A persistent object-representation layer for AI. The AI no longer edits PDFs and screenshots — the shadows of shadows. It edits the origin, and the system proves the edit was correct.
Explore locally verified cases: a real multi-host diagnostic task, deterministic proposal acceptance, and process handoff at a pre-write checkpoint. Each case states what has and has not been demonstrated.
AI Computer & cases （中文） →A real building becomes a CAD model, then a floor plan, then a PDF, then a screenshot, then OCR text. Every step travels better; every step loses something. AI today is fed this seven-layer-compressed projection of a projection — and we call the last shadow the source file.
It generates Word, Excel, PPT and CAD — yet has no stable open-see-locate-edit-verify loop. What it reads is screenshots of exports of summaries.
Chat history is treated as project state. The longer the chat, the closer to a full window — and compressed summaries drift. State and conversation are conflated.
AI emits whole deliverables — 40-page decks, full novels, chart configs. Content, format and references fail together; the only remedy is to regenerate everything.
2Origin keeps seven kinds of things together. The AI submits only a semantic transaction — what changed, what it depends on, what it asserts. A deterministic compiler validates it, applies it, and generates the artifact. Violations are bounced back with the evidence attached; nothing is written unless it passes.
“Chat is only a temporary window onto the origin. Close the chat, switch the model, clear the context — the world is still there.” — docs/01-愿景与定位.md
real entities with stable IDs — the only thing you can hold onto.
who owns, references, depends on, generates whom.
past, present, change, and the reason for each.
who created, who modified, what was inferred.
what may and may never happen — enforced at commit.
native domain data: geometry, formulas, timelines.
what this representation cannot guarantee. The boundary travels inside the package, not in a README.
The claim has been narrowed by experiment to exactly one dimension: state tracking. Task: across a 95k-character narrative, who holds the key right now, who learned the secret, which planted threads are still open.
One line of prompt (A0→A2: "explicitly maintain state") accounts for +41.3pt of the state-accuracy gain (p = 3.3×10⁻⁵, significant). Adding the full machinery on top — validator, evidence chain, compiler (A2→A3) — adds only +2.5pt, not significant (p = 0.47 / 0.72). The only surviving categorical claim is structural: prompts cannot produce evidence chains. A3 gives traceable evidence for 15.0% / 27.5% of tested fields; every prompt-only baseline stays at 0%. Full writeup: README §5 ablation.
| W3 state accuracy | Benxiang | Bare model | Vector RAG |
|---|---|---|---|
| Score | 95.8% ± 5.9 | 52.1% ± 19.7 | 58.3% ± 11.8 |
| Permutation test vs. Benxiang | — | p = 0.0024 | p = 0.0024 |
The unfavourable part first. Field by field, Benxiang scores 4/6 on
black-key.holder — worse than either control arm's 5/6 — and that is exactly the one field recoverable
from the prose the arm just wrote. Its whole advantage lives in the two fields where both controls collapse:
| Field | Bare | RAG | Benxiang |
|---|---|---|---|
black-key.holder — last handover written in its own prose | 5/6 | 5/6 | 4/6 |
bai-yao.secret_betrayal — a secret accumulated over dozens of chapters | 1/6 | 0/6 | 6/6 |
hook:shen-yan-suspicion.status — never stated in any passage at all | 0/6 | 0/6 | 6/6 |
“State you can read back from your own recent prose: retrieval and a bare model both suffice, and Benxiang has no edge. State that must be accumulated across the whole book and is never stated in any one passage: only the state machine answers it.”
| qwen-plus | deepseek-v4-flash | Permutation test | |
|---|---|---|---|
| W3 state accuracy | 98.9% | 98.9% | diff 0.0000 · p = 1.0000 |
| W1 EPC (prose, lower is better) | 0.20 | 0.55 | diff −0.35 · p = 0.0392 |
State-layer correctness is independent of the base model; prose-layer quality is not. That is the property a protocol ought to have: it does not depend on how strong the machine underneath is. And it is 11 against 11 — not an n=1.
One dialect working could be a coincidence. Two is a protocol. The same core, essentially unmodified, runs across four domains — domain knowledge lives in data (constraint tables, statute books), not in code. All four dialects together added 4 lines to the core.
Chat history committed as world state. A new session resumes in seconds. MCP server, zero deps.
Characters, secrets, foreshadowing and forbidden zones held across a million-character target without amnesia.
Citation whitelists and sentence recalculation. A 20%-over-limit mitigation is refused on the spot, with basis.
Formula dependencies as relations. “This number is decided by these 10 manual inputs” — Excel itself can’t answer that.
| Vectors | |
|---|---|
| Language-neutral test vectors (data, not code) | 68 / 68 |
| JavaScript reference implementation | 68 / 68 |
| Python second implementation (~250 lines, zero deps) | 60 / 60 core |
| Reference-implementation dependencies | 0 |
Any implementation that writes a small adapter can prove conformance on the spot — the standard is the 68 vectors, not the name.
An implementation passing its own tests proves nothing about a protocol. The mutation check deliberately breaks each protocol promise and watches who catches it. Promises caught only by the self-test are coverage gaps — the protocol guarantees only what the vectors pin down.
The longest novel ever written by a human is Venmurasu — “the white drum” in Tamil — 26 volumes, ~22,400 pages, one person, six and a half years. AI now challenges that record, with one difficulty a human author never faced: the world state of every chapter must be verifiable.
Loading stats.json…
first public milestone
a million without amnesia
beating the human record, verifiably
drift-rate — the model’s rate of misremembering prior values — last measured at 8.3% (2026-08-06; run id not retained), not a live metric. Every chapter: prose + world state + git evidence chain. Read the rules →
If this holds, a “file” should no longer be a .docx or a .pdf,
but an origin package that can be opened, seen, located, modified and re-verified.
Humans read projections. AI edits the origin. The system proves the edit was correct.