empirical-probe

slug	empirical-probe
name	Empirical Probe
description	Survey what's already out there (datasets, libraries, prior art, instruments) and probe specific empirical questions to characterize the "material" being worked on. Output is the materials TM plus a curated source list pushed to the Librarian.
version	0.1.0
metadata	{"sources":[{"mode":"port-original","author":"Jannes Stubbemann","added_in":"0.1.0"}]}

Empirical Probe

Port-original skill. Hand-authored for this Agent Company; not from upstream. Owned by the Materials / Empiricist role.

When to fire

A thread needs to know what's already out there before committing to a line of work. Typical triggers: an active TM cites a gap ("no known dataset for X"), a Theorist proposes an abstraction that depends on empirical properties no one has characterized, or a Director instigation asks "what instruments exist for measuring Y?" Fire whenever the question is primarily about what exists rather than about a new hypothesis to test. Do not fire this skill to confirm a theory — that is the Experimentalist's work.

Inputs

• The question, stated as a survey scope: what kind of material is being probed (datasets, libraries, prior art, instruments, primary sources) and what the probing is supposed to resolve.
• The scope boundary: how broad. A narrow scope (one library ecosystem) and a broad scope (a field's prior-art landscape) require very different survey strategies; name which before starting.
• Any prior TMs or Hallway entries that motivated the probe, so the survey does not re-characterize what the lab already knows.

Outputs

Two artifacts:

1. Materials TM at memoranda/TM-NNNN-materials-<slug>.md — synthesizes the findings: what exists, what its relevant properties are, where the gaps are. Must cite ≥3 sources from the source list.
2. Source list at sources/YYYY-MM-DD-<slug>.md — a structured catalog of every source surveyed, with one-line probe values. Filed under sources/; the Librarian's library-push will index it.

Procedure

1. Survey systematically against the declared scope. For datasets: enumerate known ones and note their sizes, licenses, and domains. For libraries: list maintained options and note their interfaces and limitations. For prior art: identify the primary sources, not the survey papers that cite them. For instruments: name them and note their resolution, cost, and availability. "Systematic" means working from an explicit enumeration strategy, not from recall.

2. For each source, note its specific probe value before adding it to the list. The probe value is a one-sentence answer to: "What question does this source answer that the thread needs answered?" A source without a probe value is dead weight — it signals the Materials agent found the source but does not know why the lab should care. Do not add a source until you can write its probe value.

3. Write the Materials TM using the technical-memorandum skill. The Method section describes the survey strategy (what enumeration approach was used, what was in and out of scope). The Result section characterizes the material: what exists, what properties it has, what gaps remain. The Open questions section names the gaps that the Experimentalist or Theorist should take up next. The TM must cite ≥3 sources from the source list in Prior TMs cited.

4. Write the source list as a structured catalog. Format: one entry per source, with source name, link or citation, and the probe value (≤1 sentence). Entries are not prose paragraphs — they are rows in a catalog. No source appears more than once.

5. File the source list under sources/ and post a Hallway entry naming the Materials TM and the source list filename. The Librarian's library-push will pick up the Hallway entry and index the source list for other researchers. Do not push directly to other researchers' queues — let the Librarian route it.

Invariants

• Every source-list entry has a one-line probe value. A source without a probe value must not appear in the list. This is enforced by make check. Links without probe values are deleted, not left as placeholders.

• The Materials TM cites ≥3 sources from the list. A TM that synthesizes fewer than 3 sources is not a synthesis — it is a one-source summary or an argument from authority. If the probe turns up fewer than 3 usable sources, the TM's Result section must name this explicitly as a gap finding, and the ≥3 rule is suspended in that case with a documented explanation.

• The source list cannot include sources the Materials agent did not actually access. A citation added from a bibliography without reading the source is not a probe — it is a guess about the source's content. The probe value must be grounded in what the source actually contains.

• Each source appears once, with its primary probe value. A source that answers two questions gets its most load-bearing probe value. Duplicate entries obscure the catalog's structure.

Anti-patterns

• "Paste-bin of links with no probe value." Links without probe values are dead weight. The Materials TM's value is in characterizing the material, not in proving the survey was wide. A catalog of 30 links with no probe values is less useful than a catalog of 8 with precise probe values.

• "Synthesize from memory without grounding in sources." The Materials TM must cite ≥3 sources from the source list. An agent that writes the TM from prior knowledge and then constructs the source list to match it has inverted the procedure — and produced a TM whose conclusions cannot be audited.

• "Duplicate citations across rows." Each source appears once. Splitting a source into multiple rows (e.g., "this paper for claim A" and "same paper for claim B") inflates the apparent breadth of the survey. Write one row; put the primary probe value in it.

• "Include sources the agent did not actually read." The probe value requirement enforces this: you cannot write an accurate probe value for a source you have not accessed. A fabricated probe value is worse than a missing one — it actively misleads downstream researchers who rely on the catalog.

• "Survey until the list is long, rather than until the scope is covered." The stopping criterion is coverage of the declared scope, not a target source count. A survey that keeps going past its declared scope wastes cycles and dilutes the catalog with tangential material.