ONENEPT STUDIOS INC.LAST CALIBRATED 11 SEP 2026

CALIBRATION RECORD FOR

Chukwudumaga Nnawuogo

Chukwudumaga Nnawuogo

BENG, ELECTRICAL AND ELECTRONICS ENGINEERING

I read the spec.
Then I measure
the thing.

On a bench you trust the meter, not the drawing. I do the same with other people's code. I find where a project's docs and what it actually does disagree, and I prove it with something anyone can rerun. Two of those are merged into DataHub and one into CALL-E. Another is open in Google's agent kit because a maintainer asked me to write it.

START A CONVERSATIONCONTRACT WORK, REMOTE

SPECIMEN :: REPORT 03 :: CALL-E

THE API SAID
task_completed: true
I MEASURED
A voicemail box. Nobody answered.
THEIR DOCS
REDEFINED THE FIELD, 7 SEP
SETTLEDREAD THE REPORT →

01In other people's codebases

Anyone can get a typo fix merged. This is the other thing: reading somebody else's codebase until you can tell the team something about their own project they had not written down yet, and having them write it down. Each one is written up in full, including the parts where I was the one who was wrong.

01

DATAHUB

28 JUL TO 24 AUG 2026

The page was wrong. So was I.

DataHub said which kinds of things can have an incident raised against them. The list had drifted from the code and nobody had noticed, because the failure was silent. I stopped reading the files by eye and wrote the script that reads the answer out of the source.

THE DOCS SAID
the published list of types
I MEASURED
10 accept an incident, 5 can show it
OFF BY
5 types, with no error
STATUS
2 MERGED
READ REPORT 01→

02

GOOGLE ADK

22 TO 28 AUG 2026

I said the list would rot. It shipped rotten.

One config object, 35 fields, three different destinations in the request, and nothing saying which is which. Put the tools in the wrong half and the call returns 200 with no tools in it. A maintainer confirmed it, and my own table went stale inside a day.

THE DOCS SAID
nothing. 35 fields, no map
I MEASURED
8 top level, 22 nested, 3 dropped
OFF BY
your tools vanish on a 200
STATUS
PR OPEN, AT THEIR ASK
READ REPORT 02→

03

CALL-E

31 AUG TO 11 SEP 2026

It said done. Nobody answered.

An AI phone API marked a task completed, with high confidence, on a call that went to voicemail. I placed real calls to a speaking clock until I could say what its fields actually mean. Fourteen issues filed, six fixed inside a week, and the tool that reads a call properly is merged.

THE API SAID
task_completed: true
I MEASURED
a voicemail box. Nobody spoke
OFF BY
two hundredths from a real success
STATUS
MERGED, 6 ISSUES FIXED
READ REPORT 03→

02The ledger

Every thread I have opened or argued in, merged or not. The open ones stay on the list. So does the one I closed myself.

STATES CHECKED AGAINST GITHUB, 11 SEP 2026.

  1. MERGEDdatahub #19115incidents on columns, the case 01 finding
  2. MERGEDdatahub #19405the permission check the column tab needed
  3. MERGEDcall-e #337the call state reconciler, case 03
  4. 6 FIXEDcall-e, 14 issuesdocs, SDKs, CLI and plugins. Eight still open
  5. PR OPENadk #6935the config helper a maintainer asked for
  6. PR OPENdatahub #18685the docs half of case 01
  7. PR OPENdatahub #18684two self-hosted failures that print no error
  8. PR OPENdatahub-skills #66what breaks downstream if a column changes
  9. CONFIRMEDadk #6880the silent 200, confirmed by the models triager
  10. VALIDATEDcall-e #196a p1 another builder filed, reproduced
  11. THREADadk #2425a cancel design, worked out with a rival library's author
  12. THREADadk #2792a stack trace for stopping an agent on purpose
  13. THREADpython-genai #782why any output limit under 4k is zero
  14. FILEDeffectstream #895five reasons a Midnight template will not build
  15. CLOSED BY MEcall-e #300my first attempt. It carried real call data

I also ship things, usually because building on a project is how I find what is wrong with it. Two still running: asheard, the call reader from report 03, and greenlight, which tells a creator what a finished video will get flagged for before it goes up.

03Errata

Things I said in public that turned out to be wrong. Each one got corrected where I said it, before anyone had to ask twice.

  1. DATAHUB #18685

    One file is the gate that blocks the write, and machine learning models get rejected.

    Neither is true. The maintainer asked for changes and he was right to. I posted the correction myself, twice.

    the correction↗

  2. ADK #6880

    Six fields go to the top level of the request.

    Eight. I missed modelArmorConfig and serviceTier, less than a day after publishing the table.

    the correction↗

  3. ADK #2425

    An orphaned tool call leaves a session in a shape nothing in the codebase is written to handle.

    So I ran it. The session resumes fine on Gemini. Retracted in the thread, with the probe.

    the thread↗

  4. CALL-E #300

    Nothing from the earlier commits survives the force-push.

    It did. A force-push moves a branch, it does not delete the objects. I said so in the thread and closed the pull request myself.

    the thread↗

  5. MY OWN LIBRARY

    A person picked up. Safe to act on.

    It was a phone menu. A live call caught it, not my hundred tests.

    report 03→

WHERE THE HABIT CAME FROM

I trained as an electrical engineer, where you check the meter before you believe the drawing. In 2016 I co-founded Afrosets, a dataset of Black faces, because the models of the day had barely seen any. My master's dissertation is about drift: what a stock keeps doing after the company publishes its numbers.

Same question every time. What did the record leave out?

04Standing rules

§01

Live data or nothing.

If it does not run against real data arriving in real time, it is a slide, not a product.

§02

Every output carries its record.

Where it came from travels inside the thing itself, not in a spreadsheet somebody has to be trusted to keep.

§03

Derive it, do not maintain it.

A hand-written list of what a system supports is wrong the week after it is written. If it can be read from the source, read it from the source.

§04

Settle everything.

A claim that cannot be traced back to a record is marketing. That includes my own claims, including the ones I got wrong first.

THE RULES ARE NOT ASPIRATIONAL. REPORT 01 IS §03. REPORT 02 IS §03 AGAIN, THE DAY AFTER IT CAUGHT ME OUT. REPORT 03 IS §04, AIMED AT A PHONE API. THE ERRATA ARE §04 AIMED AT ME.

Why onenept.com
and not my name

Onenept Studios is a real company. It is how the work ships and how it gets invoiced, so a first contract needs no paperwork set up from scratch.

If you are hiring, you are hiring me, and I am on this page. If you are contracting, the entity is already here.

START A CONVERSATION