synthesis-matrix
Skill1 starsChanged 11 days agoApache-2.0View on GitHub
Turn a set of paper cards into a cross-paper matrix plus research-gap notes. Use when two or more cards already exist and the task is comparison — build the comparison table, group papers along an explicit functional axis with self-tests, write cross-topic interface edges including deliberately empty ones, and derive gap candidates. Input: cards conforming to paper-card schema v1, plus a grouping axis. Output: a matrix file with comparison tables, per-cell provenance notes and pending-adjudication entries, plus a gap-notes file. It consumes schema v1 and never redefines it. Contradictions are registered, not decided; it does not create cards, does not write proposals, and does not rule on whether a claim holds.
In this repository7 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- evidence-auditSkill
- literature-radarSkill
- paper-cardSkill
- proposal-forgeSkill
- research-opsSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 26
Before building the table, write:
- Distinctive: line 32
If you cannot state (3), the axis is a description, not a hypothesis, and the matrix will encode your taste.
- Distinctive: line 48
Before two numbers may share a column, the metric family, the statistic, the measurement engine, and the implementation vintage must all be …
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
synthesis-matrix
N cards become one comparison, on one explicit axis, with the incomparabilities written down rather than smoothed over.
The output is not "a table of results". It is a claim about how a field is organised, and that claim can be wrong in a way a table cannot: a bad grouping produces a comparison that is internally consistent and meaningless.
The one thing to get right first
The grouping axis is a hypothesis, and it must be testable. A grouping is not a set of labels; it is a claim that papers in one group do the same job in the same chain, and therefore that their numbers are comparable.
Before building the table, write:
- the axis, in one sentence;
- four predictions AX1–AX4 the axis makes — what should cluster, what should be empty, which group should be smallest, which two groups should share a baseline;
- a counter-example that would falsify the axis: a paper that fits the label but contradicts the predicted structure.
If you cannot state (3), the axis is a description, not a hypothesis, and the matrix will encode your taste.
→ references/grouping-axis.md
Structure
| Table | Holds | Rule |
| A — comparison | one row per paper, columns per attribute | every numeric cell carries a source path; incomparability is written in the cell, not omitted |
| B — relations | citation direction, shared baselines, conflicts | only groups with real relations get an edge; an empty edge is written as empty |
| C — gaps | what is missing, by category | a gap needs a category, a reason, and the evidence that it is a gap and not an oversight |
→ references/matrix-structure.md for column design, the A-note rules, and why an empty edge must be explicit.
Comparability
Before two numbers may share a column, the metric family, the statistic, the measurement engine, and the implementation vintage must all be compatible — or the difference must be larger than the noise band.
| Question | Verdict |
| same metric name, different family (AUPR vs ROC-AUC) | not comparable — different questions |
| same metric, different engines, band 7–9 points | comparable only if the gap exceeds the band |
| same metric, different runs of the same model, band 1–2 points | comparable only if the gap exceeds the band |
| a re-implementation of an old model on a new toolkit | not that model's performance — label it |
| a summary sentence vs the underlying grid | the grid wins |
→ references/comparability.md — the four bands, the metric-family table, and the reconciliation procedure.
Registering, not ruling
A matrix is a record-layer artefact. When two cards disagree:
- register it with an ID, a type, both locations, and both quoted values;
- suggest a disposition;
- do not decide which is true, and do not edit either card to make them agree.
The ruling, if one is needed, belongs to evidence-audit. The register mechanics — including why a registration is not a precedent — belong to research-ops.
Output contract
matrix/<topic>.md tables A/B/C + per-cell provenance + pending-adjudication list
gaps/<topic>.md gap candidates by category, each with its evidenceBoth files name the axis in their header, and both list the cards consumed with their schema version. A matrix whose input set is unknown cannot be updated, only rebuilt.
Routing
| Task | Go to |
| design or test the grouping axis | references/grouping-axis.md |
| column design, A-notes, empty edges | references/matrix-structure.md |
| may these two numbers sit in one column | references/comparability.md |
| turn the comparison into gap candidates | references/gap-notes.md |
| produce a card | paper-card |
| decide whether a claim holds | evidence-audit |
| register a contradiction in shared state | research-ops |
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.