Public — correction log, archive
Corrections CC-2026-0061 to CC-2026-0080
Factual errors, misapplied scoring, stale data and claims the site could not back. Each entry says what was wrong, what was fixed, how it was caught and whether the published order moved. An entry whose logged date is later than its corrected date was copied here from the methodology or data changelog that first published the fix. No entry is ever deleted.
This page prints the corrections numbered CC-2026-0061 to CC-2026-0080, in full. It is closed: no entry is added to it or taken from it. 19 corrections, logged 2026-10-11 to 2026-10-28. Earlier page · All archive pages · Later page
OKX's EU spot-only rate (0.20% taker) was stored as its rate for every reader, so a reader in the UK or Asia was priced at the EU entity's fee; the exchange score read the dearest rate any entity charges, so readers were ranked on prices other readers pay; ether.fi Cash was made eligible from a country list whose page now returns 404.
The EU, US and UK rates of Bitstamp, Kraken, Coinbase, Bitvavo, Gemini, OKX and Bybit were read from their own fee pages on 28 October; each reader is now priced and scored on the rate of the entity serving their country, and a rate that could not be read is shown as "regional rate not read". OKX's EU rate applies to EEA readers only. Coinbase Staking prints the 35% in-app commission beside cb
A review of the 28 October price corrections found the dearest-rate rule and OKX's base rate; the owner ruled the same day that each reader is scored by their own entity's rate.
Yes. Exchange: Bybit stays #1; Bitstamp #2; OKX falls from #2 to #7 (no base rate). Readers in the EEA and the United States see Bitstamp first, scored on their entities' fees.
Evidence cells that a re-read could not verify (a 403 bot wall, a page gone) were published with an empty quote, no source and the re-read’s date, and their rows took that date: Revolut Crypto and Coinbase (Simple Buy), every page behind a bot wall on 28 October, were shown as verified that day, and the review dates of the categories above were too fresh. Raydium’s note named a June 2026 exploit its own incident list does not, and Balancer’s gave a loss figure its own post-mortem does not.
Each such cell keeps its last real read (quote, source and date) and says when a re-read could not verify it, a cell never read carries no date, and a row keeps its last full read. Raydium’s unsourced exploit is removed and Balancer’s figure is its post-mortem’s $121.1M.
An adversarial review of the 28 October dex and simple_buy re-read found walled cells dated 28 October over empty quotes.
No. No published score or rank changed; only read dates, the receipts behind 74 cells and two notes changed.
Hop Protocol stayed in the bridges ranking with a link to hop.exchange for 23 days after our own re-read found that the domain served an unrelated gambling page and that Hop's app and route service did not answer; on 23 October it still served that page. Synapse's record told readers to treat it as unavailable, which a read of its bridge contract on the same day did not bear out.
Hop Protocol carries a sunset date of 23 October 2026: it is kept for reference with its criteria block at zero, links to its code repository, and no page draws a Visit for it. Synapse's note now states what was read on 23 October, and a build check fails whenever a row that says it is unavailable has no sunset date or any link points at a host that served gambling or unrelated content.
A check of the live bridges page on 23 October 2026 followed Hop Protocol's Visit link to the gambling page.
No. The published bridges order and every published score are unchanged: Hop Protocol was already last (#18, 41) and stays there as a sunset row. Hop Protocol is removed from the bridges list on the country pages for Australia, Austria, Belgium, Canada, Finland, France, Germany, Hong Kong, Indonesia, Ireland, Italy, Japan, the Netherlands, Norway, the Philippines, Poland, Portugal, Singapore, South Korea, Spain, Sweden, Switzerland, Thailand, the United Arab Emirates, the United Kingdom, the United States and Vietnam, which list only products that operate.
Polygon zk
The record carries rollupType no_proof_system, as Blast and Manta Pacific have since 2026-09-13, and its notes say so.
A re-read of L2BEAT’s scaling summary API on 2026-10-21 (round-7 data freshness).
No. The published l2_chains order and every score are unchanged; in Find my best with a fast exit asked, eligible rows fall from 7 to 6 and Polygon zk
Correction CC-2026-0071 said that a gate reads every sentence of the site's process copy. It reads the trust logs, the claim types, the jurisdiction texts and the methodology specs, as string literals only. The rewrite that entry described also left role nouns for model work on live pages: the retired-category notes on six /family pages, the institutional execution FAQ and its structured data, a /methodology sentence, two guides, and about forty sentences in provider notes, among them research notes never meant for readers (“PUBLICATION GUARDRAILS FOR THE EDITOR”, a working file's name, “to me today”).
Each of those sentences now says a model run did the work, or states the fact without naming an actor, and the research notes were rewritten as statements for the reader. Widening that gate to every sentence the site renders is open work, recorded under CT-EDIT-008 in the master audit.
A re-audit of CT-EDIT-008 on 2026-10-13 read the live family, provider, FAQ and guide pages after CC-2026-0071 went live.
No. No published score or rank changed; the wording of the pages changed.
The pages and files that state our licence disagreed on what it covers: llms.txt granted "terms equivalent to CC BY 4.0" for editorial content, llms-full.txt called the datasets "functionally public domain", robots.txt licensed "ChainChoice's content", /terms named four files and the Impressum reserved every right to everything. Seventeen of the 23 files at the top of /data/ and all 14 KPI history files carried no licence, and answers.json carried the licence URL where every other file carries its id.
One scope statement, kept in src/lib/dataLicence.ts, now appears word for word wherever the licence is stated, in all six locales and in the signed Impressum (v5.2026.10.13): CC BY 4.0 covers the data (every file under /data/, /datasets/ and /methodology/spec/, the answers and the embed payload), not the site's other pages, and robots.txt adds that what was published under its earlier notice stays licensed as published. Every data file carries the licence id, third-party figures carry a note naming their source's terms, and check-dataset-has-licence reads the files the build ships.
Reading every licence statement against owner decision O-05 for the master audit's data-licence rows (SEO-AI-002, ST-DATA-010) on 2026-10-13.
No. No published score or rank changed.
The signed German affiliate disclosure (v2.2026.08.19, served at /api/v1?_route=affiliate-disclosure) said that every recommendation produces a signed decision receipt and that the guarantee can be checked in public source code. No recommendation on the portal emits a receipt (the Impressum, this disclosure and Octopus do), the source code is not public, and the five factors it said the ranking uses are not the published methodology's dimensions.
Version v3.2026.10.13 names what a reader can check: every published order is in /data/rankings.json and can be recomputed from that file alone, with its SHA-256 in the page footer; the two build checks that keep payment out of ranking and search are named, their results are in /data/build-checks.json and on /methodology, and it says the source code is private. It points to the partner list by its address and is served with a new signed receipt.
Reading every page and file that states a licence or a guarantee for the master audit's data-licence row (ST-DATA-010) on 2026-10-13.
No. No published score or rank changed; the text of the disclosure changed.
The CLARITY Act hub and our essay on the bill expanded its name as "Consolidated Lewis-Robust Asset Reform Implementation Title", which no bill carries (it is the Digital Asset Market Clarity Act of 2025, H.R. 3633), and said it was still in the House Financial Services Committee with House passage likely in 2026, when the House had passed it on 17 July 2025, 294–134. After the Senate rejected cloture on 15 September 2026, 49–50, the hub still listed both votes as "TBD 2026" and the essay, its page head and the readiness index still wrote "when it passes".
The hub, both essays, their page heads and /llms.txt give the bill's name and its dated record (House 294–134 on 17 July 2025, Senate Agriculture 12–11 on 29 January 2026, Senate Banking 15–9 on 14 May 2026, cloture rejected 49–50 on 15 September 2026; not law as of 13 October 2026), and write what it would change as conditional on passage. A test fails if the invented name or an undated "when the CLARITY Act passes" returns.
Rewriting /llms.txt for the master audit on 2026-10-13, which checked the name against the bill and found the same text on two live pages; the status was read that day at Latham & Watkins' US Crypto Policy Tracker and in Troutman Pepper Locke's note of 25 September 2026.
No. No published score or rank changed; the readiness index's scores are unchanged.
Our methodology changelog, this corrections log, the data changelog, the claim types on /methodology and several criterion texts described how the research is done as if people did it: "a three-lens panel and a judge", "an independent verifier", "an adversarial reconciler", "an independent skeptic", "the scoping analysts", "the research team" and "An editor read it at a named source". The reading, verifying and judging are done by language-model runs, as /methodology/evidence-process says, and the roster is two managing directors and six scripts.
Each of those 87 sentences now says it was a model run, with its facts, dates and quotes unchanged. A gate reads every sentence of the site's process copy and fails one that names a team, panel, judge, editor, verifier, reviewer, analyst, researcher, reconciler or skeptic without saying it was a model run or naming a person on the roster; the 32 uses that are a category's own vocabulary (a bug-bounty platform's users, a protocol's verifying contracts, a provider's own staff) are listed by name with the reason.
A refutation pass on the master audit's row for invented roles (CT-EDIT-008) on 2026-10-13 read the live pages, where the roster check had read only the roster.
No. No published score or rank changed; the wording of the record changed.
Pages named winners the published ranking does not: the US exchange guide called Coinbase "the most regulator-aligned choice" and Gemini "the best NY-native pick" (published #6 and #11 of 14) while the US country page named OKX, and the German guide and its FAQ structured data called Bitpanda, which the exchange ranking does not carry, "the strongest match" while the German country page named Bybit. A comparison guide, four reviews, an EU guide ("Bitvavo wins on fees", "the cheapest MiCA-licensed exchange"), two cards, a catalogue note and the assistant ("USDC is usually the cleaner broad default", published #2) did the same.
The US and German guides open with the answer their country pages give, generated at build by the same functions; the other passages say what each product fits. A gate reads every sentence on the site and fails on one that crowns a ranked product.
Checking 30 live head-to-head pages against rankings.json for the master audit on 2026-10-11, then searching the rest of the site for the same claim.
No. No published score or rank changed; the pages now repeat it.
The consent bar on every first visit cited the TTDSG, which has been the TDDDG since 14 May 2024. The compliance guide said the ePrivacy Directive was implemented "as TTDSG in Germany" with an ePrivacy Regulation "proposed at EU level"; the Commission withdrew that proposal in its 2025 work programme. It also said ChainChoice "uses cookieless Plausible analytics", leaving out Cloudflare Web Analytics and that both load only after a visitor accepts.
The bar names the TDDDG; the guide names the TDDDG with its former name and date, calls the Regulation withdrawn, and lists both measurement scripts and the consent they wait for. A test fails on the old Act name used as current and on the Regulation called pending.
Reading the privacy rows of the master audit on 2026-10-11.
No. No published score or rank changed.
The usage notes on /datasets said "Commercial bulk redistribution requires the partner API". The data files are published under CC BY 4.0, which allows commercial reuse with credit, and the files and their Dataset markup said so. The notes also said decision receipts can be verified against a key-set endpoint that answers 404 (recommendation receipts are hashed, not signed), and that signed receipts are immutable once issued.
The notes now state the licence the files carry, CC BY 4.0 with credit to ChainChoice, read from the one constant every file uses; say which documents come with a signed receipt and where to check one; and say only that a signed receipt's id is derived from a hash of what it states. A test holds the notes to that.
Reading /datasets in a browser to check the 2026-10-11 fix to its entries found the same claims in the notes beneath them.
No. No published score or rank changed.
The /trust page, headed "What is true today, and what is not", listed as in production that a receipt "cannot be edited after write — write-once enforced at the database layer"; the database lets a receipt's owner update it, and ChainChoice writes outcome, moderation and account fields after the write. It said a receipt is "replayable"; nothing re-runs one, and a model-written answer would not come out the same. It said a drift engine "writes a successor receipt" when a watched figure moves; nothing writes one. It said database access is gated by row-level security policies, a claim withdrawn until every policy is reviewed. It promised a PDF committee brief for Q3, which was never built. The five translations of the page said the same. Separately, /datasets called the index of the country pages' receipt IDs a feed of "signed decision receipts" with a "signing key reference", linked a key set "for verifying every signed ChainChoice receipt" at an address that answers 404, and described the client and the OpenAPI document as generated, with retry and verification helpers neither has.
The page now lists what the receipt code does: only fields outside the content hash change after the write, the engine stamp identifies what produced an answer without making it reproducible, and drift is computed when a receipt is opened and not stored. The source archive, governance log and successor receipts are listed as started and not running, the PDF brief as not built, and the translations show the English page. /datasets describes each file as it is and lists no key set. Tests hold the page to the receipt claims /institutional retired on 2026-09-14 and to what the code writes, and every /datasets link to a file, route or page that exists.
Rewriting the partner API documentation on 2026-10-11 surfaced the same receipt claims on /trust that /institutional had dropped a month earlier.
No. No published score or rank changed.
The /docs page for the partner API said the separation between the ranking and our commercial relationships is "externally audited annually"; no outside party has ever audited it. It said "EU recommendations filter to MiCA-licensed venues" and that "US recommendations align with state registrations", where the API removes the rows the engine excludes for a stated country and does nothing by licence or state. It offered a free tier and a paid tier at $24,000 a year that ChainChoice does not sell, a partner dashboard and self-serve keys that do not exist, and documented the request and response retired on 2026-09-14 (a recommendation "for you", an effective-cost percentage). The vendored TypeScript client typed every response in that retired shape, and the OpenAPI document described a deletion response and error fields no endpoint returns.
The page now describes the API that runs: the pick from the published order, the country filter, the fields read and the fields accepted and ignored, what a receipt holds, the limits, and keys by agreement with no price list. The audit claim is gone and the page says no outside party has audited the firewall. The client and the OpenAPI document match the handlers, and a test holds the spec, the client and the page to the handlers that answer.
Checking whether the partner API had an OpenAPI document on 2026-10-11 led to the /docs page, which still printed the contract retired a month earlier.
No. No published score or rank changed.
Ten /best pages said their ranking runs on four "setup signals" that neither the ranking nor any quiz reads ("The wallet path ranks on the same four setup signals every time", "Criteria evaluated before recommending a wallet recommendation", with Primary goal, Network fit, Access mode and Confidence driver listed as the criteria); the same four appeared on those categories' review and category pages as "The engine would validate this provider against these inputs", and wallet reviews said they validate "ecosystem fit" and "access mode", which the wallet methodology does not score. Separately, 310 provider descriptions named their strongest and weakest criterion in broken case ("Strongest on can i export my own keys and use them elsewhere?", "validator Count and Nakamoto Coefficient"), six of them by a criterion's old name, and nineteen lines of category copy in seven categories lowered acronyms and names ("soc 2", "1m identical calls", "nakamoto").
The four-signal block is gone from every page that carried it; the criteria and weights the ranking does use are listed on each category's /best page. Every description names its criteria by their current names and by the site's one rule for a label in a sentence, a criterion written as a question quoted whole (“Can I export my own keys and use them elsewhere?”), and the category lines keep their acronyms and names. Where a verdict's firmness note names a question criterion, it quotes it too. The two research scripts that write those descriptions and the one that reads them back now call that rule, and tests fail a description, a script or a category line that lowers a label its own way.
Reading published descriptions for cut-off sentences on 2026-10-11 turned up a lowercase "i"; tracing what else lowered labels found the rest.
No. No published score or rank changed.
195 published notes stopped mid-word or mid-quotation. The licence and regulatory notes on provider pages were cut at exactly 400 characters when each category was built ("… Date Initial MSB Regis", "… as ou"), network-coverage notes at 155, and some evidence notes and descriptions elsewhere; a few notes also showed a stray backslash before a quotation mark.
Each note now ends at its last complete sentence, or, where the cut fell inside a long quotation, marks the cut with "…" and closes the quotation; no wording was added. The script that builds a category now trims a note at a sentence boundary instead of a character count, and a build check fails any published note or refusal that stops mid-sentence.
Rewriting catalog notes for readers on 2026-10-11 surfaced six cut-off strings; measuring the pattern found the rest.
No. No published score or rank changed.
/crypto-tax-tools-dach and the Blockpit review said Blockpit offers a "PWC review" and a "PwC partnership for higher-volume audit". This site's own re-read on 29 September found no mention of PwC on any of 33 Blockpit pages, only an unnamed claim of "cooperation with Big4 companies"; the same pages named the company "Blockpit GmbH", where its terms are issued by Blockpit AG, and said it has no free tier, where portfolio tracking is free. Separately, about 200 sentences in row watchouts and evidence notes were published in the research notes' own voice ("I raised it 8 to 9 on a fact the draft never found", "WARNING FOR THE NEXT READER", "Corrected during verification: the draft said …"), and six published strings stopped mid-sentence ("… and sits under th").
The PwC claim is replaced by Blockpit's own unnamed wording, and the entity and the free tier are corrected. The watchouts and notes say what was checked and what it showed, every figure, date and quote kept, and the truncated strings end at their last complete sentence. The research voice is now gone from every catalog row; the 19 places it remains are methodology criteria, listed by name, and the list may only shrink.
Rewriting the research notes for readers on 2026-10-11: Blockpit's own record said to delete the PwC strings, and nobody had.
No. No published score or rank changed.
This site's own availability notes were printed in the wizard's warnings and on the "best for" pages under "Provider states:", so a provider appeared to say what our research had written: VALR's read "Provider states: What I found is the standard sanctions representation". The notes, and several row summaries and category pages, were also published in the research notes' own voice: four Stablecoin Settlement summaries said "Corrected during verification: the draft said …", Proof of Personhood's headline evidence said "the draft understated how uneven it is", the Stablecoin Issuance guide called the page "an authority asset rather than a revenue line", and three bill-pay notes explained the site's source code.
An availability note is now attributed to ChainChoice ("ChainChoice's note:"), and the provider's own words inside it stay in quotation marks. The availability notes, the summaries and the category copy are rewritten for a reader, keeping every dated read and quote. A build check reads every catalog row, every category page and every methodology's criteria for that voice; the 152 places it remains, mostly row watchouts, are listed by name and the list may only shrink.
Verifying batch 109 on production on 2026-10-11: the tax-software catalog still said "the draft dropped".
No. No published score or rank changed.
The /health-check picker showed a one-line description nobody had sourced under 238 of the names it offers, 115 of them names this site has never reviewed, and several of the lines were regulatory claims: "Spanish CNMV-registered crypto exchange", "Swiss FINMA-licensed crypto bank", "Italian crypto exchange (OAM-registered VASP)", "Brazilian regulated crypto exchange", "German crypto banking (status: discontinued)". A licence is a fact a register confirms or denies, and none of these had been checked; the same lines also shipped in the page's code.
Every name the picker offers now carries the state the rest of the site gives it: ranked (with the category), considered and not ranked, archived, or "Not reviewed by ChainChoice", and an unreviewed name shows no description. The descriptions are gone from the code that ships, and the type no longer allows one. A build check holds each state against the published placements and the refusal index.
Audit row CT-CENS-007 (no orphan names), read against the live picker on 2026-10-11.
No. The health scan ranks nothing, and no published score or rank changed.
This page prints the corrections numbered CC-2026-0061 to CC-2026-0080, in full. It is closed: no entry is added to it or taken from it. 19 corrections, logged 2026-10-11 to 2026-10-28. Earlier page · All archive pages · Later page