Demo recordings

The recording scripts under demos/ capture attest in use: one per major surface, a real Hermes agent calling each of the four *.ask routers, and an install on the hosting platform attestation ships through. Every recording below is real output from a real run; nothing is staged or re-typed.

The whole story, narrated

Four minutes, for experimental and computational scientists: installing on a hosted agent, catching a wrong number in a draft, asking the agent, reading the literature, and exact mathematics. Every clip is a real recording; the voice is Piper, synthesized locally. demos/film/build.py rebuilds it from demos/film/scenes.toml.

Installing on a hosted agent

The Research Assistant starter on AgentMarkit, installed on the platform's own template image the way its Worker does it: the starter package, then the install script that clones and pins attestation, then every automated smoke check. Recorded with lab/scripts/rehearse_starter.sh from the AgentMarkit repository.

The same install with a model attached (gemma4:12b), a first ingest, and three real questions. Two are answered from attestation. The third reaches feed.ask but the model cut the question to a bare topic, and the router, which does not guess, asks back:

The run ledger and claim checker

attest runs scan reads experiment artifacts off disk, runs compare ranks the arms of a sweep and says what it cannot conclude, and attest claims checks every number in a draft against the run that produced it — one contradicted, one unsupported, and a coverage lint for numbers no claim covers. No model, no network.

A lab and a simulation

The ledger and the claim checker over an experimental project and a computational one: a catalyst screen (yield, %) and a basis-set convergence study (mean absolute error, kcal/mol). The draft carries one stale number, one result from an experiment never run, one citation key no reference list holds, and a catalyst loading nothing backs. The numbers are invented for the demo, and demos/science/workspace/FINDINGS.md says so.

Recording results

attest runs record for results that exist only in a lab notebook -- here three annealing temperatures and a conductivity for each. It writes a result and a config per arm in the shape runs scan reads, declares the metric's direction, and with --scan compares the arms in the same call, caveats included.

Citations

attest claims with cite= keys, then the cite.* tools over MCP: the lint reports a key no configured source resolves, and says plainly that this is about keys, not about whether the numbers are right.

The reading graph and symbolic math

kg.* over a graph derived from real tags — the concept vocabulary, central topics, a path between two concepts and an honest "no path" — then sym.* doing algebra exactly, including a rule-by-rule derivation.

The reference library

A BibTeX file synced into the deduplicated library, searched, one reference's real citation list followed both ways (what it cites, what cites it, and which of those are already in the library), and a filtered .bib written back out. No model, no network.

Going out to look

attest research searches PubMed and CrossRef, keeps each hit in the reference library under the source it came from (two versions of one preprint merge into one reference), and attest sources add makes the same query a standing topic that every ingest re-runs. Needs the network; arXiv is left out here only because its export API throttles hard.

The feed web UI

attest serve's HTMX page: a persona's ranked feed, marking one item useful and one not, and the caveat line changing as the ranking learns.

A real agent, not a script

These drive an agent rather than calling tools directly. Asked which basis set did best, Hermes calls runs.ask and relays both caveats, including "too close to call" (gemma4:12b):

The same for a machine-learning sweep:

…and a feed question by calling feed.ask:

Acting on one item: the agent lists the reader's papers, then calls feed.ask again with that item's item_id and question="useful", which records the judgement the ranker trains on. On gemma4:e2b the reply lists the papers without saying it rated one; the second call in the trace is the rating.

The reading graph through kg.ask ("What are the main areas I read about?"), and exact calculus through sym.ask (recorded on gemma4:12b; on gemma4:e2b the model printed the call as text instead of making it):

What each one shows

directory	shows	needs
ledger/	attest runs scan/list/compare and attest claims over examples/workspace/	nothing — pure local computation
claims/	attest claims plus cite.* over MCP, over examples/citations/	nothing — pure local computation
kg-symbolic/	kg.* and sym.* over MCP — neither has a CLI command or web page	a model server, once, to seed real tags
feed/	the HTMX web UI: a ranked feed, marking items useful, the onboarding form	a model server, to seed tagged items
science/	ledger and claims over a catalyst screen and a basis-set study, with citation keys	nothing — pure local computation
record/	attest runs record --scan writing notebook results and comparing them	nothing — pure local computation
film/	the narrated film: every clip stitched with a local Piper voiceover	ffmpeg and a Piper voice
library/	attest library sync/search/related/export over examples/molecular-ai/	nothing — pure local computation
research/	attest research against PubMed and CrossRef, then tracking the topic	network
hermes/	a real agent calling runs.ask, feed.ask (listing and rating), kg.ask and sym.ask over MCP	a live Hermes install and a model server

Unlike examples/*/, these are not golden paths. They produce video rather than a pinned output line, and most need a model, so tests/test_golden_paths.py does not run them — which is also why they are verified by hand rather than by CI.

Running one

The two model-free demos need no setup:

bash demos/ledger/narrate.sh examples/workspace
bash demos/claims/narrate.sh examples/citations

narrate.sh is the inner script — it echoes and runs each command in turn. record.sh wraps it in asciinema and converts the result with agg:

uv tool install asciinema
cargo install --locked --git https://github.com/asciinema/agg
bash demos/ledger/record.sh

The graph and feed demos need a seeded database first, because a graph built from the offline stub's placeholder tags has nodes named existing and title — not worth recording:

uv run python demos/kg-symbolic/seed_kg_db.py /tmp/demo.db   # needs a model
ATTEST_DB=/tmp/demo.db bash demos/kg-symbolic/record.sh

Two things that look like bugs and are not

demos/kg-symbolic/demo.py reads its database from ATTEST_DB, never from a positional argument. A path passed as an argument is silently ignored and the demo runs against whatever resolve_db_path finds. Against a large real database, kg.communities returns an alphabetical sprawl rather than the four coherent clusters the seeded fixture gives — which reads as a clustering defect and is an invocation mistake.

The MCP server's INFO logs go to stderr. Running demo.py by hand looks noisy for that reason; record.sh already redirects them, so the recording itself is clean.

Output

Each record.sh / record.py writes to the repo-root demo/ directory, already gitignored, unless given a path as its first argument — feed/record.py always writes demo/feed.webm and ignores that argument. demos/**/*.cast, *.gif, *.webm and *.db are gitignored too, so a recording made in place is never committed by accident.

Last verified

The original six ran green on 2026-09-22 against main. record/, library/, the three new agent turns and the two AgentMarkit videos were recorded on 2026-09-28 against attestation 0.2.2. research/ was recorded after #25: its first run turned up CrossRef hits dated 2115 and PubMed hits carrying a cited paper's DOI, both fixed before it was shown. demos/README.md carries the per-demo evidence.