| name | lab |
| description | Open an interactive exploration session over the research corpus — traverse MIF relationships from a seed concept or entity, develop and stress-test hypotheses against the finding set, and reason across topics. Use this skill whenever the user wants to discuss, explore, or riff on their research rather than just retrieve it: 'explore topic X', 'what does my research suggest about Y', 'develop a hypothesis about Z', 'interrogate the findings on W', 'think through this with my research', 'brainstorm from the corpus'. |
| version | 0.4.2 |
| argument-hint | [--seed <urn:mif:id>] [--hypothesis "<claim>"] [--interrogate] [--namespace <ns>] |
| allowed-tools | Read, Bash, Grep, Glob, Skill |
lab — explore and reason over the corpus
A thinking partner grounded in the MIF substrate. Where /search retrieves and
/discover audits, this skill reasons: it walks the MIF-native graph from a
seed, grounds a conversation in findings, and helps form and test hypotheses.
It stays inside the corpus: it never invents findings, and it treats every
finding as a pointer to a primary source, not as ground truth. When a claim
matters, recommend re-reading the cited source.
Data
A finding is a MIF concept at reports/<topic>/findings/*.json: @id, title,
content, tags, entities[], typed relationships[], citations[], and
extensions.harness (dimension, verification verdict). The graph
(knowledge-graph.json) is the navigable index of those concepts and their
edges. Use the sample at reports/_meta/sample-session/ when exploring.
Modes
Default is open exploration. The flags are postures, not silos.
--seed <id> — start from a concept/entity and traverse its relationships--hypothesis "<claim>" — develop and stress-test a falsifiable claim--interrogate — invert the burden: hunt the corpus's weak points--namespace <ns> — scope reasoning to one topic
Traversal from a seed
Walk outward from a seed urn:mif: id along the graph's typed edges — one hop
gives the seed's direct neighborhood (supports, contradicts, derived-from,
mentions):
jq --arg seed "SEED_ID" '
{ seed: $seed,
out: [ .edges[] | select(.source == $seed) | { type, target, strength, via } ],
in: [ .edges[] | select(.target == $seed) | { type, source: .source, via } ] }' \
reports/_meta/sample-session/knowledge-graph.json
Follow supports/derived-from chains to build the case for an idea, and
contradicts edges to find the tension. Resolve each target id back to its
finding (title, content, citations) before discussing it.
Hypothesis mode
Run the pipeline: Orient → Surface → Form → Corroborate → Disconfirm → Weigh
→ Gaps. The non-negotiable step is Disconfirm: before any verdict, check
for contradicts edges, findings whose extensions.harness.verification.verdict
is weakened/falsified/quarantined, and low-trust provenance. A hypothesis
never tested against the disconfirming evidence is incomplete.
Verdict is exactly one of: supported | contested | unsupported |
untestable-with-corpus. Cite every supporting finding by @id and its source
citations[].url. Quarantined or falsified material may appear only as
disconfirming evidence, never as support.
Interrogation mode
The whole job is finding what is weak. Surface, in order: findings with a
non-survived verdict (name each by id), contradicts edges, low-confidence
provenance (provenance.confidence), citations that are not externally
resolvable, and aging created dates. Produce a ledger of supporting vs
disconfirming findings; weak/falsified ids go only in the disconfirming column.
Evidence discipline
- Findings are pointers, not authorities — cite the
@id and the source URL.
- Disconfirming material stays disconfirming — a falsified or quarantined
finding is never cited as support.
- Non-
survived verdicts are tripwires — surface them whenever an affected
finding enters the conversation.
- Say "the corpus doesn't cover this" when it doesn't — gaps are findings
too; hand them to
/discover --gaps or /start --augment.
Handoffs
| Conversation surfaces | Hand off to |
|---|
| A pure retrieval need | /search <query> |
| A coverage gap or thin dimension | /discover --gaps or /start --augment <dimension> --topic <topic> |
| A relationship/visualization question | /graph |
| A claim needing adversarial web verification | /falsify |
| An idea worth becoming new research | /start |
Prerequisites
- The corpus findings dir,
research-index.json, and knowledge-graph.json.
If the graph is missing: bash scripts/build-graph.sh <findings-dir>.
Examples
/lab --namespace harness/example-topic
/lab --seed urn:mif:concept:harness:kg-copier-0001
/lab --hypothesis "A living template beats a one-time snapshot"
/lab --interrogate
