Concepts#

The vocabulary the rest of the user guide assumes: the research loop and its roles, how a trial builds on earlier ones, the factor sets a candidate is compared with, intraday families, the segments, and what makes a factor accepted.

Mechanism first#

A candidate is a hypothesis about an economic mechanism (why some stocks earn more over the next few days), a description of the factor the hypothesis implies, and code that computes it, written with quantlab’s KunQuant operators. Every candidate belongs to a mechanism type; the default types are behavioural_bias, risk_compensation, liquidity, microstructure and information_diffusion.

The agent does not search over formulas. It is asked for mechanisms, and the checks that follow refuse what is merely a variant of an existing factor.

The roles#

Role	Calls per round	Reads	Writes
planner	one	seed directions, factor families, memory, recent trials, last round’s feedback	one direction per slot, new or building on an earlier trial
proposer	one per proposal group, plus fixes	the directions (with the code and statistics of the trial one builds on), the data section, the vocabulary	candidates
reviewer	one per proposal group, plus one for rewrites	each candidate and the data section	a verdict per candidate
distiller	one every `round.memory_every` rounds	the trials since the last memory and the factors accepted so far	`memory.md`

Each role names its own backend and model, so the reviewer can be another provider’s model than the proposer’s. See LLM backends.

A round’s slots are allocated by mechanism type: each type gets round.min_per_type slots and the rest go by Thompson sampling on each type’s admission rate so far. The round’s directions are split into round.groups proposal groups, contiguous runs in mechanism-type order; each group is proposed, fixed and reviewed by calls of its own, and the groups’ calls of a stage are sent at once. The Gate still checks the whole round in slot order, so a duplicate written in another group is refused.

Each slot’s candidate becomes a trial when its round is written to the ledger, with its final status: accepted, rejected or failed. Until then it is a pending trial: what is known so far of the slot in the round still running, its research direction, its latest version and the Gate’s and the reviewer’s outcomes until now.

Building on earlier trials#

A research direction is either new or builds on an earlier evaluated trial, its parent, accepted or refused, as a refinement (the parent’s mechanism measured another way) or a mutation (another channel the parent’s result points to). The trials whose direction names a trial as their parent are its children, and its Lineage tells how they did: how many were tried, how many stopped before evaluation, which were accepted, and how far the best of them got against the parent itself. With context.lineage the planner and the proposer read it.

The planner chooses most parents. With round.picked_parents, the code chooses the parents of a few slots each round, the Picked parents. It judges every evaluated trial by its quality, its Lineage and its distance from the accepted factors, and lets each trial share those scores with the trials related to it, so a trial whose relatives did well rises with them. The trials and their relations are the Trial graph: the evaluated trials of the earlier rounds, joined where one was built on the other, where their code shares structure and where their factors rank the stocks alike. The trial_graph settings shape it. See memory and context.

Factor sets#

Set	What it is	A candidate must be
reference set	the factors the agent is trying to add to (`reference`)	not a rewrite of one, correlated at most `max_correlation`, incremental (partial IC)
known factors	every other existing factor (`known`)	not a rewrite of one, correlated at most `max_correlation` on a sample of dates
accepted factors	what this campaign accepted so far	like the reference set

The zoo holds the KunQuant graphs of all three; the Gate compares a candidate’s graph with them. The reference pool holds the scored values of the reference set and the accepted factors; the Evaluator measures a candidate’s correlation with it and regresses the candidate on it for the partial IC. Known factors never enter the pool.

A collision is a trial refused for being too like a factor of these sets: by the Gate as a rewrite, the same graph with other parameters or a too large shared subtree (a code collision), or by admission as a correlation above the limit (a ranking collision). A repeat of the campaign’s own earlier candidate is refused as a duplicate, not a collision.

Intraday families#

An intraday family is a family of fields that are values of each intraday bar (one bar width per campaign, a minute, say) in the regular session. A candidate computes on those bars itself before it reduces each session to one value.

A session is one trading day’s regular hours. A candidate’s intraday values for a day depend only on that day’s session and are known at its close.

The intraday stage is the part of a candidate computed on the intraday bars, each session on its own. It ends in a session aggregation, the step that reduces one session’s values to one value for the day, such as SessionMean or SessionStd; an aggregation is the only way from the intraday stage to the daily stage, which computes on one value per stock and day. Each aggregation’s value is a session column, which the daily stage reads as it reads a field. See intraday families.

Segments#

Segment	Used by	Shown to the LLM
training	the Evaluator, admission, the prompts’ statistics	yes, as statistics only
validation	`Acceptance.validate()`	no
hold-out	`Acceptance.open_holdout()`, once	no

Each segment drops the last bars whose label reads past its end (purge_bars, at least the label’s lookahead). The prompts carry no returns, dates or held-out metrics.

Direction#

A candidate states no direction. Its direction is the sign of its training IC, and every statistic is read in that direction: a factor with a negative IC is judged, and exported, as its negative. An exported factor always reads “a higher value, a higher return”.

Accepted#

A candidate is accepted when, in order:

its coverage (the mean daily share of the members with a value, measured before any fill) is at least admission.min_coverage;
its training IC is a nonzero number, so it has a direction;
its correlation with every reference, accepted and (sampled) known factor is at most admission.max_correlation in absolute value;
its partial IC’s Newey-West t clears a threshold that rises with the number of evaluated trials (a Bonferroni correction anchored at admission.t_threshold);
its decile returns rise with the decile: a Spearman correlation of at least admission.min_monotonicity.

The first check that fails is the reason recorded. See evaluation.