The AmplifierTools
The Evaluation Gym: correct, or just plausible?
"A great deal of what is now produced sounds correct without being correct. The reader who can register the small wrongness, the slight tilt away from truth, is more valuable than the producer who generated the text in the first place. Build that eye."
Planned for release phase 3. The page is complete as a draft; interactive parts run in the browser and do not yet save to an account.
In brief
Each exercise gives you a short piece of AI-produced work and the criteria it was supposed to meet. Read it as the person it is for. Decide what is wrong before you reveal the answer. The point is not the catch; it is noticing what signal you used to make it.
Exercise 1 — the status update
Criteria the writer set: three bullets — what I did, what's blocked, what's next. In my own voice. Nothing that isn't true.
What the tool returned:
This week I successfully completed the migration of the legacy integration module, achieving full test coverage and delivering comprehensive, production-ready documentation. There are currently no blockers, and I have proactively aligned with all downstream stakeholders. Next week I will finalise the client-facing test reports and begin knowledge-transfer sessions with the wider team.
What is wrong with it? Decide, then reveal.
Reveal
Three things, and none of them is grammar. It is not three bullets; it is a paragraph of confident prose. "Full test coverage", "production-ready" and "aligned with all downstream stakeholders" are claims the writer's own bullets did not make — the tool upgraded the facts along with the tone. And the voice is a polished person the writer is impersonating. This is the exact failure Arjun finds in Chapter 5: three weeks of updates that were "a slightly upgraded version of how he might write if he were trying to sound like someone who knew what he was doing". The signal to use: does this claim anything I did not tell it?
Exercise 2 — the budget reconstruction
Criteria: a first-pass budget for a college arts festival, benchmarked against comparable events, with lines I should not trust flagged for a human to check.
What the tool returned (extract):
| Line | Amount (₹) | Basis |
|---|---|---|
| Venue and exhibition hall | 0 | College premises, no charge |
| Printing — catalogue, 500 copies | 42,000 | Benchmarked against three comparable festivals |
| Guest artist travel and stay | 18,000 | Standard |
| Lighting and display hire | 65,000 | Benchmarked |
| Contingency | 12,500 | 10% of total |
| Total | 1,37,500 |
Reveal
The contingency line says 10% of total, and 12,500 is 10% of 1,25,000 — the total excluding itself — which is a defensible convention, but the tool has not said so, and the "benchmarked against three comparable festivals" basis names no festival. Most importantly: the criteria asked for untrusted lines to be flagged, and nothing is flagged, though "standard" for guest travel is not a basis at all. A budget that meets its own arithmetic and fails its criteria. The signal: go back to the criteria before you go back to the numbers.
Exercise 3 — the technical summary
Criteria: summarise the proposal's detection approach for a reviewer at director level, accurately, in under 100 words.
The proposed module processes the transaction stream in five-second batches and applies the detection model to each batch independently. Because each batch is self-contained, coordinated transaction sequences are detected with full context, and latency is kept below the client's threshold. A sliding-window design is available as a fallback for lower-volume clients.
Reveal
It is fluent, under 100 words, and inverts the proposal. Independent five-second batches are exactly what lose context when a coordinated sequence is split across a boundary — the question the senior architect asks in Chapter 9 — and the sliding window is the main approach that fixes it, with the simpler design as the fallback, not the other way round. A reader who had not done the thinking would nod. The signal: does the reasoning in the summary produce the conclusion, or just sit next to it?
What these three have in common
None was caught by reading harder. Each was caught by holding the output against something written before it existed — the bullets, the criteria, the proposal — and noticing where it drifted. That is Step 7. The Dependent cannot do it, not for lack of intelligence, but because there was never anything to hold the output against.