Efficiency
The three-pillar waste score, what it deliberately excludes, how coverage is stated, and where the line between telemetry and a rating sits.
Efficiency
Efficiency is the diagnostic lens on waste still on the table -- retries, repeated prompts, cache busts -- attributed to people, teams and services.
Efficiency answers "where is spend inefficient, and who needs help". Savings answers "what did the platform already save you". They read the same telemetry and are never pooled.
The three pillars
The score is computed from exactly three directly measured token pillars.
| Pillar | What it counts |
|---|---|
| Retry waste | Tokens spent on retried requests. |
| Prompt repetition waste | Tokens beyond the first occurrence of an identical prompt, per person, over the window. |
| Cache waste | Tokens that missed the provider's cache due to prefix drift. |
A retried request's tokens count as retry waste alone. Repetition and cache waste are not exclusive of each other.
Waste share and the score
Waste share is the fraction of an entity's tokens in the window that were waste. It is token-denominated, so it holds in both billing regimes.
The efficiency score is its complement:
score = 100 x (1 - wasted tokens / total tokens)| Score | Label |
|---|---|
| 90 and above | Healthy |
| 75 to 89 | Watch |
| Below 75 | High waste |
The score is absolute, never relative to peers. It does not rank your team against anyone else's, and there is no "configure weights" control because waste tokens simply sum -- there are no weights.
Labels are diagnostic rather than accusatory. Accusatory phrasing was deliberately retired from this surface.
Absolute waste is reported in dollars for metered traffic only, from the retry and repetition pillars. On subscription traffic waste has no dollar denomination.
Each row also names its dominant pillar -- retries, repetition or cache -- which is simply the largest of the three rather than a concept of its own.
Recoverable waste is the prioritization metric
Waste share multiplied by volume.
A poor score on tiny volume is noise. Recoverable waste is what orders the "where to look first" list, and it is why that list is not simply the worst scores.
Avoid reading it as money recoverable under a subscription, where waste has no dollar denomination.
What is deliberately excluded
A known consequence, accepted deliberately: the score understates inefficiency in the colloquial sense. A person whose traffic is wasteless but whose prompts are poor scores 100.
Coverage is stated, not implied
The page reports two coverage facts in place:
- Since when the cache pillar has been measured. Earlier requests predate the columns it depends on, and those rows default to zero -- which is not evidence of health.
- What share of your team's tokens are identity-attributed. Unattributed traffic is not in per-person scores.
The uncovered share is unmeasured, not merely unlabelled. Do not read coverage as completeness.
A person carrying only an attribution tag, with no vendor identity behind it, has no efficiency score at all rather than a score of zero.
Methodologies are never mixed inside one pillar, and there is no backfill from a different methodology to close a gap.
The ratings visibility boundary
This is the line between measured telemetry and judged content, and it is enforced server-side rather than in the interface.
| Visibility | |
|---|---|
| Measured per-person waste -- the efficiency score, the pillar token counts, recoverable waste | Operational telemetry. Visible to page viewers. |
| Anything a judge produced about a person -- coaching summaries, rubric scores, automation-cluster recommendation text, prompt examples | A rating. Self-only by default, behind a per-team flag, and fail-closed. |
The same table can sit on both sides. An automation cluster's measured token count feeds the repetition pillar openly, while the judge-written recommendation text beside it is gated.
Coaching therefore appears only in a person's drill-down, behind the flag, and in a self-coaching card that is always available for your own data.
Reading "enabled" as "active" is a trap worth avoiding: a global admin can make ratings visible without the team being able to turn them off.
Three score-shaped numbers that are not this score
If a score does not match what you expect, check which of these you are reading.
| Figure | What it actually is |
|---|---|
| Legacy composite score | A superseded per-person computation that still fills the person detail response, using a different formula, different thresholds and a different label vocabulary. It is indistinguishable from the real score on the wire. |
| Service efficiency label | A three-valued verdict on a service computed from its retry and cache-hit rates, by a heuristic the efficiency authority neither uses nor covers. There is no per-service efficiency score. |
| Spend band | A three-valued label ranking a team member by raw spend. It carries no output-value term and measures no return, despite having shipped under an ROI name. |
A team's efficiency figure is a token-weighted mean of its members' scores. An unset value means unknown and is rendered as an honest blank; a present zero is a genuine all-waste score.
On screen
The Efficiency page, under Usage, carries:
| Section | |
|---|---|
| Efficiency map | Spend against score. Big dots are heavy users; low and to the right needs attention first. |
| Where to look first | People ranked by recoverable waste in the window. |
| Your coaching | Your own data, always available to you. |
| Team efficiency | |
| Model fit | The separate advisory. |
| How the score works | The formula, the three pillars, and the two coverage statements -- with a link to link identities if attribution is thin. |
Savings
Measured against estimated, effective tokens against raw ones, and why a sub-percent share of effective context is not a small result.
Attribution
How spend is attributed to people, teams, departments, services and seats -- and which of those are records you curate versus values observed in traffic.