Cash to ash?

A prompt asking a model to respect a limit is a suggestion. The limits are applied afterwards, in code, to whatever it proposed. These are the shipped defaults. Each one is a setting.

• No position above 20% of the portfolio.The batch is checked as a whole, so five small buys cannot do what one large one may not.
• One trade up to €100. One week up to €100.An approved buy reserves its cash until it is executed or rejected.
• Only cash that is actually there funds a buy.Planned contributions are noted, never spent.
• EXIT needs a condition you wrote down to have happened.A bad week is not one. Anything less becomes TRIM.
• A price older than 4 days, or a missing FX rate, blocks any trade.A review still goes through. Money does not.
• Advice expires after 7 days.No decision can be recorded against advice a later run withdrew.

Every default lives in src/config.py beside a note on how it was chosen. main.py process prints the ones in force.

• sourcedCites a supplied evidence item by ID and quotes it exactly. URL and date come from the item, not the model. An unknown ID or an invented quote demotes it to unanswered.
• backgroundWhat the model brought with it. Never enough on its own to move money.
• unansweredStays unanswered, in storage and on your phone. An open question is a finding.

A quote check proves the excerpt exists, not that the conclusion follows. That part is still reading.

Your thesis is yours. Research that finds it weakened writes a proposed version beside it and stops. The baseline the system is judged against stays the one you wrote.

The portfolio this was built for is small enough that a weekly pipeline over several models can cost more in a year than any excess return it could earn. That is the accepted trade: a better research habit and an interesting system, not a return.

So models route cheap by default, and the expensive ones are opt-in and metered. Nothing a model says carries investment authority. Its stated confidence is not a probability and is never stored as one.

The evals measure the process, because that is what can be measured here:

• the output has the shape it was asked for;
• every claim that says it has a source actually has one;
• the rules were obeyed, and each refusal was recorded;
• no figure was invented;
• a rerun gives the same answer.

Whether the advice was good is not measurable at this scale. Outcomes are recorded beside the same money left in a FTSE All-World (accumulating) tracker. Recorded, never claimed.

Type in what your broker shows: ticker, quantity, value, blended return. Cost basis is derived from those and the total is reconciled against the broker's figure, so a typo cannot become your cost basis. The file stays in your home directory, never in the repository.

# one profile, one snapshot, one database
uv run python main.py profile add adam --telegram-id 123456789 --operator
cp seed_snapshot.example.toml ~/Documents/cash-ash/profiles/adam/seed_snapshot.toml
uv run python main.py init --dry-run   # does it reconcile?
uv run python main.py init

# then, each week
uv run python main.py weekly --telegram

Every command, the weekly cycle and the evidence format are in the docs. What is not built yet is on the roadmap.