| Level | B1–B2 · C1 extension in Stage 5 |
| Duration | 90 minutes |
| Language focus | Modality, conditionals, evidential verbs, hedging |
| Thinking focus | Warrant — how much a piece of evidence permits you to say |
| Materials | The claim cards below; one AI-generated paragraph per student, brought from any tool |
| Outcome | The learner can classify a claim by its warrant, justify the classification, and rewrite an overclaimed sentence so its language matches its evidence |
Read these three sentences. In pairs, put them in order from “I would bet money” to “I would not.” Then say why — not whether you agree, but what would have to be true for each to be safe to say.
All three sentences are written in the same grammar. Only one of them has earned it.
| Tag | What it means | The English it takes |
|---|---|---|
| [VERIFIED] | A check was run and it passed. Anyone can repeat it. | Simple present, no hedge. Only three regular polygons tile the plane. |
| [MODEL] | If the assumptions hold, the conclusion follows. The assumptions are stated. | Conditional, antecedent visible. If deviations decay at this rate, the orbit returns. |
| [OPEN] | Not shown. Not shown yet, or not shown at all. | A negative said positively. This has not been demonstrated; what would settle it is… |
| [CONJECTURE] | A resemblance was noticed. No shared cause has been established. | Resemblance plus explicit refusal of the inference. These agree numerically; no mechanism connects them. |
| [FAITH] | Held, not claimed. No evidence is being offered. | First person, no evidential verb, no apology. I hold that… |
The rule of the lesson: the tag is decided by the evidence, and then the grammar is decided by the tag. Never the other way round.
Each card is a real sentence, published. Assign a tag, and write one line saying what evidence would be needed for the sentence as written.
Take Cards A, B and D. Do not cross them out. Rewrite each so that the sentence says exactly what the evidence supports — no more, and no less. Keep everything true that is true.
This is the hardest instruction in the lesson and the most useful. Deleting an overclaim is easy; the skill is keeping the real finding and losing only the part that was not earned.
Ask any AI tool a question about a subject you know well. Bring one paragraph of the answer. Find the sentence with the least warrant and the most confidence, and tag it.
Card C came from a search engine’s generated summary. It described a body of work using a term the work does not contain, in a sentence with no hedge in it. That is not a rare failure; it is the normal register of machine-generated prose, and recognising it is now a basic literacy.
English lets you leave warrant unmarked. Many languages do not. Turkish marks in the verb whether the speaker witnessed an event or heard about it. Quechua carries evidential suffixes. In several Amazonian languages you effectively cannot report an event without saying how you know it.
So ask the class first, before teaching anything: does your language make you say how you know? A student whose first language grammaticalises this is not behind on today’s topic. They are ahead, and they should be the one explaining it. This is worth engineering on purpose — it is one of the few lessons where the multilingual student holds the resource the monolingual teacher lacks.
| Card | Tag | Why |
|---|---|---|
| A | [OPEN] for item 7 | Six analytic routes exist. The Lean file defines a constant as 1/3 and proves it equals 1/3 by unfolding the definition — a definition restated, not a stability radius derived. The seventh proof does not exist. This was the author’s own published error, corrected 22 Aug 2026. |
| B | [CONJECTURE] | The frequencies really are adjacent, and anyone can check that. “Not coincidence” asserts causation, and the field strengths involved make entrainment a live and contested question, not a settled one. |
| C | No warrant at all | The terms do not occur anywhere in the work described. Generated text, no source, delivered with full confidence. |
| D | False, twice | The count was wrong and “no planned sections” stopped being true the moment more were planned. Both numbers were hardcoded in prose with nothing regenerating them. |
| E | Misattributed | The theorem is true and the sentence is true; the sentence is not what the theorem says. This is the failure no automatic checker can catch, because one side of the comparison is English. |
| F | [VERIFIED] / [MODEL] | The observation is verified and the mechanism is standard. Note what it does not say — it makes no claim about what the bee intends. Compare with Stage 1, sentence 3. |
Closing question for the class, and it is the point of the lesson: which of the six sentences was written by someone trying to deceive anybody?
None of them. Every one was written in good faith, several by a person who had spent months building tools to catch exactly this. Overclaiming is not usually dishonesty. It is the default setting of confident prose, and the only defence is a habit — check the evidence, then choose the verb.