# Research Loop · AI-Native 研究工作流模板

> Copyable artifact for《AI-Native Research》RES 13. The whole volume reduced to one runnable loop.
> Isomorphic to engineering's spec-driven loop (Specify → Plan → Execute → Verify → Integrate → Learn),
> with two research-specific load-bearing parts: **the knowledge graph comes first** (the evidence base
> is the spec) and **integration takes priority over retrieval**.
>
> Evidence discipline: claims marked 〔evidence ledger〕 carry reliability; claims marked
> 〔exploratory ledger〕 are leading indicators / boundaries / falsifiers, not proven facts.

---

## The six-step loop

| # | Step | In one line | The load-bearing move |
|---|------|-------------|-----------------------|
| ① | FRAME | Stand up the evidence base; write the criteria | Build a traceable knowledge graph (legible to agents / traceable / falsifiable / integrable); write down explicit criteria for "worth believing / worth knowing." **The base is the spec — this precedes generation.** |
| ② | GENERATE | Parallelize in-paradigm actions | Search / hypothesize / design experiments / analyze → hand to generation (RES 10 left cell). Each output carries evidence edges into the base. **What does not enter the base does not count.** |
| ③ | TRIAGE | Believability ledger; suspend paradigm-level | Triage each batch on "evidence strength × paradigm distance": drop in-paradigm noise, integrate the believable, **suspend the paradigm-distant to seek discriminating evidence — do not kill as noise.** |
| ④ | INTEGRATE | Synthesize across knowledge, not more retrieval | The human's scarce act: stitch never-juxtaposed claims into new understanding. Watch the ratio of integration artifacts to raw output. **Do not let generated-but-un-integrated knowledge pile into a mountain.** |
| ⑤ | OWN VALUE | Set direction; leave the change-the-variable door | Give "worth" an owner; keep a human-initiated channel to "change the schema / change the variable," resisting the generation layer's conservative bias. |
| ⑥ | FEED BACK | Errors become guardrails | Feed each retracted / refuted / mistakenly-killed novelty back as a new rule or node type in the base. Fewer mistakes next round. **This closes the loop.** |

---

## The believability ledger (③ in detail) — two axes, never one score

Do **not** collapse credibility into a single number. A single score lets "distance from the
existing-literature distribution" become the only proxy, which under-scores the genuinely novel.
Book the two axes separately:

|                         | **Evidence weak / untraceable** | **Evidence strong / traceable** |
|-------------------------|---------------------------------|---------------------------------|
| **Paradigm-distant (far)** | ⚠ **Do not kill yet — seek evidence.** A reframing is necessarily thin on evidence at birth (Einstein 1905, Lorentz both merely fit data at first). Suspend; go find the decisive discriminating evidence. | **Human — maybe a reframing.** Evidence holds while breaking the frame: send to a human, ask "did it switch to a new map?" |
| **In-paradigm (near)**     | **Doubt — in-paradigm noise.** Graph rules auto-downweight / block; no human bandwidth spent. | **Believe — integrate.** Traceable, base-consistent: auto-clear into the synthesis loop. |

The cell to watch most is **weak × far**: deleting it as noise is exactly hypernormal science's
killing move (RES 08).

---

## The in-paradigm / paradigm-level split (used in ② and ③)

Operational test: **if you can write a machine-checkable acceptance criterion → left (hand to
generation); if you can only appeal to "for whom, under which value frame" → right (stays with the
human).**

| Action | In-paradigm (→ generation) | Paradigm-level (→ human) |
|--------|----------------------------|--------------------------|
| Search | nearest-neighbor / RAG / citation tracing | — |
| Hypothesize | next checkable gap where data is thickest | ask whether the variable set is the right level of description |
| Design experiments | enumerate / sweep / ablate within a standard paradigm | cross-sensory analogy to embodied intuition |
| Analyze | statistics / fitting / symbolic regression over predefined variables | judge whether novelty is a reframing or noise |
| Judge | machine-checkable correctness | judge whether the conclusion is worth knowing |

---

## Starting path (do not build it all at once)

1. Pick **one** step where execution is already abundant but judgment is not yet externalized
   (literature synthesis, parameter sweeps).
2. Stand up a **minimal** traceable evidence base (① + ②).
3. Once it runs, add ③ the ledger and ④ integration.
4. **Reinvest the hours ② saves explicitly into ③④ — not into producing more papers.**
   (This is the single easiest place to go wrong; see RES 14 applicability boundary.)

Three starts, one line: **stand up a traceable evidence base → use the believability ledger to choose
where to inject human judgment → reinvest saved hours into integration and owning value.**

---

## Signals (are you running it right?)

- **Right:** human-hours on in-paradigm actions keep falling; hours on paradigm-level judgment keep
  rising; share of left-cell outputs landing in the traceable chain on first pass rises; integration
  artifacts grow relative to raw output.
- **Wrong (failure modes):**
  - forcing paradigm-level actions into the left cell ("let AI decide where to push the research") —
    handing value judgment away;
  - leaving in-paradigm actions in the right cell (humans still tracing citations by hand) — failing
    to abundify execution;
  - **hypernormal science** — output rises while topical breadth shrinks, "reframe" contributions
    stay low, citations concentrate (rising Gini): every metric green while science gets narrower.

---

## Source anchors

- 〔evidence ledger · grade Ⅱ〕Hao, Xu, Li & Evans, "AI tools expand scientists' impact but contract
  science's focus," *Nature* 649(8099), 2026, DOI 10.1038/s41586-025-09922-y. ~41.3M papers; topical
  coverage −4.63%; scholar interaction −22%; citation Gini 0.754 vs 0.690. Observational bibliometrics;
  causal claims with care.
- 〔exploratory ledger · grade Ⅳ–Ⅴ〕Djajadikerta, "Designing AI for Disruptive Science," *Asimov Press*,
  2026-03-23, DOI 10.62211/29ej-27et. Opinion/review: hypernormal science; the map / Beck / Farr
  mechanism; meta-science's "model organism." An essay's argument, not data; its cited empirics each
  need tracing to original sources for grading.