BraiinsRatchet/START_HERE.md

176 lines
4.3 KiB
Markdown
Raw Normal View History

2026-04-25 21:41:02 +00:00
# Start Here
This project now has one operator entry point:
```bash
./scripts/ratchet
```
That is the same as:
```bash
./scripts/ratchet next
```
2026-04-27 13:21:42 +00:00
It prints the cockpit: current state, exact next action, interpretation, reference commands, and ratchet rule.
2026-04-25 21:41:02 +00:00
## Your Job
Your job is not to understand every metric.
Your job is:
1. Run `./scripts/ratchet`.
2026-04-27 13:21:42 +00:00
2. Do only what it says under `DO THIS NOW`.
3. Ignore every other command unless `DO THIS NOW` tells you to run it.
2026-04-27 13:32:46 +00:00
4. If it tells you to run `./scripts/ratchet watch 2`, start it, leave the terminal open, and come back after about 2 hours.
2026-04-25 21:41:02 +00:00
5. If you manually place a Braiins canary, write down the order details outside this repo and wait through the maturity window before judging it.
2026-04-27 13:32:46 +00:00
## Who Is In Control?
If `watch` is running, the Python process is in control of that terminal.
You do not need to babysit it. It will:
1. Collect samples every 5 minutes.
2. Write the run report when it finishes.
3. Print the cockpit again.
4. Return control to your shell prompt.
If you want the technical report, run `./scripts/ratchet report`. The normal workflow intentionally shows the cockpit first.
2026-04-27 15:59:30 +00:00
After a watch finishes, the cockpit enters a post-watch cooldown. That is deliberate.
Post-watch cooldown means:
1. The current experimental stage is complete.
2. Starting another identical watch immediately is not useful ratcheting.
3. The run report is the evidence artifact.
4. The next planned touch is a later fresh sample, usually `./scripts/ratchet once`.
2026-04-27 16:12:47 +00:00
During cooldown, the cockpit shows:
1. A progress bar.
2. The earliest next action time.
3. The remaining minutes.
## Controlled Automation
If you do not want to babysit the cooldown manually, run:
```bash
./scripts/ratchet pipeline
```
The pipeline first prints a proposal like:
```text
I am going to: wait until this time, run one fresh sample, print the cockpit, then stop.
Are you OK with this? Type yes or no.
```
It only runs after you type `yes`.
It is still monitor-only. It never places, changes, or cancels Braiins orders.
## Forever Supervisor
For the full autoresearch lifecycle, run:
```bash
./scripts/ratchet supervise
```
The supervisor is the long-running engine. It:
1. Loads persisted state from `data/ratchet.sqlite`.
2. Waits through cooldown if cooldown is active.
3. Runs the next passive watch when due.
4. Writes reports and lifecycle events.
5. Re-enters cooldown.
6. Repeats until you stop it.
If it crashes or the Mac reboots, start the same command again. It resumes from SQLite.
Use this to inspect persisted state without starting the loop:
```bash
./scripts/ratchet supervise --status
```
## Native Mac App
The native SwiftUI shell is in:
```text
macos/BraiinsRatchet
```
Run it from source:
```bash
cd macos/BraiinsRatchet
swift run BraiinsRatchetMac
```
The app is a native cockpit over the same durable Python lifecycle engine.
2026-04-27 13:41:25 +00:00
## Research Pathway
The cockpit has two different time horizons:
1. `DO THIS NOW` is the only command you should run next.
2. `Ratchet Pathway Forecast` tells you what the next stages probably look like.
The forecast is not a profit prediction. It is a workload and research-flow prediction.
It is split into:
1. `Immediate`: what happens now.
2. `Midterm`: what probably happens after the current run or sample.
3. `Longterm`: what could happen after multiple reports mature.
Expect the pathway to change after each report. That is the point of ratcheting: the next stage adapts to measured evidence instead of following a rigid plan.
2026-04-25 21:41:02 +00:00
## What The Actions Mean
`observe` means do not bid.
`manual_canary` means a tiny research experiment is inside the configured loss budget. It is not a profit signal.
`manual_bid` means the stricter profit-seeking guardrails cleared. The code still does not place the order. You decide manually in Braiins.
## Where The Reports Are
The master ledger is:
```text
reports/EXPERIMENT_LOG.md
```
Each completed watch creates one run report:
```text
reports/run-*.md
```
Older sessions can be embedded with:
```bash
./scripts/ratchet retro START_UTC END_UTC
```
## The Ratchet Rule
One run is not a verdict. One run is a measurement.
Only change one knob at a time:
1. Depth target.
2. Overpay cushion.
3. Canary spend.
4. Duration.
5. Timing window.
Do not increase spend until multiple mature runs point in the same direction.