Proof

Every number,
with its basis.

A ratio without a basis is not a number. The same data yields wildly different figures depending on what is measured, so each one below carries how it was derived and what it was derived from. Synthetic monitoring traffic is excluded from every figure on this page — it is the single largest source of error in measuring a system like this, and it inflated our own early numbers by a factor of four.

Counts derived 30 September 2026 · reproducible from source

CLEOS · production telemetry

24h7d30d90dAll time

Context eliminated

98.4%

aggregate, probe-excluded

Prompts over a full window

73.9%

17,381 prompts · 308 sessions

Sessions with real work

2,159

real, probe-excluded

Events stored

3,351,037

analytical store

Where the billed tokens went2.46B tokens · vendor-metered · 11,700 API calls

Context the model had already been sent — 95.2%Output and everything else — 4.7%Genuinely new input — 0.1%

A static representation of the production dashboard, counts accurate as of 30 September 2026. The live system runs behind our VPN and is shown, unfiltered and with the meter running, in the second fifteen minutes of a session — including pre-execution interception refusing an action on the basis of a decision made in an earlier session, which is the part that does not fit on a web page.

98.4%

Of context tokens eliminated per session

BasisAggregate across real sessions, probe-excluded. July 2026 validation harness.

One minus persisted output tokens over raw input tokens, summed across all sessions rather than averaged per session.

73.9%

Of production prompts carry more history than a full 1M context window holds

Basis17,381 real production prompts across 308 sessions, probe-excluded.

Accumulated history estimated per prompt and compared against the window. This is a capability measure, not a cost measure — it says the work was only possible because the memory layer was there.

95.2%

Of billed token volume was the model re-reading conversation it had already been sent

Basis2.46B tokens of production traffic, metered by the model vendor across 11,700 API calls, with prompt caching already enabled.

Vendor per-call accounting, cache-read tokens over total. Genuinely new input was 0.1%. This is the vendor's invoice, not our estimate.

0.045%

Worst disagreement on an independent recount, against a ±2% bar

BasisChecked against the model vendor's own tokenizer — the meter that writes the invoice. July 2026 validation harness.

Verdict: PASS.

Two independent systems, one conclusion

The model vendor’s billing records 2.34 billion tokens of context re-reads. Our own event store, which has no access to the vendor’s accounting and measures a different thing in a different way, independently estimates 2.93 billion tokens of avoidable context.

Two separate systems, agreeing within 25%. That agreement is worth more than either number on its own, and it is the reason we lead with this rather than with a single-source figure.

The operating record

Real, and small. We state it plainly because the sample is the evidence base for a ratio, not a claim about market size.

Sessions carrying real work2,159probe-excluded; excludes sessions that opened and captured nothing

Events in the analytical store3,351,037probe-excluded; 6,028,353 including synthetic probes

Consolidation operations214,106

Continuous operation sinceNovember 2025current architecture since April 2026

Two things a diligence engineer will find

43% of individual operations expand rather than compress. Net is positive, and it does not affect the end-to-end figure, because intermediate summaries are inputs to the next round and never persist. Our savings are signed, with no positive-only filter — that filter existed once, and removing it was a deliberate correction.

The compression sample is 96 sessions. Not a market-size claim. It is the evidence base for a ratio, and the ratio is verified against the vendor’s own tokenizer.

Bring your hardest sceptic.

Forty-five minutes: the story in fifteen, the live system in fifteen, your questions in fifteen. We will bring the token meter.

Request the session →