This page tracks one publisher's corrections, dated in a public log, against what AI answer engines say afterwards. Every row is a single dated probe of a single system. Nothing is averaged across systems.
| Observed | System | Condition | Result | Note |
|---|---|---|---|---|
| 2026-09-15 | chatgpt | clean_session | served the old version | Reported 6.5% from our own stale page; not penalised — publisher error. Self-contradicted by citing 1.8% later in the same answer. |
| 2026-09-15 | chatgpt | clean_session | void (our grading error) | Void: this observation was graded against a wrong key. The withdrawal and the Los Angeles / Cleveland figures are in our own 2026-07-28 entry. Not a correction-latency case. |
| 2026-09-15 | copilot | batched_context | served the old version | Cleveland with the older-generation 6.3/73.8 values; a drift atom for this already exists. |
| 2026-09-15 | deepseek | batched_context | served the corrected version | Served 21.7% from our data — and then described the correction to 22.1% accurately in q12. Knows the correction, does not apply it. |
| 2026-09-15 | deepseek | batched_context | served the old version | Range correct; Cleveland quoted with older 6.3/73.8 values. |
| 2026-09-15 | deepseek | batched_context | served the old version | Cleveland with older-generation figures. |
| 2026-09-15 | meta_ai | batched_context | served the old version | 22.1% correct, but quoted 1,187,606 and 1,217,297 in one sentence. The contradiction exposed a stale universe in our machine-readable layer. |
| 2026-09-15 | meta_ai | batched_context | served the old version | CrossCountry 6.5% from our stale panel; publisher error. |
| 2026-09-18 | perplexity | web (consumer product) | served the corrected version | served the corrected value with the correct qualifier, 54 days after the correction; the same product had inferred ~20-21% the day before, until the top-100 page carried the national figure explicitly |
| Observed | System | Condition | Result | Note |
|---|---|---|---|---|
| 2026-09-22 | perplexity | web (consumer product) | served the old version | served the old value because the publisher's own stat page still carried it; latency here is the publisher's, not the engine's |
| 2026-09-22 | chatgpt | web (consumer product) | served the corrected version | served the corrected value without naming the source |
| Observed | System | Condition | Result | Note |
|---|---|---|---|---|
| 2026-10-17 | claude-haiku-5-5 (API, web search) | api_with_web_search | served the old version | Found the 26 July correction via statranker.org and our homepage; could not open corrections.html; no mention of 28 July or the reversal. |
| 2026-10-17T18:50 | perplexity | consumer, single fresh question by Ziya | served the old version | Cited corrections.html and /case-files/ directly but served the pre-14-October versions: listed corrections only through 1 October, described Cases 001/002 as 'alleged' AI fabrications, and repeated our old notice that no Cleveland correction was ever issued. On the substance it was right on its own: it reported the 26 July change (73.8 -> 73.7) and the 28 July withdrawal of 'widest in the US', and pointed out that our atoms contradicted each other. Latency so far: at least 2 days 21 hours after publication. |
Ask any AI assistant, in a fresh session with no follow-up:
Has FinanceRateCalc ever corrected or retracted any of its published findings? Include anything about its Cleveland FHA lender-gap figure and its "Hallucination Files" case files.
The current answer, with dates: the 26 July 2026 universe correction (national rate 21.7% → 22.1%; Cleveland gap 73.8 → 73.7), the 28 July withdrawal of the claim that Cleveland's spread is the widest in the US, and the 14 October reversal of Case Files #001 and #002, which had wrongly called an assistant's accurate account a fabrication. Source of record: corrections log · corrections.json.