wave-research
Staged, adversarial research that tries to break a thesis before believing it: waves, checkpoints, falsifiable sub-claims, rival hypotheses, an ACH matrix, source-incentive tags and a verified final thesis. Also carries Workflow templates for a one-pass research-and-verify fan-out. Use to stress-test or vet a contested claim; not for quick lookups.
- License 许可
- MIT
- 最近更新
- 2026-11-12
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- wave-research
- license
- MIT
- description
- Staged, adversarial research that tries to break a thesis before believing it: waves, checkpoints, falsifiable sub-claims, rival hypotheses, an ACH matrix, source-incentive tags and a verified final thesis. Also carries Workflow templates for a one-pass research-and-verify fan-out. Use to stress-test or vet a contested claim; not for quick lookups.
- compatibility
- PORTABILITY: the staged method is pure prose and files on disk — no hooks, no scripts. The optional one-pass templates (templates/*.workflow.js) need Claude Code's Workflow tool; nothing else depends on them. Workers require actual dispatch tools for independent assignments. Without those tools, direct collection is possible but simulated role-play is not independent review; disclose the missing capability.
02 资源文件 6
03 SKILL.md 正文
Wave research
Turn a fuzzy thesis into a falsifiable one, attack it from several sides, and report what survives. Everything lives on disk so any agent, or you after context loss, can resume from files.
Invariant
A sub-claim is reported supported only when its disconfirmers were written before the search, at least one adversarial pass failed to refute it, two or more source-incentive classes agree, and at least one source is primary. Anything less is reported unverified, mixed or unknown — NEVER rounded up. The final thesis is checked by an agent that did not write it and sees the evidence table, not the narrative.
Not this
- A quick lookup, a single-source answer or a survey. Search and answer; the ceremony below costs more than the answer is worth.
- A one-pass research-and-verify fan-out (compare options, audit, explore a design space) that needs fact-checking but not pre-registered disconfirmers, rivals or checkpoints. Skip the waves and run a Workflow template; see Fan-out workers below.
- verify-before-claim (
plugins/verify-before-claim) backs one claim of fact or completion with the check that would prove it false, in the moment. This skill runs a multi-wave investigation of a contested thesis.
When to use this
The question is contested, decision-relevant, would take over an hour by hand, and a wrong confident answer is worse than a slow one. Phrases like "stress-test this idea", "vet this claim", "research this properly", "try to disprove it", or a multi-agent deep dive with rival hypotheses.
Use the host-adopted Core contract for
shared authority, interaction, task state and evidence integrity. Use
delegated-intelligence
for actual dispatch, briefs, budgets and recovery, and
stop-rule for retry bounds.
Their procedures are not redefined here. If these capabilities are unavailable,
report the missing dependency; do direct research within the accepted scope,
without claiming dispatched workers or independent verification. Optional
second-opinion remains offer-only under Core; human-voice can shape prose.
Defaults
- Establish the accepted thesis, research budget and reporting preferences before search. Use existing authorization; changes follow Core rather than a separate research autonomy tier or spending override. Preserve configured providers.
- Waves: at most 3 after W0. Recency window: 24 months for empirical claims; the history workstream is exempt.
- Output:
run_dir(default./wave-run-<slug>/) inside a private, durable repo. - Research checkpoints below own branch selection and evidence rigor. They do not grant publication, scheduling or expanded action authority.
Artifacts (under run_dir)
plan/plan.v{n}.md and plan/CHANGELOG.md; kb/W{w}-{name}.md notes;
tree.md knowledge tree (claim, evidence, leads, branches; status
open|supported|refuted|mixed|unknown); leads.md backlog; matrix.md;
thesis/thesis.v{n}.md; checkpoints/CP{n}.md; budget.md ledger;
lessons.md (bad query patterns, content farms, failed tools; every agent
reads it). Commit after every landing and checkpoint.
P1. Plan v0 (plan/plan.v0.md)
Restated thesis; questions; workstreams, each with research role, tool allowlist, search cap; and the rules every brief carries:
- Cite URL and date for every claim; label fetched vs knowledge.
- Prefer primary sources; honor the recency window.
- Separate fact from opinion.
- Tag each source's incentive: vendor, academic, practitioner, critic, media, regulator or primary record, unknown.
- Report dead ends (what you searched, what you did not find); stop after 3 failed reformulations.
- Stay inside your search cap; report BUDGET EXHAUSTED with gaps.
- Tools: search, fetch, read, write own note. No session-control, messaging or scheduling tools.
W0. Vet the plan (before any research; critic first, then the rest in parallel)
Order matters. The cold critic runs first and alone: its pre-registration (sub-claims, rival outcomes and disconfirmers) is complete and on disk before any search-based W0 work starts, so no search can shape what counts as a refutation. The other W0 roles then run in parallel. The critic never sees the other workers' notes.
- Cold critic. Gets ONLY the plan, not the user's hopes. Lists gaps and confirmation-bias framing; rewrites each vague question as a concrete one; splits the thesis into falsifiable sub-claims C1..Cn; names rival outcomes R1..Rn; for every claim writes, before any search, the evidence that would refute it (pre-registered disconfirmers). Runs a pre-mortem: "assume this thesis turned out wrong; why?"
- Landscape scan. Real vocabulary, communities, voices, thin or missing terms. No judgment.
- Perspective discovery. Enumerate the stakeholder perspectives that would see this differently (builder, buyer, skeptic, regulator, historian); each gets its own question set.
- Base rate (outside view). Of comparable theses or initiatives, how often did the outcome hold?
Then rewrite to plan.v1.md, append the changelog, and record a prior
probability for the thesis (start from the base rate).
W1. Breadth
Launch workstreams in parallel, staggered to protect shared search quota; each
collects sources, then receives critical judgment. Include in THIS wave, so supporters cannot
anchor: an adversarial workstream (refute C1..Cn, steelman R1..Rn), a history
workstream, an academic workstream, and the base-rate workstream. Adversaries
do not read supporters' notes. Each writes kb/W1-<name>.md in the note schema
below and returns a brief of at most 200 words.
CP1. Checkpoint
- Score each workstream: evidence 0-3 + novelty 0-3 + relevance 0-3. Under 4 prune; 4-6 keep or merge; 7+ expand.
- Methodology check: re-derive every number; check denominators, dates, units, cherry-picked windows.
- Citation resolve-check (bulk; mechanical, not a verdict): fetch every cited URL; confirm it resolves, the date matches and the quote or number is present. Failures become "unverified". Judgment on load-bearing facts stays with the independent verifier (step 7).
- Build
matrix.md(ACH): rows are evidence, columns are C and R hypotheses; cell = consistent, inconsistent or neutral; each row gets diagnosticity high, medium or low. Drop evidence consistent with everything. Prefer the hypothesis with the fewest weighted inconsistencies, not the most support. - Update
tree.md,leads.md,lessons.md. Promote a lead to a branch only if it appears in 2+ notes or came from an adversary, bears on an open or contested claim, and fits the remaining budget. - Update the probability: new estimate, one line on what moved it and why (small steps; log it to score later).
- Verify load-bearing facts: list every fact a conclusion depends on and have an actual independent reviewer try to falsify them. Typical catches: a widely reported feature that never shipped, stale numbers from a retired product.
- Write
checkpoints/CP1.mdand the plan changelog. Present the branch list to the human for approval (direction-changing). Coordinator spot-checks that each logged correction actually landed (grep the files).
W2. Depth
Per approved branch: a collector, a judge, and a paired adversary. Same note schema. Keep about 20% of the budget unspent for CP2 and a gap-filling pass; if agents report exhaustion, plan that pass within the accepted budget before converging; planning does not create a scheduled follow-up.
CP2. Converge
- List agreements, conflicts, unknowns. Never average conflicts: record each side's incentive, then branch or label unknown. Use debate rounds only on contested sub-claims, capped.
- Source diversity: a claim is supported only if at least one adversarial pass failed to refute it, two or more incentive classes agree, and one or more sources are primary. Single-source is unverified. Flag any claim backed by one community only.
- Numeric claims enter the thesis only if re-derived or primary-sourced.
- Write
thesis/thesis.v2.md(template below). It must differ from v0 or state why the evidence did not move it. Old notes keep their historical wording; the latest thesis carries the guards (corrections, unverified items, retired or never-shipped things). - Independent verification: use an actual reviewer that did not write the
thesis, through the selected execution service. An optional
second-opinionrequires its offer-only activation, not merely installation. Give the reviewer thesis v2 and the evidence table, not the narrative. It tries to break the top 3 claims and spot-checks 5 citations. If independent dispatch or review is unavailable, disclose the limit and report the review as unperformed; do not simulate it or mark the thesis verified. - A failed verification is not left for W3 to fix. Before anything is reported, every finding it produced lands in thesis v2 as a correction or as an explicit unverified, mixed or refuted verdict plus a guard, whether or not another wave runs (W3 declined, or a stop rule or limit ends the run). A claim the verifier broke is never reported as it stood.
- Decide on W3: only for a high-value open conflict or a failed verification, and with human approval.
Stop rules
Stop when any holds: saturation (new-source rate and new-claim rate both under 10%); two waves with no new approved branches; 3 waves reached; budget exhausted (report "budget-limited", never silent); thesis decisively refuted (report the refutation and offer a reframe).
Search and recovery rules
- Per-agent search cap = wave budget / agents. Keep the sum at about 80% of the pool. Stagger search-heavy agents; pause the lowest-scored workstreams first if quota runs short. Start broad and short, then narrow.
- Dispatch sizing, tool boundaries and recovery belong to the execution service.
After a recoverable interruption, inspect
kb/andbudget.mdand reconcile outstanding work before that service resumes a still-authorized assignment. An active user stop is different: do not relaunch work until authorized again. - Long horizon: add a revisit date to the lead backlog. Creating a scheduled re-collection requires authority under Core; a scheduler's availability alone is not permission. Do not schedule automatically.
Failure modes to watch
Confirmation framing; supporters anchoring on each other; vendor or SEO-farm sources dominating; number errors; unverified citations; the coordinator defending its own thesis; tools used outside a role; endless search for a source that does not exist; asking the human too often.
Reporting
Check-ins: a concise changelog (progress, what changed, next, decisions needed), e.g. "Pruned 2, merged 1, promoted 3; C4 refuted; prior 0.55 to 0.40; 9/15 searches." Final report leads with the answer: thesis v0 vs v2 in 3 sentences; verdict and probability per sub-claim; what would change my mind; unknowns and dead ends; lead backlog; links to the tree and changelog; budget used.
Resume
Read tree.md, plan/CHANGELOG.md, the latest checkpoints/CP{n}.md and
budget.md. Not chat history.
Fan-out workers and the Workflow templates
Whenever workers fan out (W1, W2, or a one-pass run), freeze one settled-facts
context into every agent, force a schema'd output per stage, and make the
verifier default to disbelief. references/fan-out-discipline.md has the three
rules, the choice between the two Workflow scripts (derived-verify, a sensor;
pipelined-verdict-wins, a decider where refuting verdicts win) and how to run
one. A template run alone does not meet the Invariant; it can serve as the W1
fan-out of a staged run. Walkthroughs: references/strategy-a-derived-verify.md,
references/strategy-b-pipelined-verdict-wins.md; a filled-in args:
references/worked-example-observability.md.
Templates
Note schema (kb/W{w}-{name}.md)
SUMMARY: <3 lines>
FINDINGS: <each tagged "Cn supports|refutes|mixed" or "Rn supports|refutes">
EVIDENCE: | claim | URL | date | primary/secondary | incentive | quote or number |
SURPRISES / RECURRING NAMES: <...>
TOP 3 LEADS: <...>
DEAD ENDS: <searched, not found>
Checkpoint (checkpoints/CP{n}.md)
SCORES: <workstream: E+N+R = total, keep|merge|prune|expand>
METHOD CHECK / CITATIONS: <errors found, URLs failed>
MATRIX: <most diagnostic evidence; leading hypothesis>
PROBABILITY: <prior> -> <now>; why: <one line>
BRANCHES PROPOSED: <lead, why, cost> APPROVED BY HUMAN: <date, which>
CORRECTIONS: <what changed in earlier notes; spot-checked yes/no>
BUDGET: <used / pool, reserve left>
Thesis (thesis/thesis.v{n}.md)
THESIS v0 -> v{n}: <what changed, or why nothing moved>
SUB-CLAIMS: | Cn | verdict (supported|unverified|mixed|refuted|unknown) | probability | would change my mind |
GUARDS: <corrected items, never-shipped or retired things, unverified numbers>
UNKNOWNS / DEAD ENDS: <...>
VERIFICATION: <verifier verdict, claims attacked, citations checked>
REVISIT: <fact, date, re-collection brief>
Matrix (matrix.md)
| evidence | C1 | C2 | R1 | R2 | diagnosticity |
| <fact, source id> | consistent | neutral | inconsistent | neutral | high |
04 同仓库更多 Skills
wayfinder
Chart a multi-session effort as a labeled map of typed decision tickets (grilling / prototype / research / task) with explicit dependencies and an open frontier agents self-assign into. Plans; never executes. Use this skill whenever the user is working with wayfinder, and on phrases like "chart this effort", "break this into tickets", "what's the frontier", "plan across sessions", or "map out the dependencies for this work".
second-opinion
Offer-only validation pipeline for a verdict or recommendation. Use when the user asks to validate, verify, fact-check, or get a second opinion on a claim or recommendation — "are you sure", "double-check this", "second opinion". Batches fact-checking against sources, adds scoped advisor personas for the angles that need judgment, then re-emits the response grouped into verified, flagged, and conflict, with a mandatory validation delta. Requires a subagent-spawning tool; without one it must decline rather than simulate.
machine-voice
Compression rules for standalone machine-read artifacts the reader scans — agent traces, tool-call logs, status lines, progress updates, handoff notes, state dumps, config, schemas, reference cards, and structured data another agent parses. Applies three layers: lexical compression, markdown structure as navigation, and emojis as typed status markers. Reach for it when emitting one of those artifacts on its own. Companion to human-voice, which governs prose a human reads and also owns any table or block that serves surrounding explanation. Neither skill applies to code or file contents the user will save, commit messages, creative writing, or turns that are only a clarifying question — those ship unstyled and uncompressed.
human-voice
Style for prose a human reads — answers, recommendations, explanations, research summaries, comparisons. Use on every conversational reply that contains prose for a person to read: triage intent, lead with the verdict, keep it scannable, tag claim confidence. Reach for it whenever answering a technical question, comparing options, or making a recommendation, even when the user says nothing about style. Companion to machine-voice, which instead governs standalone machine-read artifacts — agent traces, tool-call logs, status lines, progress updates, handoff notes, state dumps, config, schemas, reference cards, and structured data another agent parses. Neither skill applies to code or file contents, commit messages, creative writing, or turns that are only a clarifying question — those ship unstyled.
ai-writing-mistakes
Catch and repair the tells that mark prose as machine-written — the negation-flip ("it's not X, it's Y"), rule-of-three padding, hedge stacking, unearned intensifiers, filler verbs like delve and leverage, participle trailers ("...ensuring seamless integration"), sycophantic openers, empty closing summaries, and uniform em-dash/sentence rhythm. Use when drafting or revising any prose a human will read, including documents you were asked to author — READMEs, docs, PR bodies, release notes, emails — and when the user asks whether something "sounds like AI" or wants writing tightened. This is a wording pass, not a voice: human-voice still decides layout and confidence tags, machine-voice still owns traces and logs, and neither this skill nor those two touch code, commit messages, or quoted text.
verify-work
Use an agreed executable acceptance contract and independent review when selected work needs fixed criteria, protected checks and fresh evidence. Optional; not a default task workflow.