| Mechanics |
|---|
| cutoff date | End of the training corpus — events after it were never seen. | trained through 2025-01 |
| not a switch | Knowledge fades near the boundary; the last months are thin. | late-2024 events half-remembered |
| counterfactual confidence | The model invents plausible post-cutoff answers without hedging. | fabricated 'new' API params |
| per-model variance | Cutoffs differ by model and refresh — check the model card, not vibes. | one family: 2023-04 → 2024-04 → … |
| Mitigations |
|---|
| RAG | Inject current documents at inference — the standard fix. | search → top-k → prompt |
| tool use | Let the model fetch live state (web, APIs) instead of recalling it. | web_search(query) |
| date the prompt | Tell the model today's date so it reasons about recency explicitly. | Today is 2026-09-17. |
| scoped questions | Ask about durable facts, not moving targets. | 'how does X work' vs 'latest X version' |
| What breaks without mitigation |
|---|
| library versions | Post-cutoff releases: invented flags, removed APIs 'still there'. | framework v3 APIs hallucinated |
| prices & plans | Anything that changes quarterly — model pricing especially. | stale per-token prices |
| current events | People, products, scores after the horizon. | 'who won…' → fabricated |
| its own context | The model may not know its own cutoff — ask it, get a guess. | 'I believe my cutoff is…' |