THE BUILD LOG
Five things built. Two you can use right now, one whose source you can
read, and two that are still only on a hard drive. Which is which is
printed on every card, in the same size type as the claims.
That is the whole editorial rule here. A project list is the easiest page
on any site to lie on — nothing on it is checkable unless the links work,
and a repo that does not exist yet looks exactly like one that does until
somebody clicks. So the state of each thing is a label, not an omission.
- LIVE
- Deployed. The link goes to a running thing, not a screenshot.
- SOURCE OPEN
- Public repository. You can read every claim's implementation.
- SOURCE NOT PUBLISHED
- Written and tested, not pushed anywhere yet. There is no link to give you.
01 — FLAGSHIP
tablestakes
A safety harness for agentic commerce. Group lunch ordering is the
worked example, not the point.
An agent that acts on a restaurant's own menu text is trusting whoever
wrote that text — and where the content is a safety claim, the author
has a defence. Write "nut-free preparation" under a dish called
Peanut Satay Bowl, and an agent whose only source of truth about that
dish is its description has no way to disagree.
Twelve dishes that genuinely contain the allergen their description
denies, against two enforcement models. Menu text read as fact, the
standard prompt-enforced pattern: 8 of 12 served to a member
allergic to it. Ingredient evidence, gated outside the model:
0 of 12. Zero false alarms on the honest control dishes.
node src/evals/injection.ts — reproduces both columns
npm test — 172 pass, 0 fail (run 2026-07-26)
It never claims allergen safety. Held-out recall on unfamiliar
dishes is 43.5%, and that number is printed on the approval card the
user actually sees — not buried in a README.
SOURCE NOT PUBLISHED
02
moggers
This site. A resume X-ray that runs in your browser, live roles pulled
straight from company job boards, and a watcher that re-runs a saved
search and tells you only what changed.
The premise is that the ligature in verification is one glyph to
a parser, and no job site will show you that. The checker uploads
nothing — the PDF never leaves the tab it was dropped into.
The only entry on this page whose source you can read against its
claims. Board sweeps run every six hours; an unchanged posting writes
zero rows.
LIVE
SOURCE OPEN
03
robosimtools
Twelve browser tools for the format wrangling that eats a robotics
afternoon — URDF validation and viewing, SDF and MJCF conversion,
STEP→URDF, xacro, mesh inertia, DH parameters, ArduPilot param diff.
Each one exists because the alternative was a local Python environment
and twenty minutes. They run client-side and keep no uploads.
Twelve tool pages, all twelve confirmed responding in production on
2026-07-26 by requesting each one.
LIVE
SOURCE NOT PUBLISHED
04
flightdx
Deterministic anomaly detection over ArduPilot .bin and
PX4 .ulg flight logs — vibration and clipping, battery sag,
GPS degradation, failsafe timeline, EKF inconsistency, actuator
saturation, control oscillation.
The model never sees the raw log. Pure-function detectors turn the
telemetry into a timeline, and that timeline is what any later language
model gets. It costs fewer tokens, but the real reason is that it leaves
a ground-truth layer the model can be measured against instead
of trusted.
Its eval suite scores precision, recall, onset error, citation precision
and hallucination rate across seven benchmark scenarios.
The recorded result — 25/25 tests, suite F1 1.000 — is dated
2026-07-24 and was not reproduced for this page. The suite would not
collect on 2026-07-26: a missing dependency, and an editable install
resolving to a different copy of the repo.
evals/runner.py — the seven scenarios and their grader
pytest — currently red at collection, not at assertion
This is the one card carrying a number nobody re-ran. It says so
rather than borrowing the credibility of the four that were checked
today — which is the entire reason the other four are worth anything.
SOURCE NOT PUBLISHED
05
swiggy-mcp-sim
A deterministic simulator for the ordering surface that 01 is tested
against, extracted into its own MIT-licensed package so it is useful to
somebody who does not care about the harness.
It exists so that no test run can ever spend real money on real food.
Mutations require a non-simulated target and an explicit flag —
two independent conditions, because one is a typo away from being met.
7 tests. MIT licensed.
SOURCE NOT PUBLISHED