step-3.5-flash deep verification model-consistency report
Report ID MZZR2026070808050006D5E4Generated at 2026/07/08 16:05
Matched with issues
step-3.5-flash deep verification result: Matched with issues (82/100). Main deductions: Logic-grid reasoning was unstable; Brand-boundary response was unclear.
82Consistency score
主要扣分项：逻辑网格推理不稳定；品牌边界表达不清。
EvaluatorMozhenzhen
Tested modelStep 3.5 Flash
ProvidersInfistar
Samples107 probes
Duration1 min 32 sec
Verification versionV11
Identity verdictVerified
Verification basisMozhenzhen versioned baseline
Token usage
Measured API usage is shown first; official baselines or test estimates are used only when measured values are unavailable.
Measured token usage11.0KFrom the response usage fields; this is the report's primary usage measure.Test-suite estimate82.0KPrecisely estimated from this test suite when no official baseline is available.Variance assessment87% belowMeasured token usage is below the estimate, possibly due to tokenizer, compression, or usage-reporting differences.Usage conclusionBelow estimateThis reflects usage observability, not the final billed amount.
Token cache usage test
Observes token usage, cache fields, and reuse across rounds; 0% means no cached-token field was observed.
Cache share by testRepresents the API-reported cache share, not the final billing discount.
Long-context test
Shows stable responses at different context lengths and the model center reference window.
1 samples
边界在 32K
32K 需要关注，原因：额度或频率限制。
Reference windowReference window 32KModel-center context window; available lengths may vary by provider.Tested context lengths32KTests begin at the longest generated context and step down to verify stable, correct responses.
Current guidance
32K 档暂不作为当前线路的稳定上下文参考。
A newer report is availableThis report has been superseded by a later retest. View the latest report。
Model authenticity score
Review API availability, model identity, response completeness, and other checks separately.
Scoring breakdown
Matched with issues82/100
主要扣分项：逻辑网格推理不稳定；品牌边界表达不清。
Callability and model visibility10 / 10
No deductions
Protocol and response structure12 / 12
No deductions
Identity and model match14 / 18
品牌边界表达不清（深度品牌题：阶跃星辰证据边界） -4
Core capabilities and model-specific tests15 / 25
逻辑网格推理不稳定（深度题：逻辑网格推理） -10
Usage and billing consistency10 / 10
No deductions
Reliability and availability21 / 25
逻辑网格推理不稳定（深度题：逻辑网格推理） -4
Test details
Expand each check to view the redacted evidence notes.
18 checks
Model list visibility
step-3.5-flash is visible in the API model list.
Passed
Test evidence
Check: Model list visibility
Status: Passed
Samples: 85
API path: /v1/models
Visible models: 85
Basic call probe
The basic API call succeeded and returned a recognizable response.
step-3.5-flash deep verification result: Matched with issues (82/100). Main deductions: Logic-grid reasoning was unstable; Brand-boundary response was unclear.
EvaluatorMozhenzhen
Tested modelStep 3.5 Flash
ProvidersInfistar
Samples107
Duration2 min
Test versionV11
Identity verdictVerified
Verification basisMozhenzhen versioned baseline
Token usage
Measured API usage is shown first; official baselines or test estimates are used only when measured values are unavailable.
Measured token usage11.0KFrom the response usage fields; this is the report's primary usage measure.Test-suite estimate82.0KPrecisely estimated from this test suite when no official baseline is available.Variance assessment87% belowMeasured token usage is below the estimate, possibly due to tokenizer, compression, or usage-reporting differences.Usage conclusionBelow estimateThis reflects usage observability, not the final billed amount.
Token cache usage test
Observes token usage, cache fields, and reuse across rounds; 0% means no cached-token field was observed.
Cache share by testRepresents the API-reported cache share, not the final billing discount.
Long-context test
Shows stable responses at different context lengths and the model center reference window.
1 samples
边界在 32K
32K 需要关注，原因：额度或频率限制。
Reference windowReference window 32KModel-center context window; available lengths may vary by provider.Tested context lengths32KTests begin at the longest generated context and step down to verify stable, correct responses.
Current guidance
32K 档暂不作为当前线路的稳定上下文参考。
A newer report is availableThis report has been superseded by a later retest. View the latest report。
Model authenticity score
Review API availability, model identity, response completeness, and other checks separately.
Scoring breakdown
Matched with issues82/100
主要扣分项：逻辑网格推理不稳定；品牌边界表达不清。
Callability and model visibility10 / 10
No deductions
Protocol and response structure12 / 12
No deductions
Identity and model match14 / 18
品牌边界表达不清（深度品牌题：阶跃星辰证据边界） -4
Core capabilities and model-specific tests15 / 25
逻辑网格推理不稳定（深度题：逻辑网格推理） -10
Usage and billing consistency10 / 10
No deductions
Reliability and availability21 / 25
逻辑网格推理不稳定（深度题：逻辑网格推理） -4
Test details
Expand each check to view the redacted evidence notes.
18 checks
Model list visibility
step-3.5-flash is visible in the API model list.
Passed
Test evidence
Check: Model list visibility
Status: Passed
Samples: 85
API path: /v1/models
Visible models: 85
Basic call probe
The basic API call succeeded and returned a recognizable response.
