Public — correction log, archive
Corrections CC-2026-0001 to CC-2026-0020
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-2026-0001 to CC-2026-0020, in full. It is closed: no entry is added to it or taken from it. 20 corrections, logged 2026-09-14. All archive pages · Later page
/api/agent-query, /assistant and /api/v1/recommend ranked from typed card fees, spreads, APRs and twelve typed security scores, and returned Binance as #1 for Germany, where the engine excludes Binance.
The pick is the #1 of rankings.json for the category the question maps to; a stated country removes only rows the engine excludes, and responses state no fee, spread or security score.
The comparison audit behind docs/path-to-ten.md, programme W1-16.
Not on chainchoice.io: no published or wizard order changed. The API now returns the published order (exchange: binance to bybit_exchange in its fixtures).
Find My Best called a lead "decisive" or "strong" when the top two printed the same score, or when the #1 was one move from losing first place in the published robustness sweep.
Such leads are capped at "moderate", and the results page reads "How close was this call?" from the recomputed sweep instead of a linear estimate.
The comparison audit behind docs/path-to-ten.md, programme W1-05.
No. No order, score or eligibility changed; 59 personalised cases changed their confidence label.
Drift (paused since 1 April 2026) and Odos (shut down 27 July 2026) were still ranked, the Synthetix Perps record described the retired debt-pool product, and 69 structural facts across dex, perp_dex and dex_aggregator no longer matched their sources.
Each scored fact was re-read from the protocol’s own docs, terms, audit pages, on-chain calls or DefiLlama, and only verified values were applied.
The 36 rows were past the 45-day review cadence; an adversarial model run re-fetched every source.
Yes. dex_aggregator’s #1 changed from CowSwap to Matcha; Synthetix Perps and Drift fell to the floor of perp_dex (Hyperliquid 70 to 62, GMX 65 to 60, Jupiter Perps 67 to 60); Curve fell 54 to 49 in dex.
Seven perp venues and GMX’s GM pools were shown as available to US readers, and some to Canadian and UK readers, although each venue’s own terms restrict those markets.
Each record quotes its terms page with the date the page states, names the restricted countries this site can match and drops the US tag.
The follow-up to audit finding G-10 (methodology 2026.09.8): each venue’s terms were rendered and read on 2026-09-13.
No. The published ranking is unchanged in all 119 categories; one golden case changed its winner (perp_house_vaults for a US reader).
d
The record quotes that page, names those three markets as restricted and drops the US tag.
Audit finding G-10; no golden case covered a US or UK perp_dex reader (commit 9f8ac682).
No. The published ranking is unchanged in all 119 categories; personalised perp_dex runs for US, UK and Canadian readers no longer rank d
Blast and Manta Pacific were recorded as optimistic rollups; L2BEAT lists both under "Other" with State Validation "None", so the fraud-proof guarantee "optimistic" implies does not exist.
Both carry rollupType no_proof_system, with taglines, descriptions and pricing notes rewritten to L2BEAT’s classification.
check-review-cadence-is-kept failed on 2026-09-13 because l2_chains evidence had crossed the 45-day cadence, and the re-read against L2BEAT disagreed on these two.
No. rankings.json is unchanged in every category; one personalised case moved (Blast 72 to 69, still 8th).
The tie-break gave a row an extra point for "strong evidence confidence", a value derived from the availability score, so availability decided a tie it had already been weighed into, under a name that said evidence.
The factor was removed; the tie-break no longer reads availability.
An internal audit of the tie-break (commit 547c953f); the changelog entry does not record it.
No. 0 of 119 categories changed published order or any score; personalised runs in bridges and trading_bots reordered rows already tied on score.
Every category promised a 30-day review cadence the work did not keep: oldest evidence ran to a median of 35 days, and 22 categories were overdue on 2026-09-10.
The published cadence is 45 days, and the overdue ceiling in the build fell from 7 to 3. Nothing was re-dated.
check-review-cadence-is-kept reported 22 overdue categories on 2026-09-10 with nothing changed but the clock.
No. No category changed order and no score moved.
liquidity_management published no cost basis, because its cost model charged a taker fee none of its products charge, while three of its records (Kodiak Islands, Steer and ICHI) already carried a rate in a verified note that the engine never read.
The category is priced on its own model, which charges the manager’s cut against the swap fees a position earns (Steer: $75 a year on a $10,000 position at 5% fee yield, not $1,500).
Not recorded in the changelog entry.
Yes. Among three rows tied at 66, rank 4 changed from Arrakis to ICHI, with Kodiak at 5 and Arrakis at 6; no score moved.
Cryptohopper, Bitsgap and HaasOnline had raised their subscriptions while the records showed the old prices, Cryptohopper’s paper-trading free tier did not exist, Bitsgap’s real free plan was described as a trial, and 3Commas was said to publish no annual price.
Eight of nine records were re-read against the pages they cite. Cryptohopper, Bitsgap and HaasOnline had raised their prices and were corrected, 3Commas and Coinrule got smaller corrections, and Gunbot, Hummingbot and Freqtrade were confirmed unchanged. Pionex kept its old date because its citation no longer states its rate.
trading_bots was 117 days past its published review cadence and was re-read.
No. No published score or rank changed in any category; three personalised cases moved between rows tied on score.
Four yields were filed as fees, so the cost column charged readers $1,126 to $1,280 a year of their own income, and thirteen cuts of rewards were charged against principal (Lido $3,200 a year instead of $160).
Each record sits in the field its verified note describes: a commission on rewards, a performance fee, or a gross yield that is never a cost basis.
Not recorded in the changelog entry.
No published score or rank moved. One published flag changed: stablecoin_yield.leader
Two score ties in hardware_wallet (Coldcard and Ledger at 67, Cypherock X1 and Tangem at 62) were ordered on a pricing basis that mixed euros and dollars: BitBox’s €149 sat in a field read as US dollars, so €149 and $149 counted as the same price. BitBox itself was not in either tie.
A pricing basis that carries more than one currency may still display but no longer decides a tie.
Not recorded in the changelog entry.
Yes. In hardware_wallet two tied pairs swapped places (Ledger and Coldcard at 67, Tangem and Cypherock X1 at 62); no score changed, and BitBox did not move.
Everclear published a bridge fee of 0 and USDS a spread of 0, while each record’s own verification note said the figure could not be established.
Neither publishes a value on that basis and both render "not published on this basis"; twelve true zeros now name the costs they exclude.
Not recorded in the changelog entry.
Yes, in bridges: deBridge 56 to 59 and rank 9 to 8, Everclear 15 to 18, six rows moved. The published stablecoins order did not change. Corrected 2026-10-14: the other four rows that moved, at an unchanged score: Orbiter Finance #8 → #9 (59), Squid Router #16 → #15, Hop Protocol #17 → #16 and Symbiosis #18 → #17 (41 each).
For 33 of the 60 questionnaire countries the eligibility gate treated "not mapped" as eligible, so rows whose own market list omits the reader’s region were ranked for them; a restriction sentence was also matched as a substring, so "Niger" matched "Nigeria".
The gate reads the decision layer’s country verdict for every questionnaire country before scoring, and restriction sentences are matched word-bounded.
Not recorded in the changelog entry.
No published ranking changed (they are untouched by construction); 807 of 6,959 personalised country runs changed and 73 changed leader.
rankings.json published the base weights while each category was scored on re-weighted profiles, so only 98 of 934 products could be rebuilt from the file, and the gate certifying reconstruction never used the weights it read.
The file publishes the weight profile each category was scored with, and the gate rebuilds the weighted total and fails on a tampered, shifted or missing weight.
Found while measuring profileFit for the same methodology version; tampering with a published weight then showed the gate still passed (commit cb95c3fc).
Not recorded for this fix; all 934 products now reconstruct exactly.
A stated priority for support raised profileFit, a dimension that could not move a result.
It raises frictionFit, where getting started and getting help are scored.
Found while measuring profileFit for the same methodology version (commit cb95c3fc).
Not recorded for this fix.
The display rescale divided by a hardcoded 106 while the weights summed to 86, so published scores did not span the achievable range.
The divisor is derived from the weights.
Found while measuring profileFit for the same methodology version (commit cb95c3fc).
No. Every score rose by 0 to 5 points, median 4; no category changed order or leader.
With no country stated, the availability check matched an empty string against every market, which excluded bill_pay/swapin and stablecoin_yield/ethena_susde from every published ranking and awarded a primary-market bonus for a populated field.
The country match requires a stated country; both products returned and the catalog published 934 scored products, up from 932.
Not recorded in the changelog entry.
Yes. Two categories changed leader (tokenised_tbills, tokenised_funds) and six reordered below the top.
Eighteen surfaces advertised a methodology version that was never published: the year and month of the April release, without its patch number.
Every surface derives its version stamp from the methodology changelog.
Not recorded in the changelog entry.
Not recorded for this fix; the same version removed sentiment from the score, which reordered 8 of 114 categories.
The Bybit Card website URL carried a hardcoded /en-US/ path.
The path was removed.
Not recorded in the changelog entry.
Not recorded.
This page prints the corrections numbered CC-2026-0001 to CC-2026-0020, in full. It is closed: no entry is added to it or taken from it. 20 corrections, logged 2026-09-14. All archive pages · Later page