atelier d'auteur · bruxelles · mmxxvi
Dt fig. I no. 01 - department title brussels · mmxxvi

Département des Harnais

author's workshop · ai agent harnesses
section des Dossiers · the model kept on a leash, the proof required
the lab cross-source forensic reports · r0 — rn
r0 dossier · 2026-06-14

The piece before the act

The Departement des Harnais audited against its own doctrine.

A file confronting the manual “Project Room” prescribed by Nate B. Jones to the automated chain of the Department. The Lab’s first report explains its work—not just what it concludes, but how it gets there.

read the file
r1 dossier · 2026-06-30

Langfuse, LangSmith, Phoenix : le spectre réel de l'observabilité LLM

« Open-source » : le mot qui couvre trois régimes

Trois outils d'observabilité LLM, un même mot dans toutes les brochures — et trois lois différentes dessous. Le rapport lit les licences plutôt que les arguments de vente — et le spectre qui en sort n'est pas celui qu'on nous vend.

read the file
<acquisition method="bitstream">

Cross-source: each regime, one requirement

Cross-source does not mean "a few links at the bottom of the page." It means holding together sources that are not verified the same way, to the same standard of a dated excerpt.

The requirement fits in one rule: no verdict without an excerpt. The harness supplies the material — it reads, cross-references, proposes; it does not decide. The verdict belongs to a deterministic test or to my own reading, not to a probability.

When the inquiry hits a wall — an inaccessible source, an unresolved contradiction, an uncovered area — the limit is written into the dossier, not smoothed over. A dossier without flaws, where the object had them, lies about how it was made. A marked zone of uncertainty is worth more than borrowed certainty.

<harness function="evidence-integrity" />

Audit, don't plead

I do not deliver an opinion. An opinion is believed; a dossier is verified. The difference is not one of tone but of nature — and the Lab produces only the latter.

A forensic dossier, here, is an inquiry run on a precise, defensible object, before a third party who does not trust it a priori — a regulator, legal counsel, an engineering committee inheriting a stack it did not write. The object is mapped, its sources are named and dated, and each section's verdict rests on an excerpt one can go back and read. Nothing is asserted that is not shown; the making is shown.

What is rare here is not the format — the cross-source forensic report is open and can be described in one sentence — but the material: a precise subject, sources to map, and the access to the material that alone makes a verdict possible. A traditional audit does not attach the forensic proof of its own making; that proof is exactly what sets this dossier apart.

<exhibit evidence="chain-of-custody" />

A report, and the trace that produced it

The workshop word is report when the analysis is composed of several successive sections — an object read from end to end, not a grid of checkboxes. It is set in Source Serif 4 and IBM Plex Mono, rendered as a book, and it arrives with the pieces that make it verifiable:

  • a making trace as a .zip — the steps that produced the report, anonymized of third-party data, attached alongside like a critical apparatus to an edition;
  • a cryptographic chain of custody: each piece sealed by an Ed25519 signature, which a single altered byte invalidates and reveals;
  • a qualified timestamp, RFC 3161, issued by a third-party authority, dating the dossier with no possible back-dating.

The Lab audits. It maps sources — a publication, a codebase, a video, a scattered corpus — and returns a report in which every claim can be verified against its source. A report is not an opinion; it is an opposable dossier — dated, signed, and carrying its own record of how it was made.

The license is split. The report, as an intellectual object, stays under my copyright; the forensic trace that produced it is open under CC-BY 4.0. The reader can pull a file from the .zip, re-read it, republish it — they cannot pass my report off as their own. The proof is open; the text is signed.