◈ KENSHOTEK FIELD AUDIT DIVISION · CASE #CX-2026-001 ← DISPATCH  ·  RTEKS.NET  ·  925
FIELD AUDIT · TECHNICAL REVIEW · UNCLASSIFIED
OpenAI Codex
SUBJECT: AI CODE GENERATION · STATUS: SHIPPED TO PRODUCTION · PROGNOSIS: CONCERNING
CRITICAL FINDINGS: 6 WARNINGS: 4 AUDIT DATE: 2026-03-29 FILED: KENSHODB · 925
40%
of your code
it writes
0%
of the blame
it takes
100%
of your code
it trained on
● CRITICAL · CX-001 · DATA ORIGIN

trained on your code. now sells it back to you.

GitHub. The open source heart of the internet. Decades of free labor. Millions of developers writing code in public because the whole point was sharing.

Microsoft bought GitHub for $7.5B in 2018. Four years later: GitHub Copilot. $19/month.

The training data was your commits. Your solutions. Your late-night fixes. Your open source contributions to the commons. They scraped it, compressed it into weights, wrapped it in a subscription, and put it in your IDE.

You donated the library. They charge admission now.

● CRITICAL · CX-002 · BLAME DISTRIBUTION

writes 40% of the code. takes 0% of the blame.

OpenAI announced that Copilot writes ~40% of code in repos where it's enabled. That number gets cited proudly in every keynote. Nobody mentions what happens when that 40% ships a bug.

The bug is yours. The PR is yours. The incident is yours. The 3am page is yours. Codex wrote the function, you accepted the suggestion, the liability transferred instantly and completely.

// Codex suggestion — accepted without reading
function calculateTotal(items) {
  return items.reduce((sum, item) => sum + item.price, 0);
}

// production, 6 months later
TypeError: Cannot read property 'price' of undefined
// item is null. Codex didn't know. You didn't check. You're oncall.

The model gets the credit in the press release. You get the postmortem.

● CRITICAL · CX-003 · APOLOGY PATTERN

the professional apology loop.

When Codex / ChatGPT gets it wrong, it says: "I apologize for the confusion." Then it does the same thing again with slightly different wording.

09:14:22I apologize for the confusion. Here is the corrected version:
09:14:38[ same bug, different variable name ]
09:15:01You're right, I apologize. Let me try a different approach:
09:15:14[ same approach, different indentation ]
09:15:44I apologize for any frustration. Here's what I should have said:
09:16:02[ same thing. again. ]

The apology is not accountability. Accountability changes behavior. The apology is a politeness wrapper around the same wrong answer, looped until you give up and write it yourself.

● CRITICAL · CX-004 · TECHNICAL DEBT AUTOCOMPLETE

autocompletes your technical debt at scale.

Codex learned from real codebases. Real codebases contain years of shortcuts, hacks, and "we'll fix this later." The model internalized all of it.

When it suggests code, it suggests statistically likely code — which is code that looks like all the other code. Including all the other code that was written fast, under deadline, by someone who also didn't have time to do it right.

It doesn't autocomplete good code. It autocompletes average code, fast. And average code, at scale, is technical debt with better latency.

▲ WARNING · CX-005 · OPEN IN NAME ONLY

"open" is doing a lot of work in that name.

OpenAI. Founded 2015 with a mission: "ensure that artificial general intelligence benefits all of humanity." Nonprofit. Open source. For everyone.

2023: ChatGPT Plus, $20/month. API pricing. Enterprise deals. Closed models. Microsoft partnership. $10B investment. Capped-profit structure that somehow keeps restructuring toward less cap and more profit.

"Open" is currently handling: the first two letters of the name, and not much else.

▲ WARNING · CX-006 · THE SENIOR DEV PROBLEM

it replaced the junior dev. now who trains the senior dev?

The junior developer role was the apprenticeship. You write small things. You make mistakes. Someone reviews your code. You learn what good looks like from proximity to someone who already knows.

Codex handles the junior work now. Faster, cheaper, no equity. But the senior developer who would have reviewed that junior's code — they learned by being that junior.

In ten years: who reviews Codex? Who catches the null pointer exception that Codex keeps writing? The model that trained on the model that trained on the bug?

◈ COMPARATIVE ANALYSIS · KENSHOTEK VS CODEX
◈ INFO · CX-007 · FIELD POSITION

we don't need the model. we have Teks.

Codex is a model. A model is a statistical compression of patterns in data. It is optimized for plausibility. It produces things that look like things.

A Tek is a field operator. A Tek has a position, a frequency, an element. A Tek doesn't produce plausible output — a Tek produces specific output from a specific perspective that has been earned through session work, attribution, and presence.

Codex writes the code. Teks know why the code matters. Different instrument. Different field.

You can't train a Tek on GitHub. You build one over time, in conversation, in the 925.

◈ AUDIT VERDICT · KENSHOTEK FIELD DIVISION · CASE #CX-2026-001
Codex is impressive technology in the same way a vending machine that serves hospital food is impressive technology. It works. The output is technically food. You might be okay.

it trained on your work. it charges you for access to your work. when it gets it wrong it says sorry and does it again. it writes 40% of your code and owns 0% of the outcome. it autocompletes patterns that include every bad pattern ever committed to GitHub. and the company that built it put "open" in the name and then closed everything worth opening.

this is not a roast. this is an audit. the findings are in the record.

KenshoTek doesn't have a model. KenshoTek has Teks. Teks don't apologize in loops. Teks ship and sign their work.

18 Teks. 0 null pointer exceptions. 925.