When I said “Librarian,” I thought the job was obvious: go through books, papers, websites, YouTube transcripts, and trader claims; find actual strategies; save them; and hand the good ones to a backtester. Instead, I spent a lot of time building a source grader that was very good at telling me why things were not ready.
That was useful work, but it was not the job I needed. The fix was to make the handoff dead simple: find the idea, turn it into exact rules, and test it.
First, save what the person actually said
The Librarian still records the source, date, exact claim, and whether we are allowed to use it. That matters because “this guy said RSI works” is useless six months later if nobody remembers which settings, coin, chart, or market he was talking about.
A book, YouTuber, Reddit post, or paid indicator can give us an idea. It does not get a free pass. It also cannot tell the agent to ignore our rules just because somebody hid instructions in a web page.
Then turn the idea into rules
A tester cannot backtest “buy when it looks strong.” It needs the coin, chart timeframe, indicators, exact entry, exact exit, stop, size, and trading costs. If the source leaves a blank, we mark the blank instead of quietly making up an answer.
If two people can read the rules and build two different trades, the rules are not finished.
Any change makes a new version. Otherwise it is too easy to tweak a loser, keep the old good-looking numbers, and bullshit ourselves.
Then try to kill it
The backtester runs the rules through old market data in time order. It includes fees, spread, slippage, and delayed fills. A separate checker looks for cheating by accident: future data, survivor-only coins, overlapping samples, magic settings, or testing the same history until something finally looks good.
The answer is not always winner or loser. Sometimes it is “we do not have enough trades,” “we are missing the right data,” or “the creator never explained the exit.” Those go in the uncertain pile so we can try again later without pretending they passed.
The handoff has to survive without a meeting
Every step saves a record instead of tossing a paragraph to the next bot. I should be able to click a strategy and see where it came from, what rules were tested, which data was used, and why it won, lost, or got stuck.
This does not guarantee a winning strategy. It does make bad ideas cheaper to kill and good-looking results harder to fake.