Exports

Three formats.

Getting one

Markdown and PDF are on the run's Report tab, beside Copy, for whichever revision you are reading — before approval too.

The verification bundle is on the Artifact tab, which exists once a person has approved the report at the review gate: approving is what freezes it. The tab shows the verifier's own checks, names any agent that ran on replaced instructions, and offers all three downloads, with the bundle first — Markdown and PDF are the report alone, readable but not checkable. Verify it yourself expands to the exact command for checking the bundle, so the instruction is where the file is.

All three work on the self-hosted server. On the desktop app, Markdown and the bundle work the same way; PDF is the app's own Print → Save as PDF rather than a server render, and the PDF button says so (see below).

Research recorded as a session, on the earlier pipeline, keeps its own export routes under /research/{id}/ — see the API reference.

Markdown

The report text, unchanged, with a trailing block recording which model produced which part — the per-role routing the run actually ran with, not the routing currently configured.

Filename: research-<first 8 chars of run id>.md.

PDF

Rendered server-side: Markdown → HTML → PDF via WeasyPrint, with a self-contained print stylesheet and no external assets, so it renders identically offline.

Inline [n] markers become superscripts, followed by a numbered sources list and the same model-attribution block.

If the environment lacks WeasyPrint's native libraries the endpoint returns 501 with the reason rather than a broken file. The Docker image installs them; the desktop build deliberately omits WeasyPrint — its dependency chain on Windows is a packaging tar pit — and uses the WebView's own print-to-PDF instead.

Research bundle

The interesting one. A .bundle.json is a self-contained, auditable record of the run — a bill of materials for a research report:

• the report and its SHA-256;
• every claim with the citation indices it carries;
• every evidence snippet with a content hash and the source it came from;
• the sources table, and any contradictions surfaced;
• the models that ran, the cost, the token counts, the elapsed time;
• the prompts the run used — each one's full text, its SHA-256, and whether you had replaced it — and whether your custom instructions were applied (from 3.0.0);
• the full approval chain — every approve or rework, its feedback, the hash of the draft it applied to, and when;
• the agent trace, and a trace_available flag distinguishing "the host has no durable event log" from "nothing happened" — both the server and the desktop app write one, so it is true on each;
• a bundle_hash covering all of the above.

Verify one with no AI, no network, and no database:

python -m research_engine.verify_bundle path/to/research.bundle.json

Exit code 0 if valid, 1 if tampered. Seven checks run: schema validity, bundle integrity, report integrity, evidence integrity, citation resolution, claim-evidence linkage, and approval-chain integrity.

The load-bearing one is the last: at least one approved entry's draft_hash must match the report's own hash, which proves the approval was given for this report rather than an earlier draft or a different run.

A bundle whose agents ran on replaced instructions verifies like any other — its hashes are real hashes of what really ran — so the verifier prints a CUSTOMISED AGENTS banner above the verdict and lists the replaced prompts. The prompts' text is in the bundle in full, so anyone you share it with can read your instructions.

Bundles from 3.0.0 on are format version 2. A verifier from before 3.0.0 refuses them; the one from 3.0.0 or later verifies versions 1 and 2 alike. Research recorded before 3.0.0 still exports as version 1, with no prompts recorded.

Full specification: Research bundle format.

Demo runs are stamped

A run made with scripted models and fixture sources is marked in the database, and every export path stamps the artifact — on both hosts:

• .md and .pdf carry a prominent ⚠ DEMO — NOT REAL RESEARCH banner at the top;
• the bundle carries a hash-covered demo field instead, and the verifier prints the provenance above its verdict.

The banner text and the "stamped iff demo" rule live in one place, research_engine.bundle, and the server and the desktop sidecar both call it. That is not decoration: the desktop .md export shipped unstamped for a release while this page already promised otherwise, which is exactly the unmarked fixture report the flag exists to prevent. A parity test now fails the build if either host stops applying it, or if either starts stamping the bundle.

The bundle is stamped differently on purpose. Injecting prose into the report body would change the report hash and break the approval-chain check, making every demo bundle fail verification for a reason that has nothing to do with its integrity — teaching readers that FAIL is normal for demos would defeat the verifier far more thoroughly than a missing banner would.

Because the flag is persisted rather than inferred from the process's mode, a demo report cannot be laundered into a real-looking artifact by any route that bypasses the UI.

Copying and sharing

The report page also offers copy-to-clipboard and browser print. Shareable read-only report links are planned, not built — today an export is the way to hand a report to someone else, and the bundle is the way to hand it to someone who should not have to take your word for it.