Tense Guide

How verb tense maps to paper sections, and which present-tense signal words usually mark a mistake. The goal is not blanket past tense — each section has a convention, and the most common error is narrating Methods/Results in the present tense.

The script (deai_check.py) flags a focused set of present-tense reporting verbs inside Methods/Results as [Script] LOW traces. The tables below cover the judgment calls the script deliberately leaves alone (notably is / are).

Tense by section

Section / part	Default tense	Example
Abstract — background	present	"Long-context inference is expensive for ..."
Abstract — methods	past	"We trained ... / Models were evaluated on ..."
Abstract — results	past	"The model achieved 92.3% / We observed ..."
Abstract — conclusion	present	"These results provide a basis for ..."
Introduction — known background	present	"Quantization reduces memory footprint."
Introduction — a specific prior study	past	"Vaswani et al. introduced the Transformer."
Introduction — prior conclusion still holding	present	"Attention is effective for long sequences."
Introduction — aims of this work	past (+ some present)	"Here we propose X and present a study of ..."
Methods	past (absolute)	"We sampled ... / Inputs were normalized ..."
Results	past (absolute)	"Accuracy ranged from ... / We found ..."
Results — describing a figure/table itself	present	"Figure 2 shows ... / Table 1 lists ..."
Discussion — restating a finding	past	"We found that ..."
Discussion — interpreting	present	"These results suggest that ..."
Discussion — inference / speculation	present + modal	"may indicate / could explain"
Discussion — limitations	present	"A limitation of this study is ..."
Figure / table caption	present (absolute)	"Figure 1 shows the architecture of ..."

Signal words (scan Methods / Results)

The most common slip is a present-tense reporting verb where past tense is expected:

• shows / reveals / demonstrates / indicates / presents / confirms / achieves / outperforms in Methods/Results narration → usually should be showed / revealed / demonstrated / …. The script flags these.

is / are — check by hand (not flagged by the script)

Present-tense is / are in Methods/Results is often a tense error, but it has too many legitimate uses to flag automatically. Keep it when it is:

• a definition: "Let G be ... / The loss is defined as ...";
• a standing general truth: "Cross-entropy is convex in the logits.";
• describing a figure/table: "Table 2 is organized by ...";
• a software capability: "The library supports mixed precision.".

Otherwise prefer past tense: "The threshold was set to 0.5" (not "is set"), "Samples were drawn" (not "are drawn").

Exceptions the script skips (present tense is correct)

1. Figures / tables / equations as subject: "Figure 3 shows ...", "as shown in Fig. 4", "Table 1 lists ...", "Equation 2 gives ..." — describing the artifact itself takes present tense.
2. Software / tool capability: "The toolkit provides ...".
3. General definitions and standing truths that happen to sit inside a methods paragraph.

If a flagged line is one of these, it is a false positive — leave it.

Boundary with other guides

• This guide is about tense. Wording strength (causal / firstness / universality) lives in OVER_CLAIM_GUARD.md.
• Surface grammar beyond tense lives in modules/GRAMMAR.md.

Script support

deai_check.py flags present-tense reporting verbs in Methods/Results as [Script] LOW traces (config: tense: in references/AI_TONE_THRESHOLDS.yaml, toggle with enabled). It guards against figure/table/software false positives but cannot judge is / are — use the checklist above for those.