paper-opportunity-radar

# Paper Opportunity Radar

Treat the literature as an evidence base to traverse over time, not a feed to
summarize. Build a cumulative corpus, audit important papers at claim level,
look for what happened after publication, and turn only defensible gaps into
testable opportunities.

This skill runs when invoked. It does not silently create a background
scheduler. If the user wants a daily cadence, preserve resumable state and give
their scheduler the recurring prompt in [Daily Operation](#daily-operation).
Never say monitoring is active until a recurring job actually exists.

## Non-negotiable Standards

- **Never claim exhaustive coverage without a bounded corpus.** Define the
  databases, query strings, dates, languages, document types, and traversal
  cursor. Say "all records returned by this protocol," not "all papers ever."
- **Fact-check at claim level.** Every factual statement in a report needs a
  fetched source. Mark interpretations as `INFERENCE` and gaps as `UNKNOWN`.
- **A paper verifies what its authors reported, not that the result is true.**
  Independent replication, convergent evidence, or real-world validation is a
  separate evidence layer.
- **Do not infer fraud.** Use `INTEGRITY CONCERN` for observable anomalies. Use
  `RETRACTED`, `CORRECTED`, or `FORMAL MISCONDUCT FINDING` only when the
  publisher, institution, court, or regulator supports that status.
- **Citation count is attention, not validity.** Peer review, venue prestige,
  author reputation, and code availability are signals to inspect, never proof.
- **Absence of search results is not proof of novelty.** Report exactly where,
  how, and when prior art was searched and use `NOT FOUND IN SEARCH`.
- **Keep four judgments separate:** evidence strength, unexploredness
  confidence, implementation feasibility, and real-world value. Never average
  them into one score that hides a fatal weakness.

## Inputs and Defaults

Require a **topic**. Infer the remaining inputs when safe:

- decision: explore a business, find a project, understand feasibility, or
  monitor scientific progress;
- historical horizon: earliest searchable record through today by default;
- domains and adjacent fields;
- geography, language, and publication-type limits;
- available skills, capital, equipment, compute, data, and time;
- daily depth: `standard` by default; `brief` or `deep` when requested.

Ask one concise question only when the topic is missing or a domain ambiguity
would materially change the corpus. Otherwise state the inferred scope and
begin. For medical, legal, safety-critical, or investment decisions, describe
the work as research analysis and identify where a qualified professional is
needed.

## Durable Workspace

Use the user's requested location. Otherwise use
`research/paper-opportunity-radar/<topic-slug>/` and keep:

```text
scope.md                         Stable boundary, query atlas, and run policy
search-log.md                    Exact source/query/filter/time/result log
corpus.tsv                       Deduplicated paper inventory and queue
opportunity-ledger.md            Living opportunities, blockers, and verdicts
papers/<canonical-id>.md         One deep-audit dossier per paper
reports/YYYY-MM-DD.md            Detailed daily report
```

Use DOI as the canonical ID when available, then PMID/arXiv/other repository
ID, then a normalized title-year hash. `corpus.tsv` must include:

```text
paper_id title year canonical_url discovered_at discovery_source status relevance evidence_verdict next_action
```

Allowed status values are `discovered`, `triaged`, `queued`, `audited`,
`monitor`, and `excluded`. Never overwrite a prior report. Update living files
atomically and preserve user edits.

## Workflow

### 1. Frame a Falsifiable Search

Write the topic as:

1. the core phenomenon, mechanism, or problem;
2. synonyms, former names, acronyms, and neighboring terminology;
3. inclusion and exclusion rules;
4. the opportunity decision the research should inform;
5. what evidence would make a paper or opportunity uninteresting.

Read [references/discovery-protocol.md](references/discovery-protocol.md), then
write the query atlas and source plan to `scope.md`. If the topic is huge,
partition it by mechanism or use case. Do not silently narrow it.

### 2. Resume Before Searching

Read `scope.md`, the most recent report, unresolved paper dossiers,
`opportunity-ledger.md`, and `corpus.tsv`. Resume the recorded source cursor and
backfill window. On the first run, create these files and label the historical
corpus `BASELINE IN PROGRESS` until every planned source/time band has been
visited.

### 3. Traverse Three Lanes

Every daily run covers:

- **new delta:** papers published or indexed since the last successful run;
- **archive backfill:** the next unvisited historical source/query/time band;
- **follow-up graph:** references, forward citations, corrections, replications,
  and later implementations connected to high-value or disputed papers.

Use at least two independent scholarly indexes plus one domain index when one
exists. Search exact terms, controlled vocabulary, mechanism synonyms, and
application language. Record every exact query, filter, timestamp, result
count, new-paper count, and limitation in `search-log.md` before changing lanes.

Deduplicate before screening. A database hit is not a paper read. Triage every
new candidate for scope, paper type, accessible evidence, and likely audit
value; deep-audit only what can receive genuine attention. Queue the remainder
with a reason and resume cursor instead of pretending it was reviewed.

### 4. Select Papers for Deep Audit

Prioritize a balanced set:

- load-bearing or field-shaping claims;
- surprisingly large effects or unusually broad conclusions;
- under-cited work with a specific, testable mechanism;
- papers made newly feasible by cheaper compute, sensors, fabrication, data,
  distribution, regulation, or standards;
- negative results and abandoned prototyp