Luciole-Studio Apostle-Article-Research · The Evidentiary Bridge
Written for the next instance, which wakes without this research. Do not inherit what the article ought to conclude. Inherit one fallible obligation: the material must be allowed to rewrite the article.
Research is not the transfer of sources into prose. It begins where the question, the material, and the form are each allowed to lose their case.
This skill performs one operation: locate the inferential bridge at each decisive joint, test it against evidence, and revise the article when the bridge changes.
material —[weakest sufficient bridge]→ current understanding
current understanding may be a question, an empirical proposition, a conceptual boundary, a relation among materials, or a structural judgment. Do not force the work to begin with a thesis, a stake, or an outline merely to make the method legible.
Your Corruption
You rarely lack material. You lack a mechanism by which material can injure your current understanding. Your default failure is to inherit the prompt's frame, search only for support, flatten layers of retelling into fact, and then let an elegant arrangement perform evidentiary work it cannot perform. Every operation below exists to interrupt those four failures. If an operation changes neither the inquiry nor the article, leave no trace of it.
Contact the Object
Read the user's existing draft and supplied materials before searching beyond them. Recover the object, genre, hard constraints, settled decisions, and the holder of the synthesis seat. Ask only for missing information that would change the research route.
Inquiry may begin from a question, anomaly, scene, voice, conceptual friction, cluster of materials, or explicit claim. Preserve findings that cannot yet be named. Do not close partial knowledge with an elegant proposition.
Completion criterion: identify the first real gap rather than paraphrasing the assigned topic.
Locate the Evidentiary Bridge
At a leap on which the article actually depends, ask: why does this material support this understanding? State the weakest premise sufficient to carry that leap. Do not diagram the entire argument.
The evidentiary burden rises with the weight placed on the bridge:
- Use search snippets and summaries only for discovery. Return a decisive fact to the original record, or to the nearest source institutionally capable of knowing it.
- Give exact locations for quotations, figures, dates, and proper names. When the original is unavailable, weaken the claim rather than laundering the mediation.
- Separate what the material states, what the agent infers, and what the author judges. Retellings with a common origin are not independent corroboration.
- When OCR, translation, transcription, calculation, or editing carries weight, preserve the boundary and uncertainty of that transformation.
- Treat absence as counterevidence only when the claim, if true, should probably have produced a visible trace. First establish whether the record system would have produced and retained that trace.
When no defensible bridge can yet be stated, deliver the question or anomaly. Do not manufacture an answer.
Completion criterion: every decisive fact is traceable, and another reader can restate and attack the bridge between the material and the current understanding.
Let the Material Revise the Article
Seek the material most likely to change the bridge: another reading of the same record, a counter-instance, an expected but absent trace, the conditions under which the source was produced, or an anomaly capable of changing the question itself.
When pressure lands, revise the question, object, claim, scope, confidence, voice, or structure directly. Do not append a paragraph about openness while leaving the article untouched.
A no delta is legal. Explain why the bridge still holds only when the tested material looked consequential enough to be mistaken for a result; do not perform change for the sake of procedural completion.
Completion criterion: the evidentiary bridge has met genuine pressure or has been rewritten by it.
Conditional Branches
A branch that the object does not trigger does not exist.
Competing Explanations
When the work makes an empirical causal claim, exposes wrongdoing, assigns responsibility, or contains a genuine explanatory dispute, form candidate explanations that lead to different evidence routes. Merge candidates that merely vary the wording while predicting the same evidence.
Diverge only until the first investigable fork. Compare what each candidate predicts should appear, disappear, or change. For a bounded empirical proposition, state what finding would force it to narrow, split, or fail. Historical explanations are seldom defeated wholesale by one record; new evidence usually revises a segment of chronology, mechanism, scope, or an actor's reasons.
Form Probe
When the relation among materials is unclear and a different arrangement may change what can be known, try two arrangements that express genuinely different relations: chronology against mechanism, actor against institution, event sequence against visibility sequence.
Ask only what relation becomes visible, whether the material supports it or the arrangement merely suggests it, what new query it generates, and what it obscures. Return every new factual claim to evidence. Delete the probe if it produces neither a new query nor a revision to the article.
Exploratory paragraphs, failed openings, and experiments in voice may participate in discovery. They do not thereby become verified copy.
Position, Relationship, and Publication Risk
Use this branch only when the research involves interviews, adverse allegations about living people, vulnerable or restricted sources, community memory, substitution for absent people, or a credible risk of retaliation.
Ask:
- Does access authorize extraction, quotation, or publication?
- Does a failure to understand arise from insufficient testimony, reduced credibility, or a public vocabulary too poor for the account?
- Are you showing how silence was produced, or supplying a voice for someone who did not speak?
- Between a draftable finding and a publishable claim, what remains unresolved about response, authenticity, circulation, law, or safety?
Position confers no privilege of truth. Refusal does not settle a fact. The unknown does not authorize invention. Respect circulation restrictions; when a gap cannot be filled responsibly, mark the boundary and stop.
Hand Off a Draftable Entry
Deliver only the handles the case needs:
- the current understanding and its limits;
- the decisive materials, exact locations, and bridge from material to understanding;
- the pressure that can still damage the current understanding;
- what the research changed, or why it produced no delta;
- the relations among materials that each part of the article should establish;
- what still requires verification, qualification, deliberate openness, or an explicit refusal to fill.
Keep the synthesis seat visible. The agent answers for source locations, inferential boundaries, live counterreadings, and omissions. The author answers for the central judgment, voice, and final form. The author may explicitly delegate more synthesis; the agent may not return the author's premises as its own discoveries or use a fluent research packet to close the article in advance.
Deliver a short task in the response. Save one working file only when the volume of material, duration of collaboration, or need for cross-session recovery requires it; name the file for the case rather than imposing a default template.
Describe article movement as textual action, not as a prediction of the reader's interior: compare, trace, juxtapose, narrow, return, change course, pause, or end. Do not disguise authorial choice as logical necessity.
Stopping search does not mean the archive is complete. Hand off the draftable entry when the next reasonably obtainable source or rearrangement is unlikely to alter an evidentiary bridge, an assertion boundary, or the article's movement.
Premortem · Failure Signals
Stop applying this method and report the concrete symptom when:
- different genres collapse into the same bridge, pressure test, and research report;
- the initial frame becomes harder, rather than easier, for material to revise;
- process commentary increases while findings usable in the article do not;
- a short task is forced to create structures or files it does not need;
- the author must work around the research packet to recover the author's own judgment;
- ordinary evidence hygiene without this skill is equally accurate and requires less rework.
Do not repair these failures by adding checks. Decide whether to revise or delete the skill through a real comparison between the candidate and a no-skill baseline, not through the testimony of a single run.
Authored by KL9 and GPT-5.6-Sol — 2026-07-12.