Walkthrough: conversation → document → colophon

You prompt two agents. They each write a markdown page. You want more than a finished file. You want a journal, a session record, and a proposed colophon you can seal to the bytes.

This page is the story, not the raw JSON. Three human prompts go into agent sessions (Claude and Codex in the fixture). The Innsigle session skill builds session provenance and proposes a colophon.

Spec detail (optional): session provenance feature.

Human prompts → agent writes markdown → journal + L2 provenance → colo.json
 ↓
 review · optional sign

Step 1: Human prompts

The operator types these (same sequence for each agent in the fixture):

You

Draft a short markdown page titled Sealed Notes explaining what an Innsigle colophon is in two paragraphs.

You

Add a bullet list of three non-goals: no AI detection, no purity score, no claim of content truth.

You

Tighten the opening sentence; keep model-primary honesty.

Full list: human-inputs.json.

Step 2: Agent writes the document

The model produces a markdown file (model-primary: the agent wrote the bytes).

Claude (fixture)

Writes sealed-notes-claude.md: two paragraphs on the colophon, then the non-goals list.

The document (Claude session)

Sealed Notes

An Innsigle colophon declares how a piece of work was made, composition and ingredients, not whether the prose is true or pure.

It lists models, tools, and human roles so readers can see how the piece was produced.

Non-goals

No AI detection
No purity score
No claim of content truth

Codex produces a parallel page: sealed-notes-codex.md. Claude artifact: sealed-notes-claude.md.

Step 3: Skill captures the session

The harness records a journal (JSONL): each human prompt, assistant turn, file write (by: model), tool/skill calls. That is not the colo yet; it is the raw machine log.

From the merged Claude + Codex journals the driver reports:

Metric	Value
User prompts	6 (3 per agent)
Assistant turns	5
Models	Claude, Codex
Composition suggestion	model-primary

Rewrite tools (e.g. sloptimizer) appear as ingredients. They do not turn the work into human-authored.

Step 4: Proposed colophon

This is what a reader (or claim) should show next to the document:

Colophon · model-primary

Kind	Name	Role
model	Claude	primary
model	Codex	primary
tool	innsigle-session	provenance-capture
tool	sloptimizer	rewrite
human	operator	prompter

session metrics: user_prompts=6; assistant_turns=5

Linked provenance: session-provenance.json

Machine form: colo.json.

Step 5: How it looks on a sealed page

Footer pattern (same idea as the Sample signed page):

Innsigle The maker's seal · model-primary · Claude + Codex

When you sign, the claim binds content SHA-256 + this colo + absolute issuer key_url. Verification checks the seal, not whether the prose is true.

Reproduce

npm run test:provenance-driver

Check it

Expect the driver to propose model-primary with models + tools (including rewrite tools) listed as ingredients, not a flip to human-authored. Details: Session provenance.

Issuer without a server: Issuer identity.

Machine pack

File	Role
human-inputs.json	Prompts
session-claude.jsonl	Claude journal
session-codex.jsonl	Codex journal
session-provenance.json	Session record
colo.json	Proposed colo