Skip to main content
Hypothesis-exploration is the generate skill at the hypothesis level — the first flywheel of the three-level mirror (see ../../docs/design/01-lifecycle.md §1). It runs an origin-agnostic idea pipeline: raw hunches, citation-seeded leads, and data-driven observations all land as rows in the paper’s backlog.md, get ranked, and are either promoted into a hypothesis folder (handed to hypothesis-testing) or dropped with a recorded reason. It sits on the explore side of the firewall (../../docs/design/00-meta-spec.md §2.3): exploration proposes, resolution disposes, synthesis reports. This skill only proposes. It never tests a hypothesis, never assigns a verdict, and never decides what the author pursues. The human promotes; the skill advises (agency principle, §2.1). Grounding: ../../resources/references/citation-scouting.md.

​
When to use

  • Park a raw idea mid-flow so it is not lost, without committing to work it.
  • Generate candidate hypotheses for a paper: mine the citation graph via literature scout, run exploratory data analysis, or apply the generation moves below.
  • Rank a backlog to decide what is worth a testing slot.
  • Promote a ranked item into a hypothesis folder to begin hypothesis-testing.
  • Drop an item that will not be pursued, recording why (file-drawer discipline).
Do not use this skill to test, confirm, refute, or score a hypothesis — that is hypothesis-testing’s job, gated by a human sign-off (findings.md verdict).

​
Verbs

The pipeline is a small state machine over backlog.md rows: parked → candidate → ranked → promoted | dropped. promote and drop are terminal but never delete a row — dropped ideas stay on record (see Guardrails: file-drawer discipline). park is deliberately frictionless; generate and rank do the analytic work.

​
Backlog row schema

Every row (a line in the backlog.md table under docs/research/<paper>/) carries provenance so any candidate can be traced to where it came from:
  • provenance — the origin, mandatory:
    • scouted → source-paper-id + a citing-context snippet (verbatim, from literature scout; the auditable link back to who-cited-what);
    • EDA → eda:<dataset-id> + the observed pattern;
    • own → own (a parked hunch or a generation-move idea).
  • frame — gap-spotting or problematization (see Ranking).
  • note — for a dropped row, the recorded drop reason.
Tooling. Row append/transition is the honest-scholar backlog park|add|list|rank|promote|drop CLI group (honest_scholar/exploration/backlog.py) — ensure via ensure-tooling; shared with paper-exploration. add realizes the generate verb’s row-append; list is a read-only inspection command. By hand (if the CLI isn’t available): edit the backlog.md table directly, keeping the column order above so the backlog verbs (rank, promote, list) can parse it.

​
Generation moves

generate produces candidates three ways; all land as candidate rows with provenance.
  1. Scout-seeded — call literature scout --level hypothesis (see ../../docs/design/02-literature.md §2). It returns ranked leads, each with type | source-paper-id | citing-context snippet | why-it-matters | est. feasibility. Copy that provenance verbatim into the row; do not paraphrase the snippet. Scout classifies each lead into an idea type — extension, contradiction, new-domain, transferable-technique, methodological-gap — which maps onto the moves below.
  2. EDA — surface a pattern, anomaly, or unexpected correlation in available data and phrase it as a falsifiable claim. Record the dataset id and the observation as provenance. (Exploratory by construction — the eventual test must be tagged confirmatory-vs-exploratory in strategy.md; this skill does not run it.)
  3. Generation moves — author-side lenses for turning a starting point into a claim:
Each move yields one falsifiable sentence, not a topic. If it cannot be written as “X will/does Y under Z,” it is a theme, not a hypothesis — keep it parked until it sharpens.

​
Ranking

rank scores each candidate on three axes and attaches a framing flag. Scores are advisory inputs to the human’s ordering, not an automatic gate.
  • Expected information gain (EIG) — how much a resolved verdict would move belief regardless of direction. High EIG discriminates between live rival explanations; a foregone-conclusion test is low-EIG even if easy. This is the primary ranking axis (the flywheel selects for belief-moving tests — ../../docs/design/01-lifecycle.md §1).
  • Feasibility × Interest (Alon 2009, ../../resources/references/citation-scouting.md §5) — feasibility: can it be tested with the data, backends, and time on hand; interest: does the answer matter to the paper’s thesis or the field. Score each; the product guards against both infeasible-but-fascinating and trivially-easy-but-pointless.
  • Framing flag — gap-spotting vs. problematization (Sandberg & Alvesson 2011, same digest §5). Flag each candidate:
    • gap-spotting — fills a hole others left open (incremental; fine, but common);
    • problematization — challenges an assumption the field takes for granted.
    The flag is not disqualifying — it is surfaced so the author sees the mix. When interest is comparable, problematization typically carries higher EIG. Never auto-rank problematization above gap-spotting; present the flag and let the human weigh it.
Output is an ordered backlog.md (rows set to ranked) plus a one-line rationale per top candidate. The skill recommends; it does not select the testing slate.

​
Composition

  • Upstream: literature scout (../../docs/design/02-literature.md) feeds citation-seeded candidates; EDA and the generation moves feed the rest. All origins converge on the one backlog.md.
  • Downstream — promote: scaffold docs/research/<paper>/hypotheses/<YYYY-MM-DD-slug>/ and write hypothesis.md (the free-form claim — the first staged doc of hypothesis-testing, ../../docs/design/01-lifecycle.md §2.2), carrying the backlog provenance forward. Set the backlog row to promoted with a link to the folder. Control then passes to hypothesis-testing, which builds strategy.md → design/plan → findings.md.
  • Understanding. A promoted hypothesis may be examined at the strategy stage of testing (defend on strategy: assumptions, entailments, falsifiers, rival explanations — ../../skills/defend/SKILL.md). Exploration does not examine; it hands off a claim clean enough to be defended later.
  • Progress. Only promoted items become artifacts (hypothesis.md + downstream), and those carry the status frontmatter progress rolls up — the backlog.md table itself is not projected into the dashboard (progress reads artifact frontmatter, not the backlog). Backlog depth stays visible in backlog.md directly; a large backlog or many dropped rows is not “bad” and is never scored (../../docs/design/00-meta-spec.md §3.6).

​
Guardrails

  • Proposes only — the firewall. This skill generates and ranks; it never tests, never assigns a verdict, never decides what is pursued. Promotion is a human act (../../docs/design/00-meta-spec.md §2.1, §2.3; ADR-0006 ../../decisions/0006-two-skills-per-level.md).
  • Provenance is mandatory. No candidate row without an origin. Scouted rows must carry the source-paper id and the verbatim citing-context snippet; fabricating or paraphrasing a snippet is a firewall violation (automation-bias / citation-fabrication risk, §2.1).
  • File-drawer discipline. drop records a reason and retires the row — it never deletes it. Dropped and parked ideas remain visible so the selection history is auditable and the file drawer cannot silently swallow inconvenient ideas.
  • A hypothesis is a falsifiable claim, not a topic. Refuse to promote a row that cannot be stated as a testable sentence; keep it parked.
  • Ranking advises, never selects. EIG / feasibility×interest / the framing flag are inputs the human weighs. Do not auto-promote the top row and do not rank problematization above gap-spotting by fiat.
  • No premature testing talk. Do not sketch experiments, pick baselines, or estimate effects here — that is strategy.md/design in hypothesis-testing. Exploration stops at the claim.

​
Commit attribution

When you commit artifacts produced by this skill, add these git trailers — discovery + provenance (see ../../resources/commit-attribution.md):