From agentic-research
Read, update, and promote claim statuses through the controlled vocabulary. Includes the project.yaml patch protocol and the conjectures.md update procedure.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agentic-research:claim-lifecycleThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Claims move through a controlled status chain. This skill governs all reads
Claims move through a controlled status chain. This skill governs all reads
and writes to conjectures/conjectures.md and the claims block in
project.yaml. No agent should touch claim statuses without following this
protocol.
question → hypothesis → proof-sketch → partial-proof → proved
→ refuted
proved → partial-proof) is allowed without user
approval when new evidence invalidates a prior promotion.hypothesis → refuted, proof-sketch → refuted) are
allowed with an explicit counterexample.question is the starting status for something not yet formalised.simulation-supported is a separate annotation that can co-exist with
any status (it records empirical corroboration, not proof strength).| Target status | Required evidence |
|---|---|
hypothesis | A gap brief exists at notes/gap_briefs/<slug>.md |
proof-sketch | A dated attempt in proofs/<claim-id>.md with a clear strategy and at least one justified step |
partial-proof | Proof attempt has a complete argument with explicitly labelled gaps (**G1.**, …) |
proved | (1) Zero gaps in the proof file, (2) an adversarial-proof-review pass recorded, (3) at least one small-case verification, (4) user approval surfaced and confirmed in chat |
refuted | An explicit, checkable counterexample written into the proof file |
proved always requires user approval — the Proof agent reports the
promotion request in the run summary and the Supervisor surfaces it. The
Supervisor patches project.yaml only after the user confirms in chat.
conjectures/conjectures.mdThe file uses YAML-frontmatter blocks, one per claim:
## C<n>: <one-line statement>
- **Status:** <status>
- **Added:** <ISO date>
- **Last updated:** <ISO date>
- **Linked proof:** `proofs/C<n>.md` (or `—` if none)
- **Linked gap brief:** `notes/gap_briefs/<slug>.md` (or `—`)
- **Notes:** <free text — why promoted, what changed>
Update rules:
Status, Last updated, and Notes fields.## C<n>: … line).Status: refuted.project.yamlThe claims block looks like:
claims:
C1:
status: proved
proof_file: proofs/C1.md
C2:
status: hypothesis
proof_file: ~
Only the Supervisor patches project.yaml. Other agents report
their desired status change in the run summary. The Supervisor then
applies the patch after validation (and user approval for proved).
When patching:
status and proof_file.current_model_version only with explicit user approval.mode (discovery ↔ project) without explicit user
approval.Before dispatching any proof or writer pass, the Supervisor should report:
Claims summary
──────────────
proved : N
partial-proof : N
proof-sketch : N
hypothesis : N
question : N
refuted : N
─────────────────
open (attackable): N
"Open" means status ∈ {question, hypothesis, proof-sketch,
partial-proof}.
npx claudepluginhub ramanebrahimi/raman-marketplace --plugin agentic-researchGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.