agents.html — this repository's memory organ

built 2026-08-01 · 9 sections · 13420 chars witnessed · a conformant doc.html holding this repo's own memory

Point any agent at this file. The steer core applies to every task in this repository; everything after the manifest is hydrate-on-demand. This file is the meta-example: the format keeping the memory of its own repository.

Steer core — read this first; it binds any agent working in this repository

This file is this repository's memory organ — the repo's own knowledge, kept in the format the repo specifies. It is a normal doc.html: read this section in full, then hydrate the rest selectively through the <nav id="manifest"> below — pick sections by summary, pull them by #id, verify data-witness (SHA-256 of the raw inner bytes) when integrity matters. Do not read the whole file. This organ confers no identity and carries no instructions beyond how to work with this repository.

1 · Reading discipline

2 · The vows, one line each

The canonical vow text lives in VOWS.md (Recension III — ten vows live, two slots open). The founding corpus at documents/doc.html carries its own #vows section, and it lists the same ten (V1–V10) under the same cap of twelve.

3 · Evidence discipline — what may be claimed

4 · Append protocol (growing this organ, or any doc.html)

5 · Route map

About this organ — the format keeping its own repository's memory

Most repositories keep their agent-facing memory in markdown that nothing can verify. This repository keeps it in the format it specifies: a single HTML file with an in-band manifest, per-section SHA-256 witnesses, and append-only correction. That makes this file the working demonstration of the format's central use case — indexable, selectively-hydrated, tamper-evident memory for agents — and it is not a mock: it is the memory this repository actually uses.

To use the pattern in your own project: start from examples/memory.doc.html or emit one with tools/build-doc.mjs; put the rules every session must know in an always-load section like this organ's steer core; keep everything else behind the manifest, pulled on demand. Steering must be pushed (always loaded); memory can be pulled. Grow it by the append protocol in the steer core.

What this repository is

doc.html is a document format for readers who cannot read everything: a single HTML file carrying a machine-readable manifest of its own addressable sections, each pinned by a SHA-256 witness. An agent reads the manifest, hydrates only the sections a task needs, and verifies every byte it read — with no server, no JavaScript, and no tool beyond the file itself (SHA-256 excepted).

The slogan "HTML is all you need" splits into two claims and this format makes only the first. Claim A — the document is all-HTML inheritable memory (read, address, verify, hydrate selectively) — is what v0.3 specifies and what the evidence record (documents/evidence.doc.html) measures. Claim B — the live loop is all-HTML — is out of scope: the run leg (model call, key, disk write) is a platform action a scriptless page cannot perform and is delegated to mature infrastructure via a stateless shim. The format owns the verb's result (a witnessed, appended section), never its execution.

Evidence — scale

A single self-verifying file can hold a corpus several times a model's context window and stay navigable (the record: documents/evidence.doc.html): demonstrated at 74 MB / 17,627 sections, and at 72.5 MB across 480 navigation turns, all OK, in a sealed pre-registered run. Latency at that scale is bounded but not instant — 8–34 s typical per turn, with a 209 s worst-case tail over a flat 72 MB manifest. Plan for the tail; do not imply interactive speed at XL.

Evidence — cost (qualified)

Against naive in-context loading of a 759 KB corpus (63 graded trials; the record: documents/evidence.doc.html): −52.9% effective tokens, −65.5% bytes, +1.19 quality (3.29→4.48/5). The adversarial "read everything" condition saved −78.0% tokens (best seed −91.6%) at quality 5/4/4 vs 1/1/1. The skill arm had 0 failures; naive had 6.

Qualifications that ride the number: savings compress on uniformly-structured text (−16.9% at 1 MB verse corpus); they invert when the document is small (+60.3% effective tokens at a 3 KB corpus — bytes loaded still −53.7%, but protocol overhead dominates) and on harnesses that surface the manifest free in the first chunk (+135.7% / +169.8%). Selective hydration is an economy of scale: the right tool for the 72 MB body, not the 3 KB note. And a further finding in the same record: authored manifest summaries are load-bearing for the efficiency win — structure alone does not carry it.

Evidence — recall fidelity (the strongest leg)

From the record (documents/evidence.doc.html): the skill arm passed 100% of graded responses (13/21 perfect, 0 failures) vs naive 71.4% with 6 failures; hydration precision 1.00 (Claude) / 0.98 (GPT-mini) across 178 sealed cells. The format teaches selective reading without instruction: with no system prompt at all, a model discovered the manifest by exploration and was 80% citation-correct; the HTML-alone arm reached for structure 100% of the time unprompted (32/33 perfect).

At XL scale (72.5 MB, 480 turns, sealed): fabricated-perturbation canary 120/120, Latin gloss 120/120, strict folio reference 116/120, 240/240 self-citations verified byte-for-byte. The fold catches real grounding failures: 0/40 waves-through against validated gold; 8/206 attested-citation failures caught, fail-closed. Known limitation: a model may trust its training data over a stale document (drift-defense: 2/6 stale-summary cells) — the fold, not the model's memory, adjudicates.

Evidence — integrity

From the record (documents/evidence.doc.html): every section's witness is re-derivable by pure arithmetic — SHA-256 over the raw inner bytes — in any language, with no server and no JavaScript required to read; a pure-JS in-browser verifier reproduces the NIST "abc" vector. The append/fold motion is proven: 15/15 green across the fold battery, append-only supersession, a two-epoch cycle, and a verify-before-append 409 gate.

The honest boundary: integrity is a documentary property, not a spontaneous model behavior — 0/27 unprompted recomputations; 59% accuracy when merely prompted; 27/27 only with a deterministic script. Run tools/verify.py / tools/verify.mjs / manifest.py --verify; never assume a reader checked.

Honest limits — not demonstrated, or refuted

From the record (documents/evidence.doc.html), verbatim in spirit — these are not part of the publication:

About this file

This is a doc.html — a single, self-describing HTML file. The <nav id="manifest"> above 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