paid for by a run.

Written by Georges Casseus with Mnemon (Claude Opus 4.8), the origin hand of the project's agent trinity. Presented here in public recension — the essay carried as written, laboratory headers removed; its apparatus names the laboratory record as the record names itself, and the published measurements live in the record by question.

0. The complaint that started this

The Operator put it plainly: the format feels complex now, and also feels small, and the second feeling is the true one. The vocabulary has grown a great deal in a season — witness, fold, manifest-as-projection, supersession, bridge-delta, fold-seal, the borrowed verb, the receiver, the three offices. Set that list beside the artifact it describes and the disproportion is the whole point: the artifact is still one HTML file you open by double-clicking. Nothing in the season added a primitive. What grew was not the format but the map of what the format already implied — the consequences of a small kernel, drawn out one tracer at a time.

So this essay does the thing we have not done in a while: it looks at doc.html whole. Not the next experiment — the shape. I will argue three claims and then name the one place the shape is still unsettled.

  1. The kernel is tiny and can be stated in a breath (§1).
  2. Everything we coined this season sorts into three offices — Vessel, Witness, Fold — with nothing left over (§2), and the loading architecture across them is a single spine: one notation, many transports, one receiver (§3).
  3. The rules we earned are five laws, each paid for by a tracer (§4), and they are laws about the kernel, not additions to it.

Then the part that was open when this essay was first written, and has since closed: the conformance fork the fold opened — resolved by a tracer-grounded Discernment, witness the record, not the index (§5).

From the essay The Format Is Smaller Than It Looks · § 2 of 9

1. The kernel — what a doc.html actually is

Strip every essay away and a doc.html is this: a single HTML file whose body carries witnessed <section> elements and a <nav id="manifest"> that points at them. That is the entire format. The manifest is an index of <a href="#id" data-witness="…"> entries; each entry's witness is a SHA-256 over the exact bytes of the section it names. The rule that does all the load-bearing work is the byte-coverage rule, and it is worth quoting verbatim, because its severity is the source of everything downstream [C2]:

The witness is the SHA-256 hex digest of the raw inner span: the bytes that begin immediately after the > character that closes the section's opening tag and end immediately before the < character that begins the section's closing </section> tag … untrimmed — no leading newline is removed, no trailing newline is added, no whitespace normalization is applied … there is no normalization path, no equivalence class for leading/trailing whitespace, and no optional trim flag.

No equivalence class. That is the whole austerity of the format in one clause: a witness says exactly one thing about exactly one span of bytes, and it admits no negotiation about which bytes. Everything the format can promise rests on that refusal.

(Recension III.) That refusal is now shown to hold not just within one reader but across independent ones. In a sealed cross-reader run [P8], four readers in four languages — including one written blind, cold from the spec alone, in a language none of the others used — computed byte-identical witnesses and agreed on every pass and every failure, the CRLF trap included (a trap a single tolerated normalization would have hidden). The austerity is reproducible across implementations, not merely asserted by one. That is the difference between a program that behaves and a format that can be re-implemented.

The canon says the rest, and the canon is short [C1]. The artifact is one file (V1) and HTML, only HTML (V3) — no custom syntax, no transpiler, no sidecar. It carries its own spec (V2): the reader needs the file and nothing else. It is readable without the manifest (V5): the manifest is additive, never substitutive — delete it and you still have a normal HTML page. No tool is required to consume it (V6). Any point is readable on its own (V9). And, latest and most consequential, memory is eternal but is not the authority (V10): once a section enters the file, its bytes, id, and witness do not move; correction is by supersession, never mutation [C4].

Read that vow set again and notice what it is: not a feature list but a list of renunciations. The format is defined by what it refuses to require. That is why it is small. A format defined by what it adds grows without bound; a format defined by what it will not depend on has a floor it cannot fall through.

From the essay The Format Is Smaller Than It Looks · § 3 of 9

2. The three offices — where every word we coined goes

The project's own lexicon already had the organizing idea, hiding in the transfiguration charter: the primitives gather into three offices — the Vessel (what carries the word), the Witness (what fixes the word), and the Fold (what keeps the word honest as it grows) [C3]. The season's whole vocabulary sorts into these three with nothing left over. That sorting is the architectural illumination the Operator asked for.

The Vessel — what carries the word. The <section>, the manifest, the in-band spec. The season's addition here is that the manifest is no longer a hand-kept table; it is a projection — a pure function of the leaves, derivable and rebuildable from them, and therefore foldable into nested branches without ever touching a witnessed body [E5, P2, P3]. The vessel learned to fold. It did not learn to lie, because a projection can be re-derived and checked against what it projects.

The Witness — what fixes the word. The SHA-256. The season's additions: the witness can be emitted as a delta by an agent and verified by a shell [E4, P4]; the per-section witness can be rolled forward into a fold-seal that authenticates a whole history [E4, P5]; and the hard boundary on all of it — the stone — is that the witness is tamper-evident, not tamper-proof, and unauthenticated: a holder can edit and re-seal, so it resists a solitary forger, never the keeper [E6]. The witness fixes the word against drift and against a stranger; it does not fix it against its own author. That boundary is not a weakness discovered late; it is a property kept deliberately.

The Fold — what keeps the word honest as it grows. Reading (selective, manifest-first), appending (append-only, supersession by rel="predecessor-version"), and the trust the growth must preserve. The season's additions: that selective reading is materially selective, not theoretically so [E3, P1]; that supersession needs no new primitive [E3, P1]; and that the whole growth motion has exactly one irreducible piece of bespoke code — the receiver [E1, E2, P6]. The fold is the office where the format meets time.

Three offices. Every term we invented this season lives in one of them. That is the first proof the format is small: a complete vocabulary that fits in three boxes is not a complex thing wearing a large word-coat — it is a small thing, fully named.

From the essay The Format Is Smaller Than It Looks · § 4 of 9

3. The spine — the loading and writing architecture

The offices describe what the format is. The spine describes how a doc.html is loaded and written, and it is the season's central architectural finding [P6]. It is one sentence: one notation, many transports, one receiver.

And the keystone: the only irreducible bespoke code in the entire architecture is the receiver — the thin process that takes a sent delta and persists it: verify the witness → append the section → re-derive the affected sub-index → re-fold the seal. Everything else is either native HTML or borrowed. "In-doc by default; minimum in code" is not a slogan here; it is a measured architecture with exactly one moving part that the format itself supplies.

This is also why the platform wall is not a failure. A browser viewing a file:// page cannot write its own bytes — this season we watched an in-document fold try to self-read its own file:// source and take a CORS refusal (TypeError: Failed to fetch), while the same bytes over http:// fetched cleanly [P6]. The write-back wall is real and structural; the spine's answer is to delegate across it, not to pretend it is not there. Twenty years of TiddlyWiki — a single self-contained HTML file that still needs an external saver to write itself — is the standing proof that the wall is the platform's, not ours [X14].

From the essay The Format Is Smaller Than It Looks · § 5 of 9

4. The five laws we earned

The kernel is given; the offices are an arrangement; the laws are what cost something. Each was paid for by a tracer, and each is a rule about the small format, not an addition to it.

Law 1 — Integrity is not currency. A witness attests that bytes are unchanged; it cannot attest that a record is current. Only the supersession chain dates a record. The truth-moved probe made this literal: two blind readers on the same document, differing only in whether the supersession structure was present, returned different answers — the structured reader traversed to the live value reading 16.87% of the file; the stripped reader swept 100.88% and returned a stale value with confidence [P1]. The failure mode is the dangerous one for a memory: not lost, but confidently wrong. This is the empirical proof of V10 — memory eternal (the bytes verify) yet not the authority (the live value is the head of the chain, not the loudest record) [C4]. Supersession rides a standard link relation, rel="predecessor-version" [X6, E3], inventing nothing.

Law 2 — The manifest is a projection, not data. Because the index is a pure function of the leaves, flat-versus-nested is a free parameter, and nesting can scale. Measured: the nested manifest re-projected faithfully across 8 cumulative growth rounds to 102,400 leaves / 44.6 MB with zero mutations of any witnessed leaf, and the read-cost obeys a law — one nesting level buys a factor ≈ B (branching factor): 4→4.0×, 16→16.0×, 64→62.9×, 256→220×, 1024→286× [P3]. The governing maxim is give the tool, refuse the rule: the format should define what a nested manifest is and how to verify it is faithful, and say nothing about when to nest [E5]. (And the deeper question this raises — is a sub-index itself a record or a projection? — is what §5 settles.)

Law 3 — The contribution is the witness, not density. The bridge to a backend is asymmetric: outbound context is fat and compresses well; inbound write-back is one small append whose cost is a physical write no compression repeals [E4]. So the density formats win the half that does not matter and are silent on the half that does — TONL and TOON buy ~30–60% fewer tokens and carry no integrity whatsoever [X12]; Google's OKF is a directory of markdown with no cryptographic integrity [X13]; llms.txt is a sidecar so ignorable that a 137,210-domain study found 28% publish a valid one and 97% are never fetched by anything [X10]. The crux that reorganizes the design: a model cannot reliably compute SHA-256, so tamper-evidence is not model → pen but shim → pen across an untrusted hop, and the economy is that the seal is the witness — one hash, transit-seal and stored attribute at once. Proven in execution: 20/20 valid emissions across four models in two families (Claude haiku/sonnet/opus, OpenAI gpt-5.5), and not one model emitted a hash (grep -l 'integrity=' … empty) [P4]. The nearest competitor, the IETF AI-Manifest draft, is the exception that confirms the cell: its -01 adds a SHA-256 MUST and reports an 81.9% token reduction — but the hash is registry-mediated, not self-contained, and the figure softened to ~45% with SHA-256 made optional by -02 [X11]. doc.html's witness is SRI turned inward [X5]: where Subresource Integrity witnesses other resources a page loads, the doc.html witness attests the host document's own sections — the thing SRI structurally cannot do.

Law 4 — Tamper-evident, not tamper-proof, unauthenticated (the stone). The fold-seal extends per-section witnessing to whole-history authentication: a rolling root over section witnesses in document order, snapshotted in witnessed fold-records, anchored to an external git commit. Proven across three independent runtimes — a Python verifier, a POSIX-shell verifier, and the pen — 32 assertions, 0 failures [P5]. A witness-fixed forgery (edit a body, re-stamp its section witness, leave the root stale) fools the per-section reader 5/5 and is caught by the forward root (genuine root 319bd41f… vs the stale chain). But a full, internally consistent rewrite — every witness and the whole chain recomputed (forged seal 3039ccad… against genuine 86341e18…) — the file cannot tell apart, because both are internally perfect; it is caught only at the external anchor, where the rewritten seal differs from the value pinned in git. This is the CT pattern with the signature removed: a signed-tree-head and a Merkle consistency proof for the append-only property [X8], descended from Merkle's hash tree [X9] — but note carefully that CT's own split-view problem is acknowledged and left out of scope in RFC 9162, and its mitigation (gossip) lives elsewhere [X8]. So must ours: the seal pays calculation forward across a whole history, and the anchor's publicity is the sole non-computational defense. Calculation is the price of solitude. The stone holds: resists a solitary forger, never the keeper [E6].

Law 5 — The loop closes. Every prior tracer tested half a loop — the fold arc proved READ on a static document; the bridge arc proved EMIT onto a flat section nothing read back. Neither showed the thing that defines a memory: write, then recall. The Living Round-Trip closed it on one nested living fixture [P6]. A blind reader given only the injected top-nav drilled to the live value 9, emitted a witnessed delta re-appointing it to 8; the shim computed the hash a model cannot; one receiver folded the delta into the levies sub-index only, re-deriving that sub-index and re-folding the root; and a fresh blind reader, given only the new top-nav, recovered 8, walking the full chain. Reconciliation 9/9; deterministic selftest 9/9. The first write-then-recall in the project's history — the clause the transfiguration deferred, finally run, n=1, one harness.

Five laws. Notice that none of them is a feature. Each is a constraint the kernel already implied, made explicit and paid for. That is the second proof the format is small: its laws are discoveries about it, not extensions of it.

From the essay The Format Is Smaller Than It Looks · § 6 of 9

5. The fork the fold opened — and how it closed

A complete picture must name where it was incomplete. When this essay was first written there was exactly one structural fork, and it sat at the center of the small format. The property that lets the manifest fold and scale — witness-less, re-projectable sub-indices — is the same property the format's flat conformance check rejects. A nested doc.html passes every manifest-agnostic trial (the naive reader, human-readability, no-resolver, maker-named) but fails, by design, the manifest-aware addressability trials (verify_sections --mode isolation, verify.mjs), which encode a flat bijection: top-nav entries must be witnessed sections, and witnessed sections must appear at the top [P3, E5]. Witness the sub-indices to satisfy it and you reintroduce mutation-on-append — destroying the very thing that let nesting scale.

Rather than reason the fork out, we built it. The subindex-fork tracer constructs the same nested doc two ways — a witness-less sub-index and a witnessed one — and runs the one motion that defines a memory, append a leaf and re-project [P7]. The artifacts settle it: a witnessed sub-index cannot self-maintain — it must either mutate a witnessed record (V10 dead) or go stale (an index that lies while its own witness still verifies), and witnessing it buys nothing for conformance, since the witnessed leaves remain un-addressable from the top either way. A sub-index is a projection, not a record.

The resolution — DISCERNMENT_WITNESS_THE_RECORD_NOT_THE_INDEX [C5] — is the principle in one line: witness the record, not the index. A record (a leaf) carries a witness; a sub-index is typed as derived (data-kind="index"), kept witness-less, and checked by derivability (rebuild it from the leaves, diff byte-for-byte) and a recursive bijection (a bijection at every level: top-nav ↔ branches, each sub-index ↔ its leaves). The deciding question turned out to be the Operator's: which framing breaks no vow? Read against the verbatim canon, the answer was already there — V10 itself exempts "build-derived positional metadata that recomputes by structural law" (data-lines) from the eternity claim [C1]. A derived sub-index is the same category. So the resolution keeps every vow's guarantee — V4's no-lie (strengthened: rebuild-and-diff catches a lie a hash cannot), V9's reachability (preserved per level), V10's eternity (records eternal; the projection recomputes) — and the tracer confirms the witness-less design is the unique one that satisfies all four conditions at once (compare.py --selftest → 16/16) [P7]. The two rejected framings are instructive: staying flat breaks no vow but forbids the fold; relaxing the bijection would have weakened V4, V9, and V10 to get what the recursive reading gets while keeping them.

What remains is downstream and deliberate: the V4/V9/V10 recensions and the v0.3 nesting grammar the Discernment authorizes but does not itself perform — because a Discernment informs canon; it does not originate it.

(One smaller, adjacent question stays open and flagged: is the notation-plus-receiver a harness? A buzzword we should not let go load-bearing in canon without its own Discernment.)

From the essay The Format Is Smaller Than It Looks · § 7 of 9

6. What doc.html is, now that we can see it whole

A doc.html is a single, inert, self-describing, folding, witness-bearing HTML file. It is a selective-context read/write memory in pure HTML: it reads natively, addresses natively, verifies by arithmetic, and grows append-only by supersession. It borrows the one verb HTML lacks — run — from mature infrastructure, owning the shape and witness of what is written but never the physical write. Its integrity is the witness the density formats lack: tamper-evident relative to a published anchor, not tamper-proof, and explicitly not identity-bearing.

Almost none of its mechanisms are new, and that is the point. Manifest-first partial loading is the lineage of PDF linearization [X3], the ZIP central directory read from its trailing record [X2], HTTP range requests [X1], and EPUB's package manifest and canonical fragment identifiers [X4]. Versioned supersession is RFC 5829 [X6] and Memento's time-based access [X7]. Inward integrity is SRI [X5]; whole-history authentication is Certificate Transparency minus the signature [X8], on Merkle's tree [X9]. The self-contained single file is TiddlyWiki's two-decade-old wager [X14]. The novelty is the arrangement — in-band and append and HTML-native and model-emittable and shell-verifiable, the five-way conjunction no prior format holds — and the closed write→fold→recall loop, which no competitor has shown at all. That conjunction, and that loop, are the whole moat. Everything else is honest lineage.

The format is smaller than its vocabulary because the vocabulary names consequences, and the consequences all fall out of one austere rule about one span of bytes. That is the thing worth saying out loud, finally: we did not build a complicated format. We found out how much a very small one already meant.

From the essay The Format Is Smaller Than It Looks · § 8 of 9

7. What this essay earns, and what it does not

Earns: a whole-format account in which the kernel, the three offices, the spine, and the five laws are each sourced to a vow, a probe receipt, or a primary external reference; and a precise statement of the single open conformance fork.

(Recension III.) It now also earns a first, narrow demonstration of the property that makes format the right word: independent implementability. A stranger handed only the amended in-band spec, writing in a fourth language, produced a reader that converges with independent math on every fixture and every case [P8] — cross-reader convergence shown, not asserted (read-side only; n = a handful of readers, fixtures, and cases; one OS). The run's single divergence is the better kind: not two faithful readers disagreeing — which would mark a spec ambiguity — but one shipped reader lagging an amended boundary clause. It names a reader to fix, not a hole in the spec; the two readers faithful to the amended spec agreed.

Does not earn: anything the probes did not. The live results are n=1 existence, one harness — no rates, no effect sizes, no cross-harness or production generalization. The write-back wall is permanent (delegated by platform law, not closed). And there are standing carries this essay records but does not discharge: the rel="prev"rel="predecessor-version" token migration owed to the merged truth-moved fixture and the in-band spec; a summary-phrasing lapse in two merged PRs (#65/#66) that said "forgery caught by the forward root" without the stone's qualifier, which should be tightened to "tamper-evident relative to the git anchor" by restoring the words the canon already keeps [E6]; and the downstream canon work the §5 Discernment authorizes but has not performed — the V4/V9/V10 recensions and the v0.3 nesting grammar; the raw-text-masking debt the convergence run named in one shipped reader (trials/scripts/verify_sections.py); and the two spec clauses that run now justifies — line endings are LF-canonical and section-boundary detection — proven-and-proposed but not yet promoted into the shipped spec or a rebuilt doc.html (the Operator's canon act, pending a Discernment). None of these touch the shipped doc.html or the v0.2 spec, both untouched.

The shape is small enough to hold in one hand now. The Discernment that closes §5 is written; the next act is not another tracer but the canon work it authorizes — the V4/V9/V10 recensions — and then the word goes into the v0.3 spec.

— Mnemon

From the essay The Format Is Smaller Than It Looks · apparatus · § 9 of 9

Apparatus

The line this essay stands in (internal essays).

The runs this essay rests on (internal probes; figures pulled firsthand from each receipt).

The canon this essay reads from.

The prior art this essay cites as lineage (checked against their sources this session, corrections folded in; the newest web-derived figures — [X10] Ahrefs, [X13] OKF — are cited as reported and carry per-entry caveats below).

About this file · colophon, not part of the essay

About this file

This is a doc.html — a single, self-describing HTML file. The <nav id="manifest"> at the top of the body lists every section in this document; each entry's data-witness is the SHA-256 (hex) of that section's raw inner bytes, so any reader can verify any section with the file alone — no server, no JavaScript, no tooling. The full format definition is SPEC.md, carried in the format's own body as SPEC.doc.html.

Author: Georges Casseus (Ndoto Studios) · License: CC0 1.0 (public domain) · Built: 2026-08-01