CR Lite · Syndicate

What CR Lite can do

One engine, from a single terminal to a full AI assembly line. It doesn’t just answer — it checks, repairs, proves, and delivers. Everything below is built and tested today.

It keeps improving the answer

The AI writes. The system checks. Repeat until it passes.

  • Every answer goes through a write → check → fix cycle, up to a set limit
  • It always keeps the best answer so far — a retry can never make things worse
  • It can write several drafts at once and keep the strongest one
  • If the AI starts repeating itself, the loop stops instead of wasting time and money
  • You can cancel any run mid-flight and keep the best answer it found

Rules the AI cannot break

The checks are real code, not the AI grading its own homework.

  • Each job gets a checklist of things a good answer must cover — and the AI cannot talk its way past it
  • Code climbs three tiers of checking — Tier 1 it compiles, Tier 2 it runs in a locked sandbox, Tier 3 it produces the answers the spec’s own examples expect — across 15 programming languages
  • If the sandbox is not available, code is never run — it fails closed instead of pretending to pass
  • Expansion packs add your industry’s rules, and those rules become part of the checklist

Proof behind every answer

No evidence, no pass.

  • A factual answer must cite real sources it actually gathered — a made-up citation gets flagged
  • It can search the live internet and read web pages, and what it finds becomes the evidence
  • It researches a question from several angles — not just one search phrase
  • It checks whether independent sources agree, and shows you when they conflict instead of hiding it
  • The AI’s own opinion of its answer is never the final word — the checks are

It remembers what it learns

Ask once. The proven answer sticks around.

  • Answers that passed all checks are saved, along with their sources
  • Ask the same thing in different words and it still finds the saved answer
  • A question it already answered skips the fresh research — fast, even on a small local model

Tools it can use — safely

Reading is free. Acting needs a permit.

  • Live web search and page reading — it looks things up and reads pages, and what it finds becomes the evidence
  • Reads your files and searches your code to ground its answers
  • Git-aware: it sees your changes, explains diffs, and guards commits — but never pushes or rewrites your repo on its own
  • Scans for leaked secrets before anything is saved or shared
  • Connect the systems you already use — their tools join the same rules
  • Tools that only read run freely; tools that change things wait for your approval first
  • Anything that goes wrong becomes a plain-language alarm you can act on

Your models, your choice

Free local models, paid cloud models, or both at once.

  • Assign a different model to each job: thinking, coding, tools, checking
  • Mix a paid cloud model for the hard parts with free local models for the rest
  • Every run shows exactly what it cost — and what you saved versus all-cloud
  • If a cloud model goes down, it tells you and falls back to your local model — never a silent failure
  • Local models run on your hardware: the model must FIT your device — its weights and context have to sit in your free memory (RAM or graphics card). A model bigger than your machine will crawl or fail to load. Cloud models have no such limit.

Build whole assembly lines

Connect blocks into a pipeline. Only good work flows forward.

  • Wire AI blocks together like stations on an assembly line
  • A block’s output only moves forward if it passed its checks — bad work stops at the wire
  • When several streams merge, each part stays labeled with where it came from
  • An input block reads a whole folder of client files — any file type, each one fingerprinted
  • An output block delivers the finished product to a folder, and never overwrites anything
  • Even outside programs can join the line as blocks — behind the same safety approvals

The workbench

Draw it. Run it. Watch it work.

  • Draw your pipeline on a canvas — add blocks with one click, wire them by tapping
  • Save your work and reopen it exactly as you left it
  • Run once, or run continuously until you hit stop — problems pause the line instead of looping forever
  • Open several workbench windows at once; each runs its own pipeline
  • A station screen shows history, alarms, and every pipeline that is running
  • Ships as an installable desktop app

What it doesn’t do yet — honestly

  • Building the expansion packs themselves happens in a separate system; CR Lite runs the finished packs.
  • Memory search is exact-and-similar today; a faster large-scale index comes when libraries grow into it.

Ready to run your own AI assembly line?

Start with one block in a terminal, or draw a whole pipeline in the workbench. Your models, your rules, your proof.