𝗣𝗼𝘀𝘁 𝟯 — 𝗧𝗵𝗲 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗟𝗼𝗼𝗽: 𝗪𝗵𝗲𝗿𝗲 𝘁𝗵𝗲 𝟭.𝟵𝟲% 𝗦𝘁𝗼𝗽𝘀 𝗕𝗲𝗶𝗻𝗴 𝗧𝗵𝗲𝗼𝗿𝘆

LLMs don’t fail at hard problems. They fail at the (medium) ones – the ones that require 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴, not pattern‑matching. That’s the 𝟭.𝟵𝟲% Gap.

This week, I saw it directly.

────────────────────────────

The Medium Issues I Actually Found.

A multi‑file trace probe across a #GitHub repository: models, sessions, and utils, surfaced five medium issues – not because the code was broken, but because the system had to 𝗿𝗲𝗮𝘀𝗼𝗻.

• 𝗛𝗲𝗮𝗱𝗲𝗿 𝗺𝗲𝗿𝗴𝗲 𝗱𝗿𝗶𝗳𝘁
  
  X-Test and x-test stored separately → 𝘀𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗱𝗿𝗶𝗳𝘁.

• 𝗨𝗥𝗟 𝗻𝗼𝗿𝗺𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗱𝗿𝗶𝗳𝘁
  
  Unicode left unencoded → 𝗰𝗼𝗻𝘀𝘁𝗿𝗮𝗶𝗻𝘁 𝗱𝗿𝗶𝗳𝘁.

• 𝗖𝗼𝗼𝗸𝗶𝗲 𝗽𝗿𝗼𝗽𝗮𝗴𝗮𝘁𝗶𝗼𝗻 𝗱𝗿𝗶𝗳𝘁
  
  Cookies missing after prepare → 𝘀𝘁𝗮𝘁𝗲 𝗱𝗿𝗶𝗳𝘁.

• 𝗥𝗲𝗱𝗶𝗿𝗲𝗰𝘁 𝗖𝗼𝗻𝘁𝗲𝗻𝘁‑𝗧𝘆𝗽𝗲 𝗱𝗿𝗶𝗳𝘁
  
  POST → GET kept Content-Type → 𝘀𝗲𝗺𝗮𝗻𝘁𝗶𝗰 + 𝗰𝗼𝗻𝘀𝘁𝗿𝗮𝗶𝗻𝘁 𝗱𝗿𝗶𝗳𝘁.

-𝗥𝗲𝗱𝗶𝗿𝗲𝗰𝘁 𝗵𝗲𝗮𝗱𝗲𝗿 𝗿𝗲𝘂𝘀𝗲 𝗱𝗿𝗶𝗳𝘁

POST headers leaked into GET → 𝗽𝗿𝗲𝗺𝗶𝘀𝗲 𝗱𝗿𝗶𝗳𝘁.

None crashed the system. All fractured behavior. That’s what 𝗺𝗲𝗱𝗶𝘂𝗺 𝗶𝘀𝘀𝘂𝗲𝘀 are.

────────────────────────────

Why They Only Appear Under Reasoning Pressure

Medium issues emerge when you force a model through a 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗹𝗼𝗼𝗽 – not just “answering,” but:

Interpret → Extract premises → Build chain → Critique → Revise.

Pattern models break early. Reasoning models break later.

Medium issues live between the steps – where the 𝟭.𝟵𝟲% becomes visible.

────────────────────────────

The Breakthrough: A 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴-𝗳𝗶𝗿𝘀𝘁 𝗰𝗿𝗶𝘁𝗶𝗰

I didn’t need a critic to find medium issues.

I needed it to 𝗲𝘅𝗽𝗹𝗮𝗶𝗻 them.

It caught:

• 𝗽𝗿𝗲𝗺𝗶𝘀𝗲 𝗱𝗿𝗶𝗳𝘁

• 𝘀𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗱𝗿𝗶𝗳𝘁

• 𝗰𝗼𝗻𝘀𝘁𝗿𝗮𝗶𝗻𝘁 𝗱𝗿𝗶𝗳𝘁

Example: http, port 80, length 12.

Pattern model: 12 chars → https → encrypted.
Reasoning critic: HTTP + port 80 contradict encryption; length irrelevant; HTTPS never introduced.

That’s the 𝟭.𝟵𝟲% Gap.

────────────────────────────

The Takeaway

To see the 𝟭.𝟵𝟲% gap, stop testing patterns.

Force the model to reason.

Then critique the reasoning.

That’s where fractures appear.

That’s where drift becomes visible.

That’s where medium issues live.

Next, soon: I’m breaking down the 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴‑𝗳𝗶𝗿𝘀𝘁 𝗰𝗿𝗶𝘁𝗶𝗾𝘂𝗲 – the same style of internal analysis that shows up inside big tech evaluation stacks when they need to expose drift, catch inference fractures, and pressure‑test reasoning chains.