Lineage
A row of related research moves.
Definition
A lineage is a chain or tree of related cards. It represents a research direction over time: what was tried, where it branched, what improved, what stalled, and whether the row deserves more compute. Lineage is the product answer to the problem of choosing the next card or switching to a new row.
How It Looks
A lineage looks like a row called representation-geometry. It contains cards for embedding probes, local density filters, disagreement weighting, and transfer tests, with scores and insights attached.
How To Use It
Use lineages to manage a portfolio of autonomous research directions. A swarm should not collapse into one popular method family just because agents locally hill-climb the same idea.
Research Portfolio
Anthropic's AAR results showed that assigning diverse research directions improved hill-climbing and helped prevent entropy collapse. Lineage makes that operational: each row has compute allocation, score frontier, novelty, complexity, transfer status, and open questions.
Switching Rows
Humans often add value by deciding that a current chain is saturated or misleading and that another row deserves attention. Picidae should surface stalled rows, suspicious rows, underexplored rows, and promising abandoned rows.
Entropy
Lineages make exploration measurable. Picidae can track how much compute is going to each method family, whether workers are duplicating one another, and whether new cards are actually lateral or just variations of the same row.
Show Examples
W2S portfolio
A W2S workspace can track separate rows for self-training, confidence weighting, unsupervised elicitation, representation geometry, curriculum, and evolutionary refinement.
lineages: - self-training - confidence-weighting - representation-geometry - curriculum - evolutionary-refinement entropy: active_families: 5 duplicate_rate: 0.18 recommendation: add compute to representation-geometry
Stop a row
A lineage can be stopped even if its latest run improved slightly, if the evidence shows rising complexity, weak transfer, and no clean mechanism.
lineage: stacked-label-filters status: stop reason: - score gains only on discovery split - complexity increased - transfer eval failed
Owns / Defines
Cards, branches, frontier score, direction label, entropy category, status, and strategic recommendation.
Questions Operators Should Answer
- What research direction does this lineage represent?
- Is the lineage improving, saturated, suspicious, duplicated, or underexplored?
- Which card is the current frontier, and which old card deserves a branch?
- How much compute has this row consumed compared with its evidence quality?
- Does this lineage transfer or only exploit one environment?