Claim-Evidence Boundary

A claim-evidence boundary states exactly what a result proves and what it does not prove. It is useful for papers, Korean synthesis posts, project notes, and concept pages.

The minimal form is:

$$(\text{claim},\text{task},\text{data},\text{protocol},\text{metric},\text{baseline},\text{uncertainty})\to \text{supported scope}$$

If any part is missing, the claim should be narrowed or marked to verify.

Boundary Fields

Field	Question
Claim	What exact statement is being made?
Task	What input and output space does the evidence test?
Data	Which population, benchmark, preprocessing, and split define the examples?
Protocol	How were model selection, threshold choice, early stopping, and final testing separated?
Metric	What number supports the claim, and what does it average over?
Baseline	What comparison makes the result meaningful?
Uncertainty	Is the effect larger than seed variation, confidence interval, or measurement noise?
Scope	Where should the claim be expected to hold?

Narrowing Rule

Write the broad version first, then narrow it until every word is supported.

Broad Claim	Narrower Claim
This model generalizes to proteins.	This model improves metric $M$ on a protein-family split for benchmark $B$.
This method improves docking.	This method improves pose plausibility under the reported pocket and ligand-preparation protocol.
This architecture is better.	This architecture improves metric $M$ under matched data, objective, and training budget.
This objective learns better representations.	This objective improves downstream metric $M$ under the stated frozen or fine-tuned protocol.
This benchmark proves real-world performance.	This benchmark supports the specific population, split, and allowed-information boundary it defines.

Formula View

An empirical claim usually estimates a target quantity:

$$Q={\mathbb{E}}_{z\sim p_{\mathrm{target}}}\left[m(f_{\theta },z)\right]$$

but reports a finite estimate:

$$\hat{Q}=\frac{1}{n}\sum _{i=1}^{n}m(f_{\theta },z_i),z_i\sim {\hat{p}}_{\mathrm{eval}}$$

• $Q$: target quantity the claim wants to support.
• $\hat{Q}$: measured estimate.
• $z$: evaluation unit, such as example, query, molecule, complex, pose, target, or task.
• $p_{\mathrm{target}}$: intended population or deployment distribution.
• ${\hat{p}}_{\mathrm{eval}}$: benchmark or test-set distribution actually sampled.
• $m$: metric contribution for one unit.

The claim is only as broad as the match between $p_{\mathrm{target}}$ and ${\hat{p}}_{\mathrm{eval}}$.

Molecular Modeling Checks

Claim Type	Boundary To State
Molecule property	molecule standardization, label source, scaffold or time split, metric
Protein prediction	sequence identity or family split, residue indexing, structure source, metric
Protein-ligand complex	ligand scaffold, protein family, complex-pair split, pocket definition, coordinate source
Docking pose	receptor preparation, ligand state, pose RMSD rule, geometry checks, known-pocket assumption
Virtual screening	active/decoy construction, enrichment metric, property bias, target context
Generative molecule	validity, uniqueness, novelty, diversity, constraint satisfaction, filtering rule

Use In Posts

For Korean synthesis posts, include only the boundary needed for the reader:

• What claim is being discussed?
• What evidence supports it?
• What split or benchmark limits the claim?
• Which wiki note holds the detailed metric, formula, or paper evidence?

Long evidence tables belong in Evidence table or paper notes. The post should keep the boundary readable.

Red Flags

• A random split is used to claim new-scaffold or new-protein generalization.
• A metric is reported without model-selection or threshold-selection details.
• A benchmark score is treated as deployment performance without population matching.
• A molecular result ignores standardization, assay context, or leakage.
• A small score difference is interpreted without uncertainty or paired comparison.
• A qualitative example is used to imply a broad quantitative claim.

Related

• Evaluation protocol
• Benchmark intake
• Metric selection
• Baseline
• Confidence interval
• Paired comparison
• Leakage
• Molecular modeling data and evaluation
• Claim extraction
• Evidence table