After workload commitment
Model / accelerator efficiency → IT electricity → heat removal → cooling → water strategy → grid coordination → emissions / infrastructure response
TA-14 ACA · VERSION 1.1 · PUBLIC DATA-CENTER RESEARCH SHOWCASE
Before compute becomes data-center demand.
The world is working to make rapidly growing computation more efficient after workloads exist. TA-14 Admissible Computation Architecture asks an earlier question: what must be true before a proposed computational path has sufficient standing to become workload at all?
THE GAP TA-14 ACA IS TESTING
Efficiency asks how to serve compute better. ACA asks whether all proposed compute should become demand.
+
Before workload commitment
Request → Proposal → admissibility → binding → COMMIT → computation → existing efficiency stack
ACA is intended as a complementary upstream layer. It does not replace efficient chips, better models, advanced cooling, workload shifting, renewable power, water stewardship or grid flexibility. It asks whether some demand can be avoided, bounded or resized before those systems ever have to serve it.
THE COMPUTATIONAL COMMITMENT BOUNDARY
Capability is not standing.
A system may be technically capable of generating, retrieving, simulating, routing, calling tools, spawning agents or retrying work. That capability does not by itself establish that the exact proposed branch, under the exact current conditions, should consume resources. ACA introduces an explicit decision boundary before commitment.
01Request
02Computational Proposal
03Admissibility
04Binding
05Commitresource boundary
06Computation
07Outcome
The bounded proposal has sufficient standing to proceed.
Required evidence, context, identity, scope or continuity is insufficient.
The proposal does not have sufficient standing to consume the requested resources.
The decision requires an authorized route outside the current gate.
WHAT COULD BE STOPPED, HELD OR RESIZED?
Not just bad requests. Unnecessary branches inside otherwise legitimate work.
The strongest data-center hypothesis is not that entire user requests disappear. It is that high-volume computational systems contain branches whose standing can fail before commitment.
Work that repeats an already-satisfied objective without adding required value.
Computation proposed from context that materially changed before commitment.
Retrievals, tools, agents or model calls that exceed the bounded objective.
A request that may deserve an answer, but not the amount or class of compute proposed.
Retries or continuations whose prior standing no longer survives changed conditions.
Work whose purpose, target, authority or expected consequence cannot be sufficiently bound.
Example:A request may deserve an answer, but not necessarily a large model, fifty retrievals, repeated retries and an unconstrained agent swarm. ACA can test both whether computation has standing and how much computation that standing supports.
WHY THIS COULD MATTER AT DATA-CENTER SCALE
A tiny percentage at enormous volume can become an infrastructure question.
TA-14 ACA does not claim that a denied model call equals a fixed quantity of electricity, water or carbon. The research proposition is narrower and stronger: first establish whether upstream governance reduces net downstream computational work after including the cost of the gate. Only then test whether the reduction is large, repeatable and attributable enough to matter physically.
PRIMARY COMPUTATIONAL TESTNet computational benefit = baseline total work − (ACA gate work + admitted downstream work + shifted / retry work)
L0Compute / work units
→L1IT energy
→L2Facility / thermal
→L3Water
→L4Emissions
→L5Infrastructure
Each attribution level requires additional evidence. The architecture intentionally prevents a computational result from silently becoming an environmental claim.
THE FOUNDING PILOT
One workload. Two conditions. Preserve whatever happens.
Define objective, context, resource class, quality threshold and baseline before the comparison begins.
Measure the existing architecture exactly as it operates without the ACA gate.
Apply admissibility before commitment and preserve ALLOW, HOLD, DENY and ESCALATE decisions.
Gate overhead, admitted work, avoided work, retries, latency, quality, false rejection and false admission.
Only where supported, connect compute differences to IT energy, thermal load, cooling, water or grid effects.
Positive, null, negative and inconclusive outcomes all remain admissible research findings.
WHO SHOULD TRY TO DISPROVE THIS?
The people already working on data-center electricity, cooling, water and grid pressure.
ACA becomes meaningful only if researchers and operators with real workloads, telemetry and facility data can test it. The Exchange Research Network is being built around the communities best positioned to answer the question.
Test whether workload admission can complement efficiency, cooling and demand-response systems.
Measure whether upstream computational governance produces facility-relevant reductions after gate overhead.
Examine whether avoided or reshaped compute changes demand before load-management mechanisms engage.
Stress-test the falsifiable proposition, accounting boundary and environmental attribution ladder.
Test inference, retrieval, tool-use, agent branching, retry and routing workloads.
Determine when compute reductions are actually attributable to thermal, cooling or water effects.
What must be true before computation has sufficient standing to consume resources and create consequence?
Govern before computational commitment.