The Echo Angel Platform

The Problem

Three gaps keep capable AI out of high-stakes professional work.

Confidently wrong output. A language model produces fluent text that is not always faithful to any source — and the failure runs in both directions. Fabrication: a claim or citation that sounds right and isn't. Omission: an answer that is accurate in everything it says and wrong by what it silently leaves out — the standard element never addressed, the consideration never raised. Most tooling defends against the first and says nothing about the second.

The exfiltration problem. The dominant way to apply AI to your own material is to send that material to a hosted model. For a confidentiality-bound professional, the transmission is itself the exposure — independent of whether the answer is any good.

The accountability gap. Even when the work goes well, what would you show an auditor? A chat transcript is not a record of what was checked, what was escalated to a human, or what the governance actually did. Professional work needs a trail, and AI-assisted work mostly doesn't leave one.

The Approach

Echo Angel does not try to out-verify the mainstream pattern. It is a different architecture.

1. No model inside either tool. Both tools are model-free: the application makes no model call of its own, and neither does the orchestrator. Adopting the platform adds no new model that sees your content.
2. Your assistant does the reasoning. The professional's own AI assistant — the client they already use — drives both tools. The orchestrator is the outer loop, wrapping any kind of working session; the application is the inner loop, for research and drafting work.
3. Enforcement at the point of export. The application's trust properties are not advisory suggestions on a draft. At export, the app re-checks the exact document being finalized — every claim grounded, every standard element addressed — and emits it, or refuses with reasons. A check that passed on some earlier draft cannot wave a flawed final version through.
4. Process accountability. Every session leaves an audit trail: what was intended, what happened, what was escalated to a human, and what governance fired — durably recorded as the work happens, not reconstructed afterward.
5. Local by default. Local mode (the default — fully offline, fail-closed) keeps sources, drafts, citations, and records on your machine; a step that would require the network is refused. The web mode (a loud opt-in) is a separate, clearly marked choice that announces itself on every use.
6. A content-free record when even the record is sensitive. For the most sensitive work, the orchestrator can keep a content-free work record: the structure and governance outcomes are kept; the content is never stored.

The Build — Echo Angel Research (the app: artifact governance)

The application governs the document that ships. It supplies your assistant a disciplined loop of model-free tools: load and search your sources, return a passage's exact captured text for verification, record a citation, check a draft's grounding, check its coverage, and finalize. Within the verify–cite–finalize loop, the checks are deterministic — the same document and the same sources produce the same verdict.

Grounded by construction. Every claim in a finished document is tied to a specific captured source passage. A citation cannot be recorded against a source that was not captured — the check is preventive, not an after-the-fact audit.

Omission defense. A coverage check requires every standard element to be addressed or visibly dispositioned. The element sets are a framework, not a one-off checklist — the same omission defense extends across work types.

The refuse-or-emit export gate. Finalizing is the enforcement point: the app re-checks the exact document being finalized and either renders it or refuses and says what is missing.

What it reads. Sources can be local documents, including PDFs and scanned or image files read through OCR. OCR is handled honestly: the application reports what it cannot read before you rely on it, and text recovered by OCR carries a spot-check flag that survives into the finished document — the caveat travels with the claim instead of disappearing into a log.

What it reaches. In the web mode, the application can fetch public pages, or sources you are authorized to access, which you declare and reach through your own sign-in — authorization is the user's recorded decision, not the app's claim. The application tells you, on every use, that the web mode reaches the web. A quality check rejects and reports pages it identifies as blocks, errors, or empty shells before they can enter your source library.

How it installs. The application installs as a single local executable; optional capabilities use system tools you provision; nothing phones home in v1; a planned licensing connector — which never receives your content — is on the roadmap.

The Build — Echo Angel Orchestrator (process governance)

The orchestrator wraps the working session itself — any kind of session, not only research. Its proof is the audit trail it produces.

Sessions you can name, leave, and resume. Sessions are named, resumable, and can run concurrently. Every call rebuilds its state from the durable record, so an interrupted session resumes where it stopped — restarts lose nothing.

A record that starts honest and stays complete. Every session's audit trail begins with a governance event — governance is engaged from the first call, not bolted on later. Events are written to the durable record before the governance checks run on them, and every check that fires is recorded with its outcome and reason: the log cannot quietly skip the uncomfortable parts.

Structure for the work. Work proceeds in bounded phases with declared intent, annotated with timestamped notes. When a decision needs a human, it is surfaced explicitly as a decision point — with a built-in governance check on how the accompanying recommendation is presented — rather than buried in prose.

Audit what was retained. The session record can be read back on demand, and the read-back respects the session's privacy settings — what you audit is what was actually kept.

A content-free option. For the most sensitive work, a session can keep a content-free work record: the structure and governance outcomes are kept; the content is never stored. The record proves the process was governed without retaining a word of what it governed.

Isolation, tested the hard way. Sessions are isolated per caller, and per tenant namespace in hosted form. A cross-caller isolation defect was found during real concurrent use, fixed, and verified — the kind of finding a platform only makes when it is genuinely being used.

Memory with discipline. Durable lessons persist in a cross-session observation catalog, and entries leave it only through a human-ratified retirement review — institutional memory that a single session can neither invent nor quietly erase.

In Production

The platform is not a concept piece — it is used to govern its own development. The sessions that build, test, and document Echo Angel run under the orchestrator: real, sustained production use across more than twenty governed sessions, which is exactly how the caller-isolation defect above was found and fixed.

Public, fully cited research deliverables have been produced under the platform's governance discipline. The lead example is the AI Harm Landscape compilation, live at ai-output-harms.pages.dev: nine parts, 248 citations across 43 sources — 205 of them warehouse-pinned to a captured source (most to a verbatim passage, the rest to a source-level reference), and every citation carrying a published verification status in the compilation's public register, including the minority marked provisional.

Independent of its own use, the platform has been through bounded validation — bounded meaning each check demonstrates exactly what it tested, and is described that way.

All of the above: validated at product commit 221e1a0, 2026-06-10; re-verified at each release. This is honest, bounded evidence — not certification.

Honest Limits