GSF SCI for AI · reference implementation
Energy, carbon, water and land for every call your system makes, with cost sitting outside them as the same consumption, priced. Measured per task, not per query, because an agent does not make one request.
Tetra, four. Meter, measure. The name is the specification: energy, carbon, water and land, with cost outside them as the same consumption priced.
It counts twice. A tetrameter is also a line of verse in four metrical feet — four beats, which is what the mark draws.
LLM observability measures one thing. Carbon tools measure one or two. The largest ESG platform measures three. The number in the name is the difference, and it is not one a competitor can adopt without rebuilding their product.
One report, measured
This is a real recognition report from SiteBeacon (opens in a new tab): 26 model calls across five providers, in one trace, on 1 August 2026. Nothing here is illustrative. The figures come from the same engine the product ships.
Notice that cost is exact and carbon is not. The provider bills us to the hundredth of a cent, so cost is Tier 4 evidence with no band at all. Nobody publishes the energy draw of a proprietary model, so carbon spans two orders of magnitude and says Tier 1 in the corner.
That gap is the product. A tool that printed a single tidy carbon number for the same call would not be more accurate. It would be less honest.
1.09 gCO₂e is what a single-number tool prints here.
| Derived from the same energy figure | Range | Tier |
|---|---|---|
| Energy | 0.0004 – 0.026 kWh | 1 |
| Water | 0.0004 – 0.11 L | 1 |
| Land | 0.043 – 30.8 cm² | 1 |
These three are one measurement in three units. Each is energy multiplied by a published constant, so they share a band rather than confirming each other.
Tier 1 is a class average. Tier 4 is measured. The tier is not our confidence. It is a statement about what evidence exists, and provider disclosure is the binding constraint on three of these five.
The unit
The reasonable objection to trace-level measurement is that per-call already covers everything. It does — and that is the problem. Fifteen requests behind a contract review look exactly like the two behind a document extraction. All ordinary, none remarkable.
One of those tasks costs nineteen times the other. That fact does not exist at the level of a call. It only appears once you group the calls by the thing the business actually asked for.
Which is why a per-query average is not a smaller version of the truth. It is a different number, and it gets further from the truth the more agentic your system becomes.
The number nobody shows you
5,416 of 15,289 tokens went to intermediate reasoning, discarded candidates and retries. You paid for all of them and none appeared in the output. Per-query dashboards cannot see this, because the waste lives in the space between the calls, and that space is exactly what a trace is.
Invisible tokens are not a failure. Reasoning is often worth buying. But you cannot decide whether it is worth buying until somebody counts it.
See it working
A demonstration organisation is open without a login: four weeks of traffic, five customers, five models, and the waste findings the detectors actually produced from it.
The traffic is generated rather than measured, and the page says so at the top. Our real capture carries a live customer’s account id, and a product that promises not to leak identifiers should not publish one to make a sales point. Every figure on it is computed by the engine your own data would run through.
| Traces | 254 |
| Calls | 868 |
| Tokens | 2,712,503 |
| Billed but never surfaced | 300,982 · 11% |
| Waste findings | 43 |
| Recoverable | $0.22 – 0.78 |
Findings across all four detectors: planning loops, context resend, redundant calls, and reasoning spent on trivial answers.
Seven differences
LLM observability measures one thing: cost. Carbon tools measure one or two. The largest ESG platform on the market measures three, and none of them can see a token.
Every competitor measures per query. An agentic workload makes one business outcome cost hundreds of times a chat turn, so a per-query average is not a smaller version of the truth; it is a different number. We measure the task.
Our biggest structural edgeRunaway planning loops, redundant tool calls, quadratic context re-send, multi-agent debate that changed nothing, retries without backoff, reasoning spend on trivial inputs, and provider-billed tokens you cannot inspect. Each one costed in money, grams and litres.
Average for the inventory, because that is what the GHG Protocol asks for. Marginal for every reduction claim, because that is what is physically true. Published research shows the same intervention reads 18% savings on one signal and 11% on the other, and goes negative measured a third way. Most claims in this market are artifacts of the vendor's chosen signal.
We shadow-evaluate the cheaper path against your real traffic and report the measured quality delta before recommending a switch. Nobody captures the model-choice lever because of fear of regression, not ignorance. We sell certainty, not insight.
You can hand your customer a real AI carbon number for their account. Structurally hard for cost tools, which think in cost centres, and impossible for ESG platforms, which never see a token.
A methodology statement, per-number data lineage, tier classification, uncertainty ranges, every emission factor with its source, version and retrieval date, and a restatement log for when a coefficient changes. Not a feature but a discipline, and the hardest thing here to copy.
The hardest thing here to copyNo field in our data model can hold a prompt or a completion, and the ingest endpoint rejects a request carrying one rather than dropping it quietly. That opens banks, health, legal and government: buyers that prompt-capturing tools cannot serve at any price.
Where this sits
There is roughly a hundred-fold price gap between LLM observability and enterprise carbon accounting. We serve the budget at the top of that range and cost a fraction of it.
| Category | Measures | Unit | Typical price |
|---|---|---|---|
| LLM observability | Cost, latency, quality | Per call | $29 – 2,499 / mo |
| AI FinOps | Cost | Per call | $30 – 200 / mo, or 1% of spend |
| Enterprise carbon platforms | Emissions across the org | No token visibility | $50,000 – 250,000 / yr |
| Tetrameter | Energy, carbon, water, land + cost | Per trace | $499 / mo – $60,000 / yr |
Prices as published by each vendor, July 2026. Ours are on the pricing page in full, because a product that argues for disclosure should start with its own.
Design partners, 2026
A measurement product takes about nine months to become credible enough to sit in front of an auditor. We are taking a small number of design partners who want to arrive at that deadline with their AI footprint already documented.