Mesa dos 28

The problem

People who work and study all day have no time for an AI course. And what the leading names in the field teach is scattered across videos, posts and courses, and it changes every week.

At the same time, building with AI is fast, but the result arrives without a trail: unexplained decisions, tests that never ran and "facts" nobody checked. Reviewing everything alone does not work, and calling in every expert for every question is expensive.

What it does

1. 1
   Follows those who teach

   Every Friday, AI assistants look up what each reference published that week, in that person's own sources, and update the 28 guides based on their work.

2. 2
   Calls only those who know the topic

   You write the question and, before it, what you bet you will hear. The table calls 3 to 5 guides who know that subject, not all 28.

3. 3
   Points to the file and the line

   Each guide reads the project, unable to change anything, and points to where it saw each thing. A check confirms that the place pointed to exists; whatever does not exist is removed.

4. 4
   Delivers the list, you decide

   At the end, a short list of what to improve, each item with a way to check that it is done. The decision line is left blank for you.

On screen

1Bet2Topic3Lenses4Checks5Decides

Tap outside the image to close it.

Tap outside the image to close it.

Tap outside the image to close it.

Tap outside the image to close it.

Tap outside the image to close it.

What changes

People with no time for a course learn, every week, from what the leading names in the field are teaching, and see that advice applied to their own project, with the exact place for each piece.

With all 28 at once, the round also pointed to 5 places that did not exist. With only the 5 for the topic, all 108 pointers were checked, and the cost is estimated before starting: if it goes over the limit, the round does not start without your go-ahead.

And the same approach can serve any field where it pays to learn from the references and review your own work through their eyes: the same engine already runs a table of 3D designers and a board for the thesis.

Problem, improvement and result

Problem

Learning applied AI in depth without time for a course, and reviewing projects built with AI, which arrive without a trail: unjustified decisions, tests that never ran and citations nobody checked. Keeping up with what 28 references publish every week takes hours, and convening all of them for every review is expensive.

Improvement

28 guides with citation, link and date, updated every Friday by a scheduled task that only accepts the reference's own sources, are turned by code into read-only reviewer subagents. Per topic, a panel of 3 to 5 lenses, with tokens estimated before spending and a refusal above 2 M; each file:line citation is checked on disk, and a moderator consolidates a backlog with acceptance criteria.

Result

On September 16, 2026, the round with all 28 lenses on the table itself cost 8.45 M tokens and invented 5 citations. On September 18, with the 5-lens panel: 914,000 tokens (about 89% less), 108 of 108 citations checked on disk (3 fixed on the second try) and an estimate of 1.36 M against the 2 M cap, computed before spending.

Evidence

• 108 of 108 citations checked on disk

  In the September 18, 2026 round in which five lenses reviewed the table itself: each file:line citation is checked against the file, and 3 were fixed on the second try. The check proves the line exists, not that it supports the claim.

  Measured.

• 58 test functions, with mutation testing

  Counted in the public cut of the code on October 30, 2026; pytest with ruff and mypy. Mutation testing plants one defect at a time and checks whether the suite catches it.

  Measured.

• Question topic with 85% accuracy (ECE 0.08)

  laya-tema, evaluated a single time on a locked final set of 108 questions; on the same test, the keyword rule got 50% right.

  Measured.

On screen

1Bet2Topic3Lenses4Checks5Decides

Tap outside the image to close it.

Tap outside the image to close it.

Tap outside the image to close it.

Tap outside the image to close it.

Tap outside the image to close it.

How it works

Six steps, from the question to the backlog. The model reads and writes; the code checks.

1. Bet

   The person writes the hypothesis first; skipping is possible, but it is recorded.

   Python
2. Topic

   laya-tema suggests one of 9 topics; on its own it only applies at 0.8 confidence or more.

   laya-tema85% accuracy
3. Context

   Builds the context, runs the project's tests, masks secrets and estimates the cost.

   Python2 M token cap
4. Lenses

   Three to five read-only subagents write reviews citing file and line.

   Antigravity CLI
5. Citations

   The code checks each citation on disk; one that fails gets a correction and is then removed.

   Python108 of 108
6. Backlog

   The moderator merges everything into up to 12 items, each with an acceptance criterion.

   Python

Testes do projeto: VERDE
Citações conferidas no disco: 108 de 108 existem
Tokens: 914 mil
Orçamento: estimativa 1.36 M · teto 2.00 M

Decisions

Limits

• The check guarantees that the cited file and line exist; checking whether the line supports the claim is the next step.
• The numbers come from rounds of the table on its own code; the next step is the pre-registered protocol, with 10 seeded defects and 3 decoys, which measures the table against a single reviewer.