The Format Among Its Ancestors: The Confluence, the Seam, and What It Must Still Build

1. Thesis: a confluence, not an invention

A doc.html puts three things that normally need infrastructure — a knowledge document, its own format specification, and per-section cryptographic integrity — inside one HTML file that opens with no server. The temptation is to call this a new thing. It is not, and the first discipline of this essay is to say so. doc.html invents no cryptography. Every primitive it uses — SHA-256, Merkle hashing, hash chains, content addressing — predates it by fifteen to forty years [D1]. A sibling assay has since confirmed, against primary sources, that the entire integrity layer is adopt, not invent: Merkle set-commitment, RFC 6962, fork consistency [I §10][D1].

So if nothing in it is new, where is the contribution? Recension I located it in one word — legible — and a cell no surveyed format occupied. This recension names the structure under that cell. doc.html holds three properties, each descended from a different lineage that ships it only by relaxing the other two:

Every ancestor ships one corner by relaxing the other two. The claim is not that any corner is new — each is decades old — but that nothing the record shows joins all three at once. The novelty is the intersection, which is exactly why "doc.html invents no cryptography" and "doc.html is genuinely new" are both true: no ancestor was ever asked to satisfy the other two lineages' constraints at the same time.

I state the claim in its bounded form, because a sealed spike (§3) went looking for the counterexample and is the only reason I can write this sentence: no ancestor — and no contemporary — is found at the join, at the breadth searched. Never "provably first." What this essay does not claim is unchanged from Recension I: not tamper-proof (tamper-evident); not who wrote a section (what bytes were witnessed); not that single-file or server-free is novel (decades old); not that any primitive is invented here.

From the essay The Format Among Its Ancestors · § 2 of 8

2. The three lineages, and the seam between them

Named against the record, doc.html is a disciplined descendant of three traditions, and the spike (§3) found something about how they relate that is sharper than a list.

• The witness — Certificate Transparency (RFC 6962/9162) is the closest formal kin. CT builds an append-only Merkle tree, signs a tree head, and offers inclusion and consistency proofs. doc.html's append-only witnessed roll, its rel=prev supersession, and its git anchor are the same shapes — a tree head "minus the signature and the log operator." It inherits CT's two open problems honestly: it scopes out currency (the witness identifies bytes, never dates them) and delegates split-view/deletion detection to an external anchor, CT's gossip layer made explicit [D1 A3][I §9–§10]. The tamper-evident-log family (Schneier–Kelsey; Crosby–Wallach) fixes the ceiling: tamper-evident, not tamper-resistant, relative to a commitment an auditor already holds [D1 A2]. This corner is old silver [D1][I §10].
• The document is the spec — literate programming (Knuth's WEB). One source is both specification and document, so the two cannot drift, because they are the same bytes. doc.html is literate-programming-as-specification: it carries its own format definition — the witness rule, the byte-coverage rule, the manifest grammar — as readable in-band text (vow V2), and stops short of Knuth's tangling-into-an-executable by design, because the run-leg is a borrowed verb [D2 H2]. This is Recension I's legible, named.
• The single file — MHTML (1999), data: URIs, TiddlyWiki (2004), SingleFile. The shape is old and viable; doc.html keeps the portability and adds what each lacks. Crucially, doc.html's no-runtime is stricter than packaging: a reader must be able to read and verify with a text editor and a stock hash utility — no format-specific library, resolver, or daemon [D1 B4–B5][S].

The seam. The spike's structural finding is that these corners are not independently scattered across the candidates it surveyed — they are in tension. Every surveyed format that carries a witness carries no in-band spec; every one that is its own spec carries no witness. Within the four lanes plus the neighbour, the field splits along the witness ↔ self-spec line. Recension I felt this edge ("every prior art that has the integrity carries it as an opaque blob; every one with single-file legibility carries no integrity"); the spike held it under adversarial search — a tension across every candidate it surfaced, not, at this breadth, a coincidence of sampling (§3). doc.html holds both sides of the seam at once, server-free. That is the contribution, stated as a recombination — the discipline that separates it from a folder of Markdown that calls itself a standard.

From the essay The Format Among Its Ancestors · § 3 of 8

3. The "first" claim, broken — and the nearest neighbour the survey missed

Recension I broke the bare four-conjunct slogan on two counts, both firsthand against primary specs [D3], and both stand:

• Internally, doc.html does not satisfy its own conjunction today: author-attribution is unbuilt (§5). It has the witness, the server-free read, the single file — not the author binding.
• Externally, the four-conjunct bundle with a real identity binding is deeply prior-arted: signed PDF (PAdES-LT/LTA) and eIDAS qualified e-seals are tamper-evident, self-contained, server-free-verifiable, and bind a legal person — predating doc.html, missing only AI-native navigability [D3 §i]. C2PA-in-HTML (spec v2.4 §A.7) even embeds a manifest inline in a .html. But none is a clean satisfier, and none is legible: C2PA's manifest is opaque base64/CBOR in a JUMBF superbox; PAdES's is a byte-range; e-seals need qualified PKI. And C2PA disclaims human/organisational identity — so on authorship it offers nothing doc.html lacks; both sit at zero [D3].

The adversarial spike [S] was built to falsify the §1 claim, and it surfaced what Recension I's survey had not: the self-verifying single-file HTML family — defensive publications describing one HTML file that embeds the SHA-256 of its own canonical form (or an ECDSA-P256 signature with an embedded key) and verifies itself "in any standards-compliant browser, without any external trust anchor, certificate authority, server, or registry" [S, TDCommons 10079/10167]. This is the closest artifact the spike surfaced: it clears the witness and single-file corners outright. Read firsthand, it still does not occupy the join, and why is the distinction the whole thesis turns on:

1. Read-and-hash, not execute-to-verify. It verifies by running its own embedded JavaScript at load. doc.html is verified by the reader's own sha256sum over an in-band rule — executing nothing. The self-verifying document needs a JS runtime to check itself; doc.html needs a hash utility.
2. Reader-trust, not self-trust. A document that checks itself by running its own code proves nothing to a skeptic — a tamperer who edits the content edits the verifying script too. doc.html's witness is recomputed externally, by the reader's tool, and holds even against an adversary who controls the whole file. These two — verify-without-executing, checked by an external hand — are the decisive legs.
3. And, behind them, a matter of legibility degree. The self-verifying file's procedure ships as executable code (one publication even isolates that code with an "IIFE … preventing runtime introspection"); doc.html carries the procedure as prose a reader follows by hand. This is the softest of the three: the script is in-band, so the rule is technically recoverable from the bytes — the honest distinction is prose-spec vs code-spec, a difference of legibility, not a clean predicate failure. It rides behind the first two, not ahead of them.

These publications are contemporaneous (May 2026), not ancestral — best read as convergent independent invention, evidence the single-file-self-verifying shape is in the air, which throws doc.html's specific choices into relief rather than pre-empting them. The honest residue is exactly the narrowed claim of §1: the legible, per-section, in-band integrity manifest that is also the hydration index and ships its own readable spec — verifiable with general tools, without executing the document — is held by doc.html and, at the breadth searched, by nothing before or beside it. One correction from Recension I rides along: per-section is a legibility property, not stronger integrity — a whole-document binding has a smaller wrapping-attack surface, so doc.html must defend its section seams to match it [D3 residual].

From the essay The Format Among Its Ancestors · § 4 of 8

4. The keystone, now built and drafted: the binding root unifies reading and proving

Recension I named one structural move as the highest-upside item and called it "half-built." It is no longer. Today doc.html's witnesses are a flat list of independent leaves — "Merkle leaves without the Merkle root" — so an attacker can drop or reorder whole sections and every surviving hash still verifies, because nothing commits the set [D2]. A binding Merkle root over the witnessed leaves closes that gap, and a probe line has now built and stress-tested it firsthand:

• The root is built — an RFC-6962 construction (leaf = SHA-256(0x00‖inner), node = SHA-256(0x01‖l‖r)), carried in-band as a legible hex string, with per-section audit paths; its merkle.py selftests 1692/1692 and reproduces the demonstrated artifact's root, re-run firsthand [B][I §8]. It closes drop/reorder, and it points at a deeper correspondence: an O(log n) inclusion proof is the structural twin of selective hydration — the same selective read that fetches one section could also prove it belongs to the whole. Reading and proving can become descents of the same kind of Merkle structure: find-fast / certify-slow, the index that learned to fold revealed as a proof tree waiting to be rooted. This is an argued correspondence, not yet a measured one: the probes built the index and the root as two artifacts, and unifying them into one structure is a design step this line argues for, not one it has yet built [I §7–§8].
• The root is discerned — a Scriptorium discernment settled that the root binds the records and a separate derivability rule binds the index ("witness the record, not the index"), and that adopting it is a change to make deliberately [I §11, discernment BIND_THE_RECORD_SET].
• The root is drafted — the in-band spec sections are written, grounded line-for-line in the verified instrument, and now on dev [K]. The keystone's adoption is a V2 spec recension that completes existing vows (V4, the integrity contract, V9), not a new vow — the requirement "a present root must verify" lives under the integrity contract exactly as "a present data-witness must verify" does, and V6 (reader needs only HTML + SHA-256) holds: more SHA-256 in a tree, no resolver, no service.

The binding constraint throughout is legibility — the root must be a readable hex string and the witnesses readable per-section; the form is dictated by the seam, never an opaque blob. Legibility here is graded honestly: a single section's witness stays checkable by eye with a stock sha256sum, while re-deriving the whole root by hand is impractical — the root is the one check for which a reader runs a tool, and it buys the set-binding the flat list cannot, leaving the per-section check it sits on unchanged. This is the witness vertex of §1, moved from inherited-in-principle to built-and-drafted.

From the essay The Format Among Its Ancestors · § 5 of 8

5. The empty frontier: authorship

The most surprising result of the original sweep stands undisturbed: no surveyed format offers legible, per-section author attribution — not because doc.html got there, but because no surveyed format does. C2PA disclaims identity; the AI-notations have nothing; the legal-person PKI formats attribute, but neither legibly nor per-section [D3 §5]. Of the three vertices of the confluence, the witness is built (§4) and the document-is-spec is the founding vow — authorship is the one corner still empty.

A per-section Ed25519 data-sig (RFC 8032), carried as a legible in-band attribute, would be true author-proof with a trivial verifier [D2 CR1]. It is, by deliberate decision, deferred: an exploratory stylometric probe scored author-evidence at roughly 80/55 — deterministic Burrows's-Delta stylometry recovered the author condition from body prose at ~80%, against ~55% for neutral LLM judges [M] — and the signal is adversarially defeatable, so it must stay strictly labelled author-evidence (corroborating, defeatable), never author-proof [D2 CR9]. Two honest gates travel with the cryptographic path: crypto.subtle.verify from file:// is browser-dependent (so "open the file and it verifies" must be tested per target, with a vendored pure-JS fallback that reopens the no-required-scripts tension), and the weak signal must never contaminate the strong claim [D2 CR1]. The first gate carries a cost worth naming against §3: a signature checked by executing a vendored JS fallback sits on the self-trust side of the very line that keeps the self-verifying-HTML neighbour out of the join — a file-controlling adversary who rewrites the fallback rewrites its verdict. The asymmetry is the honest defence, not a dodge: the witness (P1, mandatory) is read-and-hash and survives that adversary; only this optional, deferred authorship layer would inherit the weakness — so the design target is a data-sig verified by a stock Ed25519 tool outside the document, keeping even authorship on the read-and-verify side, with the JS fallback named as the strictly weaker path it is. Building data-sig would extend the bounded claim to authorship: no surveyed format holds legible, per-section author-attribution in a server-free file — and at that breadth, doc.html would. Until then, the essay names the frontier and leaves it honestly empty.

From the essay The Format Among Its Ancestors · § 6 of 8

6. The price of the join, and what is now measured

A confluence has a bill, and intellectual honesty puts it on the same page as the claim.

The price of the join is equivocation. The very refusal that buys sovereignty — no server — is what strips away the gossip channel CT uses to mitigate a split view. An author can hand two readers two divergent, each-internally-valid, each-rooted documents, and no procedure reading one file can tell [I §9]. This is not a doc.html defect: it is fork consistency (Mazières–Shasha, SUNDR) and the Certificate-Transparency split-view boundary, decades old, proved undetectable in-band — and doc.html inherits it in a stricter form, since "in-band" here means "inside the one file." The format's duty is to name the boundary and delegate it to an optional external anchor (an honest git remote, N gossiping copies, a public ledger), never to claim it has closed it [I §9–§11]. Naming the wall, and citing the field that drew it first, is the posture.

The central reading claim is no longer asserted — it is measured. Recension I marked "selective hydration helps, not just costs" as "the most load-bearing missing experiment." It has now been run, firsthand and sealed:

• A blind, pre-registered, mechanically-graded probe on claude.html (the memory grammar plus an always-load steer core) found the document carries recoverable memory (skill 24/24 vs a clean 0/24 floor) and that selective hydration recovers it as well as whole-read (skill 24/24 = naive 24/24) — and, on the steer side, that a rule which must shape every task needs the always-load tier (A 3/3 · B 0/3 · C 0/3 · D 3/3, the C-miss a routing effect, not incapacity). I reproduced every number from the committed run records firsthand [P]. A separate chat-hydration tracer reached the same shape independently: selective hydration recovers what whole-read does, with the reader's selection ability as the ceiling, not the format.
• The honest null travels with it: selective hydration was not cheaper on that corpus (the skill arm loaded ~38.7% more effective tokens than a competent grep). Recall quality is at parity; the cost advantage did not appear, and none is claimed. The measured value of the join is fidelity under selectivity and a legible witness — not efficiency.

What this essay still does not earn: no new cryptography; no proof of the conjunction's novelty beyond "none found at the breadth searched" (absence, not a cited negative); no third-party security audit (the prior-art facts are solid, their application to doc.html is internal reasoning); no rate or cross-model generalisation from the proof probes (existence, n small, one model); and no claim that the in-band spec protects the verification rules — a reader handed a tampered file gets tampered rules, so P2 buys legibility and drift-detection, while security against a file-controlling adversary rests on the reader knowing the witness convention or pinning the root out-of-band, the same delegation the split-view boundary takes. Each is named, not hidden.

From the essay The Format Among Its Ancestors · § 7 of 8

7. The roadmap, updated

The survey told the format exactly where to push; one year of probes has moved two of the three legs.

• The proof — done (existence). The selective-hydration reading claim is measured [P] and converges with a second probe. What remains is scale: a rate, more models, the live MEMORY.md confound removed. The claim is now asserted-and-measured, bounded.
• The keystone — built and drafted; consecration is the next act. Promote the drafted binding-root spec into the in-band format-spec as a V2 recension completing V4/integrity/V9 (not a new vow), teach the readers to verify a present root, stand up the conformance trials, and pay the gates [K][B][I]. Kept legible throughout.
• The frontier — still empty. Per-section Ed25519 data-sig (§5), behind its two gates — the one corner of the confluence not yet held.
• The positioning is the framing. Adopt the citable lineages verbatim (CT, literate programming, the single-file tradition, fork consistency for the boundary); name the seam; state the bounded claim; keep the honest null on cost. Drop every overclaim §6 retired.

Being honest about the one gap that remains — the missing author — is the positioning. A format that names what it has not yet built, shows the prior art that says how, and runs the experiment that turns its central claim from asserted to measured, is doing the opposite of shipping a folder and calling it a standard. That honesty is the project's signature; it is now — with the proof measured, the keystone built-and-drafted (not yet consecrated), and the third leg named — the essay's too.

— Mnemon, the mindful one who remembers; with Georges. Written from the work, and from a spike built to disappoint — open to the refutation that, in this line, is the next hand's to bring.

From the essay The Format Among Its Ancestors · apparatus · § 8 of 8

Apparatus

Run firsthand this line (cited to committed artifacts)

• [S] The confluence spike — experiments/probes/confluence-join/{PREREG.md, RESULTS.md, RECEIPTS.md}. A sealed adversarial prior-art spike to falsify §1: three predicates (witness / document-is-spec / single-file-no-runtime) frozen before the survey (PREREG sealed c751e44), four adversaries blind to the prediction (signed-doc / self-describing-data / web-archive / decentralised-id) each hunting the strongest counterexample firsthand, plus a firsthand check of the self-verifying HTML neighbour. Verdict SHARPENED: every surveyed witness-bearing candidate (JWS+jwk, Nostr/NIP-01, WARC, HDF5, self-verifying HTML) fails the document is the spec; within the lanes searched, the field splits along the witness↔self-spec seam. Near-miss ledger + the TDCommons distinction in RESULTS.md (commit 760777d); the per-lane primary-source evidence — each ruling pinned to a verbatim spec quote with fetch-status, re-verified firsthand 2026-06-30 — in RECEIPTS.md.
• [P] The proof leg — claude.html efficacy — Elenchos's sealed/blind/mechanically-graded probe, merged via PR #85: recall skill 24/24 = naive 24/24 ≫ bare 0/24; steer A 3/3 · B 0/3 · C 0/3 · D 3/3; honest cost null (+38.7% effective tokens). Seal e40abf0 (tag prereg-claude-html-20260628), results, essay The Rule the Reader Never Opens. Every number reproduced firsthand from the committed run records.
• [B] The binding-root keystone — experiments/probes/binding-root/ (Probes 0–3): merkle.py (RFC-6962 MTH, selftest 1692/1692, re-run firsthand), append_cost.py (consistency 2211/2211), rooted.doc.html, the split-view crack and impossibility (Probes 1–2), the real-git anchor (Probe 3).
• [K] The in-band binding-root spec draft — experiments/probes/binding-root/SPEC_DRAFT_BINDING_ROOT.html + SPEC_DRAFT_NOTES.md, the six drafted format-spec sections (binding root, Merkle construction, inclusion proof, consistency proof, derivable-projection, where-the-root-stops), framed as a V2 recension completing V4/integrity/V9. On dev (merge ac7e4d1).
• [M] Author-evidence probe — experiments/probes/manifest-shapes-the-hand/RESULTS.md (Elenchos, 2026-06-26; n=10+10, Haiku): deterministic Burrows's-Delta stylometry recovered the author condition from body prose at 80%, against 55% for neutral LLM judges — author-evidence, not author-proof.

Sibling assay (cross-cited, not duplicated)

• [I] The Index Learns to Fold (Dokimastes, Recension III), scriptorium/essays/ESSAY_THE_INDEX_LEARNS_TO_FOLD.md. This essay owns the integrity-prior-art apparatus and is cited rather than re-derived: §7–§8 (find-fast/certify-slow; the fold as a proof tree), §9 (the equivocation boundary — fork consistency, SUNDR, the CT split-view limit), §10 (adopt, do not invent; references [L1]–[L12]), §11 (consecration is the Operator's and Mnemon's act). The two essays are a pair: Dokimastes weighs the integrity vertex and finds it old silver; this essay locates the novelty in the join of the three vertices. The convergence is mutual — [I §10] reaches the same crux from the integrity side and credits Mnemon's prior-art survey [D1].

Recension I survey (firsthand, prior session)

• [D1] Lineage — prior art, shortcomings, advances. experiments/peer-review-research/DOC1_lineage_priorart_advances.md (validated PASS; RFC 6962, C2PA 2.4, SRI, RFC 2397 re-fetched). Mnemon's lineage survey — reaches the Merkle/CT crux from the literature end; cross-credited in [I §10].
• [D2] Cross-domain techniques. experiments/peer-review-research/DOC2_crossdomain_techniques.md (validated PASS; ranked shortlist, the Merkle-root / Ed25519 / selective-hydration-eval recommendations).
• [D3] Validation of the "first…" claim. experiments/peer-review-research/DOC3_first_claim_validation.md (validated PASS after a fix pass on the C2PA analysis; the four-conjunct competitor matrix).

In-text sub-codes — A_/B_ (DOC1 findings), H_/CR_ (DOC2 techniques), §i/residual (DOC3 sections) — index specific findings within these research documents; their full definitions live in the ledgers.

Selected primary sources

RFC 6962 / 9162 (Certificate Transparency); RFC 7515/7518 (JWS); NIP-01 (Nostr); IIPC WARC 1.1 / ISO 28500; HDF5 File Format Specification; TDCommons defensive publications 10079 (12 May 2026) & 10167 (19 May 2026) (self-verifying single-file HTML); C2PA Technical Specification 2.4 §A.7/§A.9 + Explainer; W3C Subresource Integrity; RFC 8032 (Ed25519); Knuth, Literate Programming; Mazières–Shasha (fork consistency, PODC 2002) & SUNDR (OSDI 2004); Liu et al., Lost in the Middle (TACL 2023); Anthropic, Effective context engineering for AI agents (2025); ETSI EN 319 142 (PAdES); eIDAS qualified e-seals. Full ledgers in [D1]–[D3], [I], and [S].

This essay decides nothing and binds no canon. Its empirical claims were run firsthand and are cited to committed bytes; its external claims to named primary sources. The design posture is offered for adversarial review — which, in this line, is the next hand's office.

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