Fulcrum RVx Calculator

Kinetic Efficiency (E)

useful work / total time. 1 = no overhead.

0.60

ⓘ How to get E for a real boundary

Measure from distributed traces (AWS X-Ray or OpenTelemetry). For a representative sample of transactions crossing the boundary, E = sum(useful local compute time) / sum(total transaction time), where waiting on remote calls, serialization, and network are overhead (not useful). Code that computes it: run_study.py (AWS Lambda) and run_testbed.py (local HTTP testbed).

Semantic Distinctness (S)

1 = evolves independently. 0 = always changes with neighbors.

0.60

ⓘ How to get S for a real boundary

Measure from version-control co-change over a trailing window (e.g. 12 months). S = 1 - (change sets touching both sides of the cut) / (change sets touching this service). Use intent-unit change sets and drop mechanical repo-wide commits (monorepos deflate S). Code: gen_git_history.py.

Cognitive Load (L)

0 = easy for the team. 1 = overwhelming.

0.40

ⓘ How to get L for a real boundary

L = clamp(static complexity / team capacity, 0, 1). Complexity from LOC, file count, and (for serverless) code size / memory / config; capacity from the owning team's effective size. With no trustworthy capacity source, hold it constant and report the run as structural (requirement N16). Code: run_study.py (measure_L from deployed Lambda config).

0.00

-

Bands: below 0.4 distributed monolith - 0.4 to 0.7 at risk - above 0.7 optimal (illustrative defaults, calibrated per domain).

E (efficiency)

-

S (distinctness)

-

L (load)

-

RVx_raw

-

Formula: RVx = raw / (1 + raw), raw = (E^1.2 × S) / (L^0.8 + 0.1)

Number of tools shown (n)

how many tools the agent must choose among.

20

ⓘ How to get n

Count the tools/functions exposed to the agent in one request, i.e. the length of the MCP tool list (or the function schema array) the model must choose among. Model: exp2_agentic.py.

Tool Distinctness (S_a)

1 = no overlapping/redundant tools (paper Sec 11.5).

0.90

ⓘ How to get S_a

Estimate S_a = 1 - fraction of redundant/overlapping tools (tools with near-duplicate names, schemas, or purposes the model can confuse). Judge from the catalog or cluster tool embeddings.

Token-and-Latency Efficiency (E_a)

1 = little prompt bloat / few round-trips (paper Sec 11.5).

0.90

ⓘ How to get E_a

Paper: E_a = useful tool work / total interaction cost in one stated common unit (tokens and/or latency). Low when tool schemas bloat the prompt or the agent needs many turns. Quote the cost unit with the score.

0.00

-

Higher is a healthier tool boundary (proposed RVx-A). Agent Context Load L_a rises with tool count.

Agent Context Load (L_a)

-

Modeled selection accuracy

-

Accuracy model calibrated to published tool-selection-collapse data (RAG-MCP, tool-count studies). Illustrative, not a live LLM.

Transaction complexity (C)

0 simple, 1 deeply branched.

0.50

ⓘ How to get C

Rate the workflow's branching and conditional depth on 0 (a single linear happy path) to 1 (many branches, retries, and conditional compensation). See paper Section 9.

Business risk (R)

0 low stakes, 1 financial/compliance.

0.50

ⓘ How to get R

Rate the stakes of a failed or inconsistent transaction on 0 (cosmetic) to 1 (money, compliance, or safety). This is a business judgment, recorded per workflow. See paper Section 9.

Service interactions (i)

number of steps across services.

4

ⓘ How to get i

Count the cross-service steps (local transactions) in the saga. Read it from the workflow definition or a trace of one end-to-end transaction. SCS = C × R × i/(i+5).

0.00

-

Formula: SCS = C × R × i/(i+5). Higher favors a central orchestrator; lower favors event choreography.

Distributed-Monolith Cost (DMC) prices wasted transaction time from low Kinetic Efficiency: W = N × t × (1 - E) (seconds of overhead per period).

Requests in period (N)

e.g. monthly request volume.

Avg transaction time t (ms)

end-to-end wall time per request.

Kinetic Efficiency (E)

same E as RVx; sync from RVx tab below.

0.60

Cost per wasted second ($)

optional dollar overlay (illustrative).

Wasted time W

-

Overhead share

-

Illustrative $

-

Dollar rate is illustrative only. Prefer W (wasted seconds) as the measurable core from the same traces as E.

Paste one boundary per line as name,E,S,L (values 0 to 1), or load a CSV file. A header row is optional. This scores your real boundaries and ranks them.

RVx uses your measured E, S, L. Boundary names are rendered safely (no HTML). Thresholds are illustrative defaults. Do not paste account IDs or credentials.