Token signal-to-noise ratio.
Token Signal-to-Noise Ratio measures the quality of signal in an AI operator's token flow — specifically, how much productive output the operator generates relative to the input and reused context that flow through the session. It is one of five canonical derived metrics in the MO§ES™ evaluation framework.
The formula.
Token SNR is the ratio of output tokens to the sum of input, output, and reused context tokens. It captures how much of the non-construction token flow is productive generation.
Where:
- O = Output (tokens generated by the model)
- I = Input (fresh tokens sent to the model)
- R = Cache Read (reused context tokens)
Note that Token SNR excludes Cache Write (W) from the denominator. This distinguishes it from Yield, which includes all four signals. SNR focuses on the signal quality of the ongoing token exchange — input, reused context, and output — without counting the one-time investment of building new context. An operator who builds a lot of new context (high W) will not be penalized in SNR for that investment, though they will be in Yield.
What Token SNR tells you.
Token SNR is the signal quality lens. It measures how much of the ongoing token exchange — the input the operator provides, the context the system reuses, and the output the model generates — is productive output.
A larger share of the input-output-reuse token flow is output. The operator's sessions generate more signal relative to the input and context they consume. This may indicate efficient prompting, minimal redundancy, or workflows where output dominates the token budget.
A smaller share of the token flow is output. More tokens are spent on input and context reuse relative to generation. This may indicate verbose prompting, heavy context dependencies, or exploratory workflows where input dominates.
The distinction between SNR and Yield matters when an operator is actively building new context. An operator in a context-construction phase may show low Yield (because W inflates the denominator) but normal SNR (because W is excluded). SNR isolates the ongoing exchange efficiency from the one-time context investment, making it a complementary signal to Yield rather than a replacement.
Token SNR in the evaluation pipeline.
MO§ES™ computes Token SNR from canonical telemetry for every observed session. The metric enters the pipeline at three stages.
Computed per session from INPUT, OUTPUT, and CACHE READ token counts. No prompt content required. CACHE WRITE is excluded by design.
Operators are ranked by Token SNR within their cohort. Percentile bands show where each operator falls relative to peers under the same conditions.
Token SNR contributes 20% of the AI Operator Development Index, the 0–100 composite score that summarizes overall operator behavior.
In the synthetic demo cohort of 50 operators across 1,668 observations, Token SNR values track closely with Yield for operators who build little new context, but diverge for operators in active context-construction phases. This divergence is one reason both metrics are retained in the canonical set — they capture different aspects of the token economy.
What Token SNR does not tell you.
Token SNR is a derived metric labeled DEVELOPMENTAL. It is a structural signal quality measure whose relationship to actual performance outcomes is still being tested.
- Not a quality measure. High SNR means more of the token exchange is output. It does not mean the output was correct, useful, or valuable.
- Not a productivity score. SNR measures signal efficiency in the token flow, not work output or business value.
- Not an employee ranking. DEVELOPMENTAL labels mean results route workflows, not people. No adverse employment actions are permitted from pilot data.
- Excludes construction by design. SNR deliberately omits Cache Write from the denominator. This makes it complementary to Yield, not redundant. Reading SNR without Yield can miss context-building activity.
- Validation required. The relationship between SNR and validated business outcomes has not been independently established. Correlations are labeled ASSOCIATION, never CAUSATION.
Every SNR measurement carries provenance: source telemetry window, operator cohort, evidence label (DERIVED), decision-use label (DEVELOPMENTAL), and synthetic-data flag where applicable.
Read alongside.
Productive output share of total token flow including construction. O / (I + O + R + W).
How much context the operator reuses and builds relative to new input. (R + W) / I.
The 0–100 AI Operator Development Index. Token SNR contributes 20%.