Crypto Signal Performance Claims: 16-Provider Evidence Audit

Eleven dossiers stop at provider-published claims without a complete record

The largest posture covers 11 of 16 providers. This does not mean the claims are false. It means the frozen dossier did not establish the complete chain from original alert through eligible denominator, entry and exit execution, costs, losses, open positions, edits, and account-level outcome. The two partial-arithmetic rows show why arithmetic and evidence are separate: a multiplication or division can reproduce a displayed number while the underlying observations and realized portfolio result remain provider-published.

The simulation or third-party posture also prevents category error. A platform can offer a backtesting engine without claiming one platform-wide live return. A marketplace can display connected signal sources without independently owning every source result. A copied bot result can depend on strategy version, exchange, pair, fees, fill model, capital, and start date. The row-detail posture recognizes information gain without overstating it: visible wins, losses, skips, or predictions help scrutiny, but independent reproduction still needs stable definitions, timestamps, versions, and downloadable records.

The alphabetical matrix keeps every result boundary and selected proof gap visible

Each provider appears once, sorted by display name rather than claim size, source count, commercial status, or editorial posture. The result-boundary text is reproduced exactly from the frozen dossier. The missing-proof bullets are also exact source-record strings, selected by the provider-specific indexes in this generator. Selection narrows the broader dossier queue to performance-relevant requests; it does not rewrite the wording or imply that unselected identity, billing, security, or route questions were resolved.

The source count is the number of source assignments attached to the provider record. It is not a count of sources that independently corroborate the result boundary, because the frozen schema does not map each sentence to stable source IDs. A ten-source packet can contain primary pages, historical context, terms, registry material, and independent boundary sources serving different questions. It must not be translated into ten confirmations.

Dossier links open the broader provider evidence files. Coverage remains non-ranking and non-endorsing. A row that names a large percentage is not promoted above a row with no percentage, and a row with partial arithmetic is not declared more investable than a row with simulation evidence. The matrix is designed to preserve the decision boundary, not collapse it into a leaderboard.

A percentage can be arithmetically correct and still fail as an account return

Performance language uses the same percent sign for different objects. A post can describe the underlying asset’s price move, a leveraged move before costs, the return on one allocated slice, a summed target figure, a win rate, an average per signal, a monthly net figure, or a cumulative promotional total. Those quantities are not interchangeable. Before comparing two percentages, the reader needs the unit, observation window, eligible population, capital basis, leverage, target allocation, cost treatment, and status of every included trade.

The denominator decides what a win rate means

A basic win-rate expression is wins divided by eligible closed observations. The hard part is defining both terms. Does a partial first target count as a win if the remaining position later stops? Are break-even calls wins, losses, or exclusions? Are canceled alerts removed? Are duplicate updates counted? Are still-open trades excluded from both sides while their favorable target movement appears elsewhere? Are all pairs, products, free calls, paid calls, copied bots, and market regimes included? A headline cannot answer those questions by itself.

Changing exclusions changes the denominator even when the underlying calls do not change. A defensible record should publish stable signal IDs, original timestamps, eligibility rules fixed before outcomes are known, and a reconciliation showing wins, losses, break-even calls, cancellations, invalidations, duplicates, and open positions. The period must have a clear start and end. A rolling snapshot should state when an observation enters or leaves the window. Without those rules, two correctly calculated win rates can describe different populations.

Leverage multiplication measures a trade move, not the whole account

Multiplying a price move by leverage can reproduce a displayed trade-level percentage. For example, a three percent favorable move multiplied by ten gives thirty percent before fees, funding, slippage, liquidation constraints, and target allocation. That arithmetic does not mean the subscriber’s account gained thirty percent. If only a small share of capital was allocated, the account effect is smaller. If leverage applies to a futures position with adverse movement before the target, margin and liquidation risk matter. If the entry is a range rather than one fill, the starting price also changes the result.

A complete method should state position size as a share of capital, leverage, isolated or cross margin, entry-fill rule, stop execution, maximum concurrent exposure, re-entry and averaging treatment, and what happens when an alert cannot be filled. It should distinguish return on position margin, return on allocated capital, and return on the whole account. Reporting one while implying another is a measurement problem even when every displayed multiplication is numerically correct.

Partial exits prevent one target percentage from representing one realized outcome

Multi-target signals need an allocation rule. If part of a position exits at target one, another part at target two, and the remainder stops or stays open, the realized result is the weighted combination of those paths. Counting the furthest touched target as the outcome overstates what a subscriber following earlier partial exits could realize. Counting every touched target as a separate win can also expand the numerator without adding independent signals.

The record should fix target weights before the trade, preserve any stop movement, show the remaining position after each fill, and state whether percentages are summed across targets or weighted by actual allocation. Averaging entries and later edits require the same treatment. Without this, a target-hit post can document market movement while remaining insufficient for account-return reconstruction.

Fees, slippage, and funding belong inside the result

Entry and exit fees reduce every realized trade. Slippage can be material around volatile alerts, thin books, stop orders, and large subscriber bursts. Perpetual futures funding can add a cost or credit over time, while spreads and borrow costs matter in other products. Exchange tier, order type, region, pair, and execution latency can change those values. A performance figure that excludes costs should say so and should not be presented as net subscriber return.

Reproduction needs an explicit fill model: signal timestamp, exchange timestamp, reference venue, bid or ask convention, order type, allowed latency, partial fills, failed orders, and fallback when the quoted price never trades after publication. Costs should be applied trade by trade rather than subtracted as a vague disclaimer. Independent verification is not created merely by linking an exchange if the verifier cannot reconstruct the same fills.

Edits, deletions, and open trades can change both numerator and denominator

An edited entry, stop, target, leverage value, or timestamp can change the measured outcome after the market moves. A deleted losing call removes evidence from the apparent population. A forwarded result may preserve the result message without preserving the original alert. Open trades create another asymmetry: favorable movement may enter a promotional total while unresolved downside remains outside realized losses. A robust archive therefore needs original alerts, edit history, deletion or tombstone records, cancellation reasons, closure timestamps, and a declared report cutoff.

Immutability does not require a particular technology. It requires a preservation method that makes later changes detectable, such as exported platform identifiers, trusted archives, signed files, or versioned datasets with hashes. Corrections should remain visible rather than overwrite the prior record. The goal is an auditable chronology, not a claim that no human error occurred.

Drawdown exposes the path that cumulative profit hides

A cumulative or summed profit figure can hide the capital path needed to survive it. Drawdown measures decline from a prior equity peak and should specify whether equity includes open positions, how concurrent trades are combined, and whether the calculation uses account equity, closed balance, or a model portfolio. Maximum drawdown, recovery time, losing streak, and peak margin use can matter more to a subscriber than the sum of positive trade percentages.

Summing leveraged percentages across signals is especially weak as a portfolio measure because it can ignore capital reuse, overlap, sizing, and losses. A reconstructable account curve needs starting capital, allocation rules, concurrency, compounding, deposits and withdrawals, costs, and mark-to-market treatment. Without those inputs, a large total can describe signal-level movement without showing whether a real account could follow the sequence.

Backtests measure a model under assumptions, not live execution

A backtest can be useful when its rules, data, versions, and costs are visible. It still answers a hypothetical historical question. Selection of pairs, date ranges, parameters, and surviving strategies can introduce overfitting and survivorship bias. Look-ahead leakage, revised data, optimistic fills, omitted outages, and repeated tuning can raise historical results without improving future execution. A provider-wide claim should not be inferred from a tool that lets each user define a different strategy.

Useful evidence separates in-sample development from out-of-sample testing and later forward results. It preserves strategy and model versions, parameter changes, data sources, timestamps, fees, slippage, latency, failed orders, and inactive periods. For machine-learning or LLM-assisted predictions, the scoring code, prompt policy, model version, feature timing, threshold, and revision behavior also matter. A disclaimer that past results do not predict the future is necessary context, but it does not make an otherwise irreproducible backtest reproducible.

Claim language appears often, but overlapping words are not performance scores

The audit applies five literal, case-insensitive language tests only to the exact resultBoundary field. Ten of the 16 boundaries contain a percent sign. Eleven contain the defined accuracy, win, loss, or profitable-recommendation family. Nine contain profit, return, gain, PNL, or monthly-net language. Eight contain leverage, target-allocation, sizing, or partial-exit language. Two contain backtest or simulation language.

These counts overlap: one boundary can match several families, and one provider is still counted once within each family. The tests do not determine whether a claim is prominent, current, true, false, misleading, or material. A relevant concept can also be present with different wording and remain outside the match. The counts exist to describe the authored cohort consistently, not to produce an allegation rate, risk score, keyword rank, or estimate of the wider market.

Thirty-one selected proof requests define what evidence would change the record

The generator selects 31 exact missing-proof strings across all 16 providers. These are concrete requests such as a complete alert export, loss-inclusive ledger, execution assumptions, target allocation, fee and slippage treatment, source timestamps, model versions, correction policy, or independent support for a provider-published claim. The selection is deliberately narrower than each dossier’s full missing-proof list so this page stays focused on performance evidence.

The lexical themes overlap and are not exhaustive. One request can concern both archive integrity and execution accounting; another can be performance-relevant without matching any listed term. The authoritative output is the exact 31-item selection shown in the provider matrix and summary files, not the theme totals. A proof request is not proof that manipulation occurred. It states what the frozen record did not establish and what evidence could make a later review more conclusive.

A useful evidence packet would preserve every eligible alert with a stable ID and original timestamp; retain edits, deletions, cancellations, and open status; define wins, losses, break-even calls, duplicates, and exclusions; record fills, position size, leverage, target weights, stops, fees, slippage, funding, and failed orders; reconcile trade-level results to an account curve and drawdown; and identify any simulation, third-party source, model, or backtest version. Independent review should reproduce the calculation from that packet rather than repeat a headline.

The source packet has 152 assignments, while 102 lack a resolved publication date

The frozen cohort contains 152 provider-level source assignments resolving to 148 unique URLs across 42 hostnames. Each provider has between 7 and 10 assignments, with a median of 10. All 152 assignments carry the common access date 2026-07-10. Access date records when the research checked a route; it is not the publication date and does not show whether content changed before or after capture.

The 102 unresolved publication-date assignments are a freshness limitation, not a judgment about reliability. Some live pages do not expose a publication date, some records store null or blank, and some explicitly remain unresolved. The audit keeps those states visible rather than replacing them with access dates or inferred dates.

Record-level provenance prevents sentence-level source claims

The frozen research schema attaches sources to the provider record. It does not provide stable source IDs or direct references from each resultBoundary sentence and each missingProof item to one or more source rows. This generator can prove that the exact authored fields and source packet coexisted in a committed dossier. It cannot prove a one-to-one citation for every sentence, classify all 152 assignments as independent corroboration, or tell a reuser which single URL supports each clause.

That limitation is preserved in the JSON summary, CSV boundary, Dataset description, LLMS text, and this article. A record with many sources is not awarded a stronger posture because of count alone. Source types describe research roles, not votes. Primary material can establish what a provider published; independent-boundary material can constrain interpretation; historical material can establish a past state; registry material can answer identity or domain questions. None automatically supplies a complete live performance ledger.

The generator reproduces the frozen aggregation, not current provider performance

The generator requires a named Git ref, resolves it to a commit, and reads the four Best21 JSON blobs only through git show. It validates schema version, research date, provider count, unique slugs, source URLs, source types, access dates, publication-date formats, and the exact taxonomy coverage. It does not read mutable working-tree copies of the Best21 files. Provider rows are then sorted alphabetically.

The performance taxonomy is explicit and fixed in this source file. For every provider it assigns one evidence posture and zero-based indexes selecting performance-relevant strings from the frozen missingProof array. The generator preserves the exact resultBoundary and selected strings. Claim-language and proof-theme counts use the documented literal regular expressions. They are descriptive lexical measures, not semantic adjudication.

Generation asserts the requested frozen totals before writing: 16 providers; 152 source assignments; 148 unique URLs; 42 hosts; 31 selected proof items; posture counts of 11, 2, 2, and 1 in the documented order; claim-language counts of 10, 11, 9, 8, and 2; and 102 assignments without a resolved publication date. It also requires one Article, one Dataset, one BreadcrumbList, zero FAQPage, one H1, one literal primary action, 16 dossier links, 16 provider rows, the immediate decision boundary, and at least 3,000 visible words.

node tools/generate-crypto-signal-win-rate-accuracy-claims-audit.mjs --git-ref 5dcd30b2b1a0da9bacbaeec08244190dae19a49f --distribution-ref e4bc8b29e561f6137cdd2af0d0de5411d6e05941 --distribution-json-url https://gist.githubusercontent.com/TheCryptoSimon/5f175f59c906a29bc64f8054a264071d/raw/93aaa994333ac2528d4c3145aa492468b09c8eab/summary.json --distribution-csv-url https://gist.githubusercontent.com/TheCryptoSimon/5f175f59c906a29bc64f8054a264071d/raw/cecebaf71b61608e9c47aac36640b57d5fdc12ee/summary.csv --distribution-version eedd90183ecbfafded1d58070068806c3d630fc4 --distribution-page-url https://gist.github.com/TheCryptoSimon/5f175f59c906a29bc64f8054a264071d --published-at 2026-07-11 --modified-at 2026-07-11

Reproduction shows that the aggregation matches the committed CSR research objects and this fixed taxonomy. It does not independently verify the providers’ source pages, current offers, original trade records, fills, account returns, or later changes. Fresh performance verification remains a separate evidence task.

The aggregate release preserves the denominator and evidence boundary

The machine-readable release contains the exact aggregate metrics and alphabetical provider matrix used here. It preserves result-boundary wording and selected missing-proof text, but it does not reproduce third-party pages, private records, ratings, recommendations, or profitability conclusions. Any reused count should retain the 16-provider denominator, cohort date, frozen source commit, and record-level provenance limitation.

Release state: The public distribution arguments pin the release page, version, JSON, CSV, and matching repository commit so later site changes do not silently redefine this dataset.

• Download aggregate JSON

  30,170 bytes; SHA-256 232ebb7f4e93290b0c813fd53698b4806e5d42c5cbf9ae43457b33c2d354175e.

• Download aggregate CSV

  16,966 bytes; SHA-256 2181003e0a08b259807a19f1bcba53e5d641aca109a0e7abbf72aa90017369d3.

CryptoSignalsReview Evidence Desk. Crypto Signal Performance Claims: 16-Provider Evidence Audit. Cohort captured 2026-07-10. Frozen source commit 5dcd30b2b1a0da9bacbaeec08244190dae19a49f. https://cryptosignalsreview.com/crypto-signal-win-rate-accuracy-claims-audit/

Public release eedd90183ecbfafded1d58070068806c3d630fc4; matching repository release commit e4bc8b29e561f6137cdd2af0d0de5411d6e05941.

Aggregate reuse terms: CryptoSignalsReview permits quotation and reuse of the aggregate JSON and CSV with attribution to the Evidence Desk, canonical page, 2026-07-10 cohort date, and frozen source commit. Reuse must preserve the non-representative, non-ranking, non-endorsement, non-verification, non-accusation, and non-performance boundaries. This permission does not grant rights in third-party names, marks, source material, or personal data.

Official guidance supports scrutiny, not conclusions about this cohort

The following official sources provide general context for hypothetical performance, execution costs, independent verification, financial-promotion substantiation, and social-media investment claims. They were not used to rate, rank, accuse, endorse, or validate any provider in the matrix. Their presence does not turn a provider-level source packet into regulator confirmation.

• CFTC: commodity trading systems sold on the internet

  Official context for distinguishing hypothetical from actual trading, checking independent verification, and accounting for spreads, execution, commissions, fees, losses, and margin.

• FCA: 2026 review of financial-promotion approvers

  Official context for accurate, clear promotions and the FCA’s finding that some sampled approvals contained unsubstantiated claims.

• Investor.gov: social media and stock tip scams

  Official 2026 context for independently researching social-media investment claims and treating high-return, low-risk promises as warning signs.

Related evidence should remain separate from this performance audit

• Open the 21 provider evidence files

  The source cohort and narrative overrides cover broader identity, terms, route, security, and result questions. Coverage is not endorsement.

• Evaluate a provider’s accuracy and performance

  Use the existing decision guide to examine denominators, losses, execution assumptions, and account-level evidence before relying on a claim.

• Audit account access and execution authority

  A separate 16-provider analysis distinguishes manual signals, API trade permissions, copy trading, Cornix automation, and deposit or UID gates.

• Read the commercial terms audit

  Pricing, renewal, refund, and access evidence should not be collapsed into a win-rate conclusion.

• Check backtest assumptions

  Use a separate educational workflow for overfitting, data leakage, costs, and forward-testing questions.