Skip to main content
Paper-exploration is the generate skill at the portfolio level — the paper-level mirror of hypothesis-exploration one level down (see ../../docs/design/00-meta-spec.md §3.1 and ../hypothesis-exploration/SKILL.md). Where hypothesis-exploration proposes hypotheses into a paper’s backlog.md, paper-exploration proposes candidate papers — application spin-offs and follow-ups — into the portfolio’s docs/research/portfolio-backlog.md. A paper starts as a one-line pitch here and, once the human promotes it, is handed to paper-synthesis (the resolve skill) which drives it through pitch → positioning → outline/plan → decision → sections (ADR-0006, ../../decisions/0006-two-skills-per-level.md). The portfolio is the second of three nested loops: hypotheses accumulate into a paper; papers accumulate into a portfolio; the portfolio (optionally) covers a thesis’s aims. The portfolio sits under the optional thesis level (see ../thesis/SKILL.md) — in a repo that is not a thesis, the portfolio is the top. This skill proposes only; the human decides what gets written (agency principle, ../../docs/design/00-meta-spec.md §2.1).

​
When to use

  • A paper’s hypotheses are resolving and you want to see what next papers the evidence, the mechanism, or the leftover limitations suggest.
  • The portfolio backlog is empty, thin, or stale and needs a generation pass.
  • A reviewer comment, a rival’s result, or a new application domain lands and you want it parked as a candidate before it is lost.
  • You want the backlog re-ranked by feasibility × interest, or a candidate promoted to paper-synthesis (or dropped).
  • You are registering a newly promoted paper into docs/research/papers.md.
Do not use this skill to develop a paper (that is paper-synthesis), to decide whether to publish (that is the decision.md gate inside paper-synthesis), or to make any commitment on the author’s behalf.

​
Verbs

park, generate, rank, and drop never commit the author to writing a paper. Only promote moves a candidate out of exploration, and only on an explicit human pick — the exploration/resolution firewall (../../docs/design/00-meta-spec.md §2.3).
Tooling. Row append/transition is the shared honest-scholar backlog park|add|list|rank|promote|drop CLI group (honest_scholar/exploration/backlog.py, shared with hypothesis-exploration) — ensure via ensure-tooling, operating on portfolio-backlog.md at the paper level. 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 portfolio-backlog.md table directly, keeping the column order.

​
Generation lenses

Three lenses turn finished or in-flight work into candidate papers. Each emitted candidate names the lens it came from, so the backlog is auditable.
  • Mechanism-transfer — take a mechanism that worked here and ask where else it applies. “The construction that gave monotonicity in domain A — does it transfer to domain B?” Application/spin-off papers. Cross-check novelty with literature scout --level paper before parking (someone may already have transferred it).
  • Limitation-driven — read the limitations / threats-to-validity / future work of a resolved paper (and its refuted or inconclusive hypotheses) and turn each into a follow-up that removes the limitation. Refuted load-bearing hypotheses are especially fertile: the negative result often is the next paper.
  • Result-driven — take a confirmed result and ask what new question it now makes askable — a scaling study, an ablation-as-its-own-paper, a head-to-head against a rival the result now beats, a new dataset the result unlocks.
Shaping heuristics (applied here, not by scout): gap-spotting vs. problematization (Sandberg & Alvesson 2011 — prefer candidates that challenge an assumption over those that merely fill a hole) and feasibility × interest (Alon 2009). See ../../resources/references/citation-scouting.md §5. Prioritization — feasibility × interest. Score every candidate on two axes and rank by their product:
  • Feasibility — do we have (or can cheaply get) the data, backend, and compute? Does an existing bound backend already cover it? How many unresolved hypotheses stand between here and a submittable claim?
  • Interest — venue fit, novelty vs. prior work, size of the audience, and whether it advances a thesis aim (if the portfolio sits under a thesis).
Neither axis is a productivity score; the product orders proposals the human triages. A high-interest / low-feasibility candidate is parked, not discarded — feasibility often changes.

​
Seeding from literature scout

generate may call the literature skill in scout mode at --level paper (recall-oriented: large citing set, skim metadata, co-citation clustering + research-front / burst detection dominate — see ../../docs/design/02-literature.md §2 and ../../resources/references/citation-scouting.md §4). Scout returns ranked idea rows, each carrying mandatory provenance (idea | type | source-paper id | citing-context snippet | why-it-matters | est. feasibility). A scout-seeded candidate keeps that provenance line so every parked paper traces back to who-cited-what. Scout proposes leads; this skill folds them through the lenses and the human triages — scout adjudicates nothing.

​
The papers registry

Paper-exploration owns docs/research/papers.md — the portfolio’s index of paper roots and their experiment-backend binding.
  • What it holds. One row per paper: paper-id → its root path under docs/research/<paper>/ + a backend: field naming the bound experiment backend (required — the plugin bundles no default; each repo supplies its own implementation of the contract; ADR-0013, ../../decisions/0013-experiment-backend-contract.md; ../../docs/design/04-substrate-and-contract.md §3.1). Pipeline skills resolve the backend from this binding and contain no backend-specific logic, so backend: is set here at registration time.
  • When it is written. promote adds the registry row and scaffolds the paper root (hypotheses/, backlog.md, paper/). A candidate in portfolio-backlog.md is not in papers.md — the backlog is proposals; the registry is committed papers.
  • paper-id. A stable, human-readable slug (kebab-case), assigned at promotion and never reused; it keys the paper across the backlog, the registry, the dashboard, and progress.

​
Composition

  • Down-mirror: hypothesis-exploration (../hypothesis-exploration/SKILL.md) is the same shape one level down — same verbs, same firewall, same proposes-only stance. Reading either explains the other.
  • Hand-off: promote → paper-synthesis (../paper-synthesis/SKILL.md), which develops the promoted candidate to a publish/no-go decision.md.
  • Seeding: literature scout --level paper (../literature/SKILL.md) feeds candidate rows; defend’s cited-work target later audits their citations.
  • Up-mirror: the portfolio feeds the optional thesis level (../thesis/SKILL.md); when a thesis exists, interest scoring weighs coverage of a thesis aim.
  • Reporting: progress (../progress/SKILL.md) reads the papers.md registry and each artifact’s status frontmatter — not the portfolio-backlog.md table; it never re-ranks or scores.

​
Guardrails

Hard rules — load-bearing, not preferences.
  • Proposes only; the human promotes. Every verb but promote stays inside exploration. promote requires an explicit human pick and is the only path out. The skill never decides which paper to write, and never touches a decision.md publish verdict — that gate lives in paper-synthesis and is a material decision with a named human sign-off (../../docs/design/00-meta-spec.md §2.1, §2.3).
  • Firewall. Exploration proposes; resolution (paper-synthesis) disposes; synthesis reports. This skill sits wholly on the propose side.
  • File-drawer discipline. drop records a reason; candidates are retired, not silently deleted, so the exploration history stays auditable (rigor kit, ../../docs/design/01-lifecycle.md §3).
  • Provenance is mandatory. A scout-seeded candidate keeps its citing-context snippet + source-paper id. A hand-parked candidate records its origin (which result / limitation / mechanism it came from).
  • No scores. Feasibility × interest orders proposals; it is not a productivity metric and never rolls up into one (anti-Goodhart, ../../docs/design/00-meta-spec.md §3.6).
  • Domain-neutral. No monotonic-network or ML assumptions live here; the lenses and the registry are generic. Domain content comes from the consuming repo.

​
Commit attribution

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