User reference / Read the reports
Influence matrix
Read calculated matches, historical agent declarations and analyst interpretations in the same matrix.
Recorded: executor 0.1.5 · analyst 0.1.3 · report 0.1.16
Captured 2026-09-21 · Capture provenanceExplore this view ↗
Choose a refinement round, then Influences. A row is a current-round recipient. A column is a previous-round possible source. The diagonal compares an agent slot with its own earlier version. Headers link to the underlying proposals. Round 0 has no previous-round influence matrix.
| Element | What it counts |
|---|---|
Round n ↓ / round n−1 →Reviewed E 0.1.5 · A 0.1.4 · R 0.1.16 |
Explicitly identifies the recipient round and source round. |
| Proposal / agent row labelReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | The new proposal whose steps or ideas are being examined. |
| Proposal / agent column labelReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | An earlier proposal the refiner had available to read. |
| KeptReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Current steps attributed to the same agent's preceding proposal by title matching. |
| Same titlesReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Current steps assigned to this peer's preceding proposal by the title-origin heuristic. It is an exclusive attribution count, not every pair of titles above the threshold. |
| New stepsReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Current steps with no qualifying assigned previous-round source. A renamed idea can count as new. |
Agent says +a / −rReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 |
Count of adopted/rejected idea-list entries in legacy notes written by the recipient about that source. Only off-diagonal declarations are shown. |
Analyst sees +a / −rReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 |
Count of structured taken/rejected idea entries the saved analyst attributed to this source. Only off-diagonal annotations are shown. |
Within a row, the title counts across source columns plus new steps equal the proposal's step count. The idea counts do not have to add up to that number: an idea can span several steps, and several ideas can live in one step. Do not add title matches, declared ideas and analyst ideas together.
Recorded: executor 0.1.5 · analyst 0.1.3 · report 0.1.16
Captured 2026-09-21 · Capture provenanceExplore this view ↗
| Popover | Detail shown |
|---|---|
| Same titles / KeptReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Explanation of title-only matching and a list of current step numbers/titles paired with source step numbers/titles. |
| New stepsReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Which current steps have no title source, with the caveat that rewritten or combined ideas may be missed. |
| Agent saysReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Who wrote the note, which earlier proposal it concerns, the verdict and adopted/rejected lists. The author had read the proposal, but this is still its own account. |
| Analyst seesReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Analyst provider/model, method explanation, taken/rejected ideas, source-step references, and what / why evidence. It was produced after execution. |
| Zero / missing lineReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Zero means the available source supplied no counted entries. An absent analyst/declaration line means that evidence source is unavailable, not a measured zero. |
Hover or focus to read a number's explanation, click to pin it, drag the pinned header to move it, and close with ×, outside click or Escape. The interactive example's long explanation can scroll within its popover.
Titles are lowercased and compared with SequenceMatcher. A candidate needs similarity at least 0.6. A peer wins only if its match qualifies and the own-version match either does not qualify or is beaten by more than 0.1. Otherwise a qualifying own match wins. Each current step gets one source or “new”; the method can reuse a previous step for multiple current steps.
For a worked numerical example and tie/order behavior, see Attribution methods. This is a deterministic textual heuristic, not conceptual lineage or a measurement of causal influence.
In the captured round-2 example, GPT's row has 0 title-attributed steps from Opus, while the analyst identifies 5 taken ideas and 3 rejected ideas from Opus. This is not an arithmetic inconsistency. It directs you to read the cited ideas: the step wording can change while a concept survives.
The analyst also receives the title matches as evidence. Its interpretation is not an independent ground-truth label. Use the pair of views to inspect a hypothesis about influence, then check the full proposals.
Source files used for this reference
Reviewed at 7361ea0b71a5.