// corrections
Corrections
Every factual error we have published and fixed, with the wrong claim quoted rather than paraphrased. A review site with no corrections page is not more accurate than one with a list this long. It is quieter about it.
Our editorial policy has said since September 2026 that when a published fact turns out to be wrong we fix it in public and sweep every instance rather than editing quietly. This is where that gets checked. If you have found one we have not listed, email [email protected].
16 corrections logged
14 found by our own re-checks
1 found by a reader
1 found by the company covered
That last split is the number worth watching, and it is the reason we publish it. A log where we found everything ourselves would mean our checks are good or that nobody is reading closely enough to catch us, and we cannot tell those apart from the inside.
What goes on this page
A correction is a claim you could have acted on that turned out to be false: a price, whether something runs on your phone, what a product can actually do, a ranking, a score, a quotation. Each entry says what was published, when we found it, where it appeared, what it says now, and how it was caught.
Two things deliberately do not appear here. Broken links, layout defects and tooling bugs are maintenance — they never told you anything untrue. And a score that moved is not a correction: a review is a live verdict rather than a record, so when a re-test changes a number the review itself is rewritten and carries a dated note saying what it used to score and why. How dated pieces are handled covers the rest.
Corrections are also marked on the pages themselves, dated, at the point where the error was. The wrong sentence is rewritten rather than deleted, and where the piece is a dated record the false line is struck through above its correction instead of being swept away.
The log
Newest first, by the date the error was established rather than the date the fix shipped.
-
Fable Forge’s entry said we had no commercial relationship while it held a sponsored placement
- We published
- That “no commercial relationship exists between Fable Forge’s developer and Arcanum”. The sentence came from the standard disclosure block added to every directory entry on 10 September 2026. The block knew about free access but not about sponsored placements, so it printed that line on Fable Forge’s entry while Fable Forge held a paid placement in our directory.
- It now says
- The placement itself was labeled on the entry and in the directory throughout, and Fable Forge carries no score. The disclosure block now reads the placement dates: while a platform holds a sponsored placement or has bought design work from us, the block names that relationship and never prints the “no commercial relationship” sentence.
- Where it appeared
- 1 page · live from 10 September 2026 to 14 September 2026 Fable Forge entry
- How it was caught
- Found by our own re-check Found while adding a disclosure for paid design work to another entry. The block was built from one template for every platform, and none of our checks compared its wording with the placement dates.
-
Our directory page still said no platform had a commercial relationship with us
- We published
- That “no platform has a commercial relationship with Arcanum”, stated in the closing section of our AI RPG directory page. It was true when the page went up on 28 June 2026. It stopped being true on 22 July, when a platform we cover bought an engine audit from our design service, and again on 3 August, when a listed platform bought a sponsored placement. The sentence was not updated when that happened.
- It now says
- Both relationships were disclosed in our coverage of each platform from the day they began, and neither platform carries a score. The directory page now says what is true: no game or platform can pay for a listing, a review or a rating; Arcanum sells clearly labeled sponsored placements, only to platforms already listed, and design services; and any commercial relationship or free access is disclosed on the pieces it affects.
- Where it appeared
- 1 sentence · live from 28 June 2026 to 14 September 2026 AI RPG directory
- How it was caught
- Found by our own re-check Found while re-reading every claim the directory page makes about Arcanum itself, during a sweep of how that page presents our own games.
-
We described Voyage’s free tier as it was in April
- We published
- That Voyage is free to play on a “starter usage pool, free daily usage claims” — the model Latitude announced for its expanded beta in April 2026, stated as current on seven pages from 31 August, five days after the open beta replaced it. One of those pages also told groups that four players was the ceiling for Voyage multiplayer.
- It now says
- Per Latitude’s own open-beta release notes and press release of 26 August 2026: free players get a limited usage demo for single-player and can play public servers as much as they want, with parties of up to eight, and paid Voyage Unlimited plans include unlimited play on core features plus a usage pool for premium ones such as voice narration. Private co-op still takes up to four invited players, with only the host using usage. All seven pages now say so.
- Where it appeared
- 7 pages · live from 31 August 2026 to 11 September 2026 Voyage entry, Voyage review, Voyage vs AI Dungeon, AI RPG apps, What an AI RPG costs, Best AI RPG for a long campaign, AI RPG with friends
- How it was caught
- Found by our own re-check Found while researching a Voyage guide. Latitude publishes dated release notes on voyage.io, and its April and August posts describe two different free tiers; for the 31 August review we re-checked the store listings and the pricing ladder, and never the release notes.
-
We said Voyage could not be installed on an iPhone
- We published
- That Voyage, the highest-scored platform in the directory, was Android-only and had no App Store listing at all — stated as fact on eleven pages and built into the argument of one of them, whose section heading read “The Best Platform in the Directory Is Android-Only”.
- It now says
- Voyage AI RPG Platform, published by Latitude Inc., has been on the App Store since 26 August 2026 — five days before we published the claim. Every line that carried the error is rewritten rather than deleted, the section built on it says so, and the dated August note keeps the false line struck through above its correction.
- Where it appeared
- 11 pages · live from 25 August 2026 to 2 September 2026 AI RPG apps, Voyage review, Voyage vs AI Dungeon, August 2026 note, Best AI roleplay platforms, Voyage entry
- How it was caught
- Found by our own re-check We read the Google Play listing and Latitude’s own announcements and never opened the App Store listing itself — the exact check we had written up as reusable one day earlier, on another platform. It took one request to settle. Three sentences carrying the same error survived that sweep — two in the Voyage review and one FAQ answer in the long-campaign guide — and were found and rewritten on 11 September 2026.
-
Nine rankings still described the board as it stood before Voyage
- We published
- Nine ordinal claims left standing by a scoring pass: Friends & Fables called “our top-rated AI RPG platform” when it is third, Tabled called “joint-best Memory” when it is a quarter-point behind, “the only two platforms scoring above 3 on both” when there are four, and the methodology page saying the highest-rated platform holds no axis alone.
- It now says
- All nine are rewritten against the benchmark data rather than by hand, and comparison claims are scoped to the set the page actually covers so the next entry does not break them again.
- Where it appeared
- 9 claims · live from 31 August 2026 to 1 September 2026 Best AI RPG for a long campaign, Methodology, Deep Realms vs Friends & Fables
- How it was caught
- Found by our own re-check None of our three automated rating checks could see these: every one names a rank rather than a score, and all the numbers on the pages were correct. Four were inside structured data, where an error is invisible in a browser.
-
Eight claims described score ceilings that had already moved
- We published
- Eight published sentences said the Memory axis caps at 4 with Hidden Door on top, after Voyage took it outright at 4.25; that Player Agency stood “above every structured platform”, after Tabled matched it; and that “the most restrictive platforms are among the least fair”, after one platform scored 1 on agency and 5 on fairness.
- It now says
- Each is rewritten to the current state of the board, with the mechanism named rather than the sentence merely renumbered. Two of the eight were FAQ answers, so the error was in structured data as well as in prose.
- Where it appeared
- 8 claims · live from 31 August 2026 to 1 September 2026 AI Dungeon vs Hidden Door, AI Dungeon alternatives
- How it was caught
- Found by our own re-check Found by re-deriving every claim from the benchmark data and asserting it back against the corpus. A sentence that names a ceiling contains no score, so no rating check can reach it.
-
Tidefall’s Unlimited tier was published at $17 a month
- We published
- That Tidefall Unlimited costs $17/month. The figure came from a July first look and was carried forward onto the directory entry, the review and the cost guide.
- It now says
- Unlimited is $14.99/month, billed monthly, with Paddle as merchant of record. All three pages now cite the pricing page as of 28 August 2026 rather than repeating a dated July figure. The $0 ninety-minute demo and the $299 founder tier were confirmed unchanged.
- Where it appeared
- 3 pages · live from 22 July 2026 to 28 August 2026 Tidefall entry, Tidefall review, What AI RPGs cost
- How it was caught
- Found by our own re-check Found by re-checking our own pages against the vendor’s live pricing page. It is the reason we now check a price against the checkout rather than the pricing page, and date every figure we publish.
-
Our own games were carrying star ratings we had awarded ourselves
- We published
- The Starter Kit page rendered ★4.9, ★4.8 and ★4.6 against three Arcanum Originals — our own games, which we have always said we never score — alongside two pre-v1.0 scores for third-party platforms that no longer matched the site.
- It now says
- The ratings are gone from those cards. Arcanum Originals are not scored anywhere on the site; the reason is on the methodology page.
- Where it appeared
- 1 page · live from 14 July 2026 to 25 August 2026 Starter Kit
- How it was caught
- Found by our own re-check Found by eye during an unrelated sweep, not by tooling. All three of our rating checks read content files and this was hardcoded in a page template, which is a blind spot we have documented rather than closed.
-
We said Tabled’s iOS app was still in App Store review
- We published
- That Tabled’s iOS app was “in App Store review”, on both the directory entry and the review.
- It now says
- It cleared review and listed on 15 August 2026 — six days before we published the sentence saying it had not.
- Where it appeared
- 2 pages · live from 21 August 2026 to 25 August 2026 Tabled entry, Tabled review
- How it was caught
- Found by our own re-check Turned up by checking all eighteen platforms against the App Store and Google Play for a piece about phone apps. Three separate corrections came out of that one afternoon of checking.
-
We described Janitor AI as browser-only on a phone
- We published
- That mobile play on Janitor AI meant the browser, with no app of its own.
- It now says
- Janitor AI ships its own iOS and Android apps. The iOS app has been listed since November 2025 — six months before we published the claim.
- Where it appeared
- 1 page · live from 22 June 2026 to 25 August 2026 Janitor AI entry
- How it was caught
- Found by our own re-check Same store-by-store check as the Tabled correction above.
-
A model table we published in the present tense had quietly stopped being true
- We published
- That the second model on each of Voyage’s three unadvertised Shadow tiers was Raven, at 32k, 64k and 128k of context.
- It now says
- That slot now reads GLM 5.2 at 16k, 32k and 64k — half the context at every tier, with prices and memory allowances unchanged. The directory entry is corrected and dated; the interview that recorded the old figures keeps them verbatim with a dated note, because it is a record of a conversation held on 10 August.
- Where it appeared
- 2 pages · live from 10 August 2026 to 24 August 2026 Voyage entry, Nick Walton interview
- How it was caught
- Found by our own re-check The change was the platform’s, not ours, and it was made without announcement. Our error was publishing a live figure in the present tense without a date on it, which is what turns a silent change into a false claim on our site.
-
We presented AI Dungeon’s context ceilings as flat per-tier guarantees
- We published
- That an AI Dungeon subscription tier gets you a stated context window — 64k on the top plan, for instance — with three pages treating the number on the pricing page as a promise that applied to every model.
- It now says
- Context is allocated per model, not per plan. The figure on the pricing page is a ceiling several models never reach and one exceeds, and the site now carries the full per-model table across all five tiers instead of tier-level claims.
- Where it appeared
- 3 pages · live from 17 June 2026 to 10 August 2026 Best AI Dungeon model, AI Dungeon entry, AI Dungeon review
- How it was caught
- Found by a reader A reader called our AI Dungeon review “completely inaccurate” and declined to say why. We measured it ourselves rather than guessing, and the per-model table that came out of it is the best original data we published that month.
-
Our AI Dungeon model roster was stale, and one figure in it was invented
- We published
- A four-model roster long out of date, base-model attributions we could not verify, and a parameter count — “Nova (70B)” — that had no source at all.
- It now says
- The roster is 8 models on the free tier and 18 on the top tier, verified against Latitude’s in-app model list. The unverifiable attributions and the invented parameter count are gone, and model rosters now carry the date they were checked because they change without announcement.
- Where it appeared
- 1 page · live from 1 July 2026 to 10 August 2026 AI Dungeon entry
- How it was caught
- Found by our own re-check Found while measuring the context question above. It is the worst kind of error on this site — not a fact that went stale, but a number that was never true — and it is why nothing here is written from a marketing page or a model-generated summary.
-
Three quotations in the Nick Walton interview were wrong
- We published
- That the turning-point model was GPT-5; that Voyage’s open beta was “released in a few weeks”; and that the 1.0 release was coming “in early 2027”. A graphics analogy was also presented as his phrasing when it was ours.
- It now says
- The model was GPT-3.5, which moves the capability unlock to late 2022 and changes the surrounding analysis — the five-year gap becomes design iteration rather than waiting for a good enough model. Both quotations are corrected to what he said, the analogy is attributed to us, and a fuller answer he sent afterwards is marked as a post-publication addition rather than folded into the live account.
- Where it appeared
- 2 pages · live from 10 August 2026 to 10 August 2026 Nick Walton interview, Voyage vs AI Dungeon
- How it was caught
- Found by the company covered Nick Walton read the published piece and sent seven notes on it. Two of the quotations were also mirrored into structured data on another page, so they were swept together.
-
We published a Voyage pricing table that did not exist
- We published
- A “Standard / Premium / Ultimate at $15 / $30 / $50” tier table for Voyage, assembled from beta materials rather than read off a real account, and repeated across the directory entry, the review and two comparisons.
- It now says
- Voyage is not sold separately. One Latitude subscription covers both Voyage and AI Dungeon — Journey $14.99, Legend $29.99, Mythic $49.99, Ultimate $99.99 — and the FAQ asking which is cheaper now answers that neither is, because it is one subscription.
- Where it appeared
- 5 pages · live from 17 June 2026 to 10 August 2026 Voyage entry, Voyage review, Voyage vs AI Dungeon, AI Dungeon review
- How it was caught
- Found by our own re-check Caught by comparing our own two pages against each other: the tier ladder we had published for Voyage matched AI Dungeon’s exactly, and so did the model roster. The site had been wrong about this for a month, and the previous week’s attempt to fix it was wrong too.
-
We said the hidden Shadow plans covered Voyage only
- We published
- That the unadvertised Shadow tiers — Banshee $246.66, Reaper $496.66, Apocalypse $996.66 — were exclusive to Voyage, stated in six places including a “the one asymmetry is at the top” framing that a whole comparison rested on.
- It now says
- They are part of the same single-subscription ladder and cover both Latitude products. The asymmetry framing is removed rather than reworded, and the AI Dungeon entry now lists the Shadow plans it had correctly omitted under the wrong fact.
- Where it appeared
- 6 places · live from 10 August 2026 to 10 August 2026 Voyage vs AI Dungeon, AI Dungeon entry, Voyage entry
- How it was caught
- Found by our own re-check Found the same day as the pricing table above, by continuing to check rather than stopping at the first correction.
Before this log
The log starts with the corrections we can document properly. Anything corrected before then was fixed on the page at the time and is not reconstructed here, because an entry assembled from memory would be exactly the kind of claim this page exists to rule out. The record runs forward from here.
If we got something wrong
Email [email protected]. You do not need to prove it — telling us where to look is enough, and we would rather check ten things that turn out to be right than miss one that isn't.
Disagreeing with a verdict isn't a correction and we won't pretend otherwise. But if the reasoning behind a score doesn't survive contact with your experience of the same product, the method is published so you can say exactly where it goes wrong.