Prompts

The prompts describe the config’s data and research, not one market. They are built from market, the label’s horizon (its span_bars()), the scoring, the vocabulary and the mechanism types, through six templates you can replace.

market: adjusted daily bars of every US common stock    # required: the data, as the prompts say it
mechanisms:                                              # optional: the default is the five types below
  - {name: behavioural_bias, description: investors' systematic errors that prices correct later}
  - {name: liquidity, description: the cost and the pressure of trading}
prompts:                                                 # optional: templates replaced from files
  review: prompts/review.txt

Mechanism types

mechanisms are the types the planner allocates slots to and a candidate may belong to, each with a one-line description the planner reads, in plan order. round.min_per_type times their number must fit in round.candidates. Without mechanisms the types are:

Type	Description
behavioural_bias	investors’ systematic errors (over- and underreaction, anchoring, limited attention) that prices correct later
risk_compensation	a return paid for bearing a risk investors demand to be compensated for
liquidity	the cost and the pressure of trading: illiquidity premia, liquidity demand and its reversal
microstructure	the mechanics of trading within and between bars: order flow, the bid-ask bounce, the open and the close
information_diffusion	news that reaches prices gradually, across investors, related firms or time

Templates

Template	Role	Must use	May also use
data	the data section, the same every round	market, inputs, scoring, horizon
plan	planner	data, context, quota	round, mechanisms, actions (the actions a direction may take and how it names its parent), picked (the picked slots, empty without round.picked_parents; required with it)
propose	proposer	data, context, code_rules, vocabulary, fields, n, directions	round, mechanisms
retry	proposer’s fix request	code_rules, vocabulary, fields, n, failures	round, data, mechanisms
review	reviewer	candidates, data	round, mechanisms
distill	distiller	memory, history	round, data, mechanisms, accepted

Every template may use market and horizon. A template is a Python format string; each {placeholder} is filled when the prompt is made. When the config is read, each replaced template is checked for the placeholders its role needs and for any it may not use, and every problem is reported at once (PromptError). A replaced template is copied into the run’s snapshot like the field and operator files, so changing it needs a new run directory.

What the model is never shown

The prompts carry training-segment statistics only: no return, no date, no validation or hold-out metric. Papers in seed directions are cited without their year, so no prompt names a date after the training segment.