Same demand. Different work.

How much state computation is genuinely reusable? Preserve the requests, the estimators and the local acceptance checks. Count only the work whose inputs and requirements actually match.

ANALYTICAL MODEL · NO DEVICE TELEMETRY

Consumers, N Source versions per second, f Versions with shareable work, p

Cost, isolation and baseline assumptions — expose the costs that could reverse the result

Required view / replica groups, K CPU per transform, c (µs) Added CPU per request, h (µs) Other added CPU, B (ms/s)

Baseline already shares the same eligible computations

One consumer request per source version. Same per-transform cost on both sides. K groups preserve required distinct views or replica domains. h includes additional per-request lookup / delivery work; B captures other additional maintenance cost. All costs here are assumptions.

01 / Follow the executions

□ source version○ state use● transform execution

Baseline integration

separate computation

Proposed shared contract

reusable materialization

Every mark is one modeled event. Both panels use the same rows, source versions and time scale. Consumer-local policy checks remain on both sides; this view isolates the transform operation, not the entire autonomy stack. Mark positions are logical source epochs, not measured execution latency.

Inspect a modeled event without selecting a small chart mark

Select a mark to inspect its modeled meaning. Source-version identity is preserved through reuse; receiving a reused result does not make it new independent evidence.

02 / Count demand separately from work

Selected workload / rate	Baseline	Proposed

CPU totals include only the selected transform and the modeled added overhead. Upstream estimation and unchanged consumer-local work are outside this cost subtotal. Operation-count reductions are not whole-system speedups.

03 / Find the break-even point

Sweep only the shareable fraction p. Hold all other inputs fixed. Positive values mean less CPU in this scoped model; negative values mean sharing costs more.

04 / The computation

W_base = Nf
W_shared = f[(1 − p)N + pK]
ΔCPU = pf(N − K)c − Nfh − B

In the equation, c and h are expressed in ms, B in ms/s. When the baseline already shares, W_base = W_shared, so this model offers no compute savings and retains the added overhead.

Grouping identical operations only identifies an opportunity. It does not prove that a shared result can arrive before each consumer's deadline, preserve required isolation, or satisfy uncertainty and freshness requirements.

05 / Replace assumptions with traces

Instrument operation starts, completions, exclusive CPU time, consumer delivery, deadlines, evidence versions and acceptance results. Keep unrelated state, deliberate independent checks and unknown semantics out of the reuse count. Compare paired replays before changing a control path.

Inspect the candidate signature

operation + algorithm version
entity + immutable evidence / state versions
valid time + time scale + clock-model version
source / target frame + transform version
parameters + calibration + numeric contract
evidence policy + permission partition
independence + placement domain + cache epoch

Read the supporting methods

LLVM: common-subexpression elimination / Early CSE — reuse computationally equivalent work.
ROS 2 tracing — runtime instrumentation; application semantics require additional tracepoints.
Materialize — incrementally maintained views, a design precedent rather than a proposed dependency.
PX4 multiple EKF instances — preserve fault-detection redundancy.
Edward Tufte, Envisioning Information — small multiples and layered comparisons.

This local artifact makes no live requests, collects no telemetry and changes no production system. No latency, error-rate, safety or device-performance improvement has been measured.