Over-Claim Guard (Audit Lane Reference)

Reference for the claims_vs_evidence lane: how to spot wording that outruns the evidence, and the conservative wording it should be revised to. The goal is calibration, not timid prose — strong evidence earns strong wording; weak evidence must use weak wording.

Boundary with CLAIM_EVIDENCE_CONTRACT.md

• CLAIM_EVIDENCE_CONTRACT.md decides whether the visible evidence supports a claim (the strength ladder unsupported → observed → supported → strong).
• This file decides how a claim should be worded once its strength is known (which verb / qualifier keeps the sentence inside the evidence).

In a finding, set claim_strength from the contract and put the bounded rewrite in allowed_wording / the overreaching phrasing in forbidden_wording. Where they conflict, the contract wins (substance over phrasing). Emit these as comment_type: claim_accuracy.

Certainty ladder (verbs, strongest → weakest)

demonstrate / prove                         ← intervention + controlled experiment
reveal / identify / find                    ← strong effect, multi-method or replicated
indicate / suggest                          ← significant but single-method
support / be consistent with                ← trend, agrees with prior work
may indicate / could suggest / appear to    ← marginal or predictive
hint at / point toward                      ← very weak signal or hypothesis

Flag any claim whose verb sits above the rung its evidence can reach.

Substitution tables (flag ❌ → suggest ✅)

1. Causal (correlation stated as causation — the most common over-claim)

❌ over-claim	✅ conservative
caused by	associated with / linked to
drives / determines	contributes to / is associated with
responsible for	implicated in / associated with
proves that	indicates / provides evidence that

Causal wording is justified only with a controlled intervention (ablation, randomized assignment, A/B test), an instrumental-variable design, or a reproduced established mechanism.

2. Novelty / firstness (reviewers verify these in seconds)

❌ over-claim	✅ conservative
the first to	the first, to our knowledge / among the first to
novel (self-labeled)	name what is new; drop the label
unprecedented	substantial / notable

3. Universality (one setting studied, all settings claimed)

❌ over-claim	✅ conservative
always / never	generally / rarely
in all cases / universally	in the cases studied / across the benchmarks evaluated
any dataset	the datasets sampled

4. Effect size (vague magnitude word, no number)

❌ over-claim	✅ conservative
strong / large improvement	reduces error by X% / β = X.XX (95% CI: …)
significant gain	improved from X to Y (p = …)
robust	consistent across N runs / stable under [perturbation]

If the number carries the weight, the adjective should be dropped.

5. Temporal / inferred order (present data, past mechanism)

❌ over-claim	✅ conservative
X drove the change	the change is consistent with X
occurred at time T	estimates suggest ~T (CI: …)

6. Application / impact (downstream uses not demonstrated here)

❌ over-claim	✅ conservative
will revolutionize	has potential implications for
solves the problem of X	addresses one aspect of X
ready for deployment	provides a candidate approach for [setting]

7. Comparison (disparaging prior work)

❌ over-claim	✅ conservative
previous methods failed to	previous methods were limited by
outperforms all prior work	compares favorably with [specific methods]

High-frequency trap phrases

trap	safe rewrite
"Our results demonstrate X." (X causal)	"Our results are consistent with X."
"This is the first work to …"	"To our knowledge, among the first to …"
"X plays a critical role in Y."	"X has been implicated in Y / may contribute to Y."
"These findings have important implications for …"	"These findings provide a basis for further study of …"
"Strongly supports"	"Is consistent with / provides evidence in line with"

When NOT to flag (reverse calibration)

Do not flag strong wording that the evidence earns:

• a controlled intervention (ablation / RCT / A-B) → demonstrate is fine;
• multiple methods / datasets / seeds replicate the result → robustly (with evidence named);
• a reproduced established mechanism → confirms / validates;
• a large effect with a strong statistic stated → strong wording plus the number.

Flagging earned strong wording is a false positive — leave it.

How the lane uses this

The claims_vs_evidence lane flags over-claim wording as claim_accuracy findings with allowed_wording / forbidden_wording filled in. This is the LLM-judgment complement to the writing skills' deai_check.py over-claim script, which only catches a few unambiguous phrases; the tables above cover the contextual judgment the script cannot make.