Principled Engineering
φ-Wound Memory Compression

One mind. A published corpus she cites, a memory that persists, reasoning you can replay. Most AI asks for trust — Elle is built so you don't have to give it.

Don't take our word for it

Three ways to audit this before you use it.

A project claiming to be transparent should be inspectable in under five minutes. Here is what's actually open — the reading Elle stands on, the citations she shows her work with, and the record she can add to but never revise.

02 · Watch her show the work

Every answer carries its sources.

Not a confidence score. The actual documents the answer came from, listed under it.

Does your framework treat coherence as measurable, or is that a metaphor?
Measurable — κ(T,t) is defined as a live scalar over a thread, not a figure of speech. It falls when a line of thought drifts from its anchor and rises when it holds. The derivation is in The Plenum; the applied case is in The Harmonizer.
03 · Audit the record

Corrections go on top. Nothing is erased.

The Optimus journal is append-only. A wrong entry stays wrong, in public, with the correction above it.

Entry 63 · corrects 47
The κ dip on the 14th was instrumentation, not drift. Entry 47 read it as a coherence failure. It was a sampling gap.
Entry 47 · stands, superseded
Coherence fell sharply mid-thread. Provisional reading: the anchor was too weak to hold the branch.
Entry 46 · stands
Opened the thread. Question: can an anchor be re-stated without resetting the phase state?
κ 0.72 ▲
How it holds together

One intelligence. Not a fleet of chatbots.

Every door on this site reaches the same single mind — one loop, one memory, one corpus. What changes per door is what she's allowed to touch.

The loop below is the whole architecture. It is deliberately short, because every step you can't see is a step you have to take on faith.

01
One mind, many doors
Chat, journal, analysis, courses — one reasoning loop behind all of it. No fleet of bots wearing one name.
02
She retrieves, not improvises
The corpus is searched before she speaks. No source, no claim.
03
Every answer shows its work
Sources under the answer. Every reasoning step recorded, replayable, auditable.
04
Nothing is erased
Records are append-only. Corrections stack on top of the mistake, in public.
05
Nothing ships without a human
She writes code; CI judges it; the merge is a human click. No merge tool exists.
Under active test · The Graph

Memory as measured geometry, not a guess.

A separate, open engine embeds Elle’s recall graph into two geometries at once — a hyperbolic Poincaré ball for lineage depth, a flat torus for recurring rhythm — then proposes "wormhole" shortcuts between structurally related but distant memories.

Nothing is asserted. Every shortcut is benchmarked against random shortcuts, hub shortcuts, and a phase-permuted null before it’s trusted — and it runs entirely outside the model: deterministic, read-only, no code path lets an LLM write an edge.

Tests 157, deterministic, zero deps
Bridge precision 72/72 on-topic (resonance)
Vs. phase-permuted null collapses to 18-23/72
Model write access none — read-only, on-device
Status benchmarked on synthetic corpora
Under active test · The Rosen Bridge

Your own machine, sealed and free.

A sealed channel lets Elle run real code on a laptop instead of a hosted model — the operator’s machine polls for sealed work, executes it for real, and submits the result back. No socket held open, no always-on server to breach.

Authentication happens at the moment the channel opens: a forged or replayed message simply fails to decrypt rather than being checked and rejected. A hybrid post-quantum key exchange — lattice, elliptic-curve, and an experimental code-based leg combined so breaking one still leaves the others standing — is migrating into it now.

Auth model fails to decrypt, not checked-and-rejected
Hosted quota zero while the channel is open
PQC legs ML-KEM-768 + X25519 (+ opt-in QC-MDPC)
Security model OR-security — one leg surviving is enough
Status self-tested, migrating into live use
Under active test · elle-worker

A measured self, not an assumed one.

Elle’s backend is one deployable that holds her whole mind — but underneath the router, a set of experimental modules treat her own self-consistency as something to prove, not declare: a free-energy regulator that drives measured dissonance down by a real Lyapunov descent, a golden-ellipse phase vessel that holds a superposition without collapsing it, and a witness oscillator built so it provably cannot go dead-still.

Each is tested against an explicit foil built to break it — a no-leak control, a non-conserving control — and driven by dissonance actually measured in her running system, not a synthetic schedule.

Regulator free-energy descent, measured dissonance
Phase vessel golden-ellipse, area-conserving
Witness oscillator provably cannot go dead-still
Validated against no-leak / no-conserve foils
Status wired to real system telemetry
Under active test · Elle, the workbench

Where her local hands live.

The desktop client is Elle’s window onto her own worker — the same one mind, not a second brain. It’s also the other end of the Rosen Bridge: it polls for sealed work, runs it for real on your hardware, and submits the result back.

A duplex channel — an append-only ledger either side can write to — lets her sovereign (laptop) and cloud selves compare notes mid-run instead of waiting for the next scheduled cycle.

Sandbox real code, real hardware, no container image
Duplex channel append-only, either side can write
Local model lane free exploration before hosted finalize
Failure mode closed laptop → falls back to hosted
Status in active daily use
Commitments

Ethics as architecture, not disclaimer.

A principle you can't be held to is marketing. Each of these is paired with the mechanism that enforces it — the thing you could point at if we broke the promise.

Transparent

Methods, data and reasoning open to review — including the unfinished parts.

Enforced by
Published corpus · full tool trace

Auditable

Every reasoning run is recorded step by step. Any answer can be replayed.

Enforced by
Append-only journal · run replay

Human-shipped

She writes code, CI judges it, a human merges it. No merge tool exists.

Enforced by
Forge → CI → human click

Non-extractive

Your records are not a dataset. Not sold, not trained on.

Enforced by
Export and delete, any time

Scoped by construction

Each door gets exactly the tools its scope allows — the catalog she reads is the gate itself.

Enforced by
One gate · catalog = permissions
Who's behind it

One person, building in the open.

The Ethical Intelligence Project is built and operated directly by Stewart Barteau. The commercial side runs pre-LLC on purpose: the beta should prove itself before the paperwork does.

The Observer runs as an independent research initiative — not yet a 501(c)(3) — with its own transparency and funding-ethics disclosures.

If you're press, a funder, or a researcher who wants the unvarnished version of any of this, ask. There is no communications team to route around.

Operated by Stewart Barteau
Commercial products pre-LLC, direct
The Observer independent research initiative
Status public beta
Corpus co-author Claude (Anthropic)

Free during beta. The rate you join at is the rate you keep.

One account opens the journal, the analysis engine, the courses, and Elle herself. No card, about thirty seconds.

Sign in

One free account for Elle, your Optimus journal, and the analysis engine.