> ## Documentation Index
> Fetch the complete documentation index at: https://lattora.puneetchandna.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Protocol and assumptions

> Understand pairing, measurement units and the limits of Research claims.

Lattora **2.1.0** uses frozen protocol/schema **2**.
Stress has a separate schema **1** and descriptive scope.
The product version does not replace these scientific identifiers.

## Model

Placement happens before release. Each case starts with fresh CloudSim Plus 8.5.7 state.
Hosts reserve PE/MIPS, RAM, bandwidth and storage. Overcommit and migration are absent.
Bandwidth/storage limits describe reservations, not simulated topology or contention.

One cloudlet runs per VM at constant full utilization for a seeded 60 to 120 seconds.
Space-shared schedulers and the 0.125-second minimum event interval are fixed.
A separate witness checks generated feasibility without guiding search or repair.
Required failures stop the run; inputs are never resampled for convenience.

<Accordion title="Frozen hardware and workload constants">
  Hosts have 16 PEs at 3000 MIPS/PE, 32768 MiB RAM, 10000 Mb/s bandwidth and 1000000 MiB storage.
  Host IDs cycle idle/max power pairs (175,250), (210,300) and (280,400) W.
  VM PE count is 1 or 2. MIPS/PE is 1000/2000/3000 and RAM is 1024/2048/4096 MiB.
  Each VM reserves 1000 Mb/s bandwidth and 10000 MiB storage.
  Release is zero relative to submission. Dynamic arrivals, host transition energy and empirical calibration are outside the model.
</Accordion>

## Pairing and seeds

The default master seed is 123456.
Explicit `java.util.Random` streams use SHA-256 over length-prefixed identity fields.
Scenario/workload seeds exclude algorithm and N/T, so competitors share immutable inputs.
Optimizer seeds identify the algorithm. Main and sensitivity workloads use separate phase seeds.
Keep replication identities and canonical hashes when comparing results.
Equal row counts alone do not establish pairing.

## Objective and measurements

| Quantity | Meaning |
| - | - |
| `objective_w` | Estimated steady power in watts; minimized during placement |
| `energy_j`, `energy_kwh` | Event-time integrated power; 1 kWh = 3,600,000 J |
| `sla_rate` | Violating cloudlets / all requested cloudlets; a proportion |
| `allocation_wall_ns` | Placement wall time in nanoseconds; secondary/descriptive |

Used hosts stay on until the common terminal/censor horizon; unused hosts stay off.
The simulator integrates the linear power curve. It does not substitute optimizer watts for measured energy.
Watts and joules can rank placements differently.

SLA counts failed/censored cloudlets and slowdown **strictly above 1.10** against isolated simulated references.
The denominator includes all requested work.
Strict reservations and constant utilization can make SLA identically zero.
Zero variation does not prove equivalence or successful SLA optimization.

## Claim decisions

Research uses 30 paired main replications.
It compares HO with three baselines across three scenarios: **18 composite claims**.
Paired energy log ratios and SLA rate differences enter wild-bootstrap/BCa/Holm analysis.
BCa intervals are unadjusted estimates; Holm controls the fixed familywise claim gate.

| Claim | Practical benefit bound | Other metric's noninferiority bound |
| - | - | - |
| Energy benefit | Energy interval upper ≤ `ln(0.95)` | SLA one-sided upper \< `0.01` |
| SLA benefit | SLA interval upper ≤ `-0.01` | Energy one-sided upper \< `ln(1.05)` |

Both require eligible inference and Holm-adjusted p ≤ 0.05.
Degenerate variance, undefined studentization or invalid bootstrap/jackknife intervals produce **NO\_CLAIM**.
The table records a reason, p=1 and empty unavailable intervals.
Missing required cases prevent complete analysis.

The [2.1.0 release](https://github.com/puneet-chandna/Lattora/releases/tag/v2.1.0)
reports **0 CLAIM / 18 NO\_CLAIM** under the degenerate SLA model.
Descriptive energy differences do not override that gate.
Do not add jitter, replace seeds or switch tests to force significance.

Explore, sensitivity and Stress are descriptive. Sensitivity does not establish optimal settings.
Runtime is secondary; it does not identify a winner across machines or JIT states.
These synthetic observations do not establish production savings or universal superiority.

[Executable constants](https://github.com/puneet-chandna/Lattora/blob/v2.1.0/src/main/resources/protocol.properties) ·
[Effective configuration](https://github.com/puneet-chandna/Lattora/blob/v2.1.0/src/main/java/org/puneet/cloudsimplus/hiippo/runtime/RunConfig.java) ·
[Independent validator](https://github.com/puneet-chandna/Lattora/blob/v2.1.0/scripts/statistics_validator.py).
The [experiment contract](https://github.com/puneet-chandna/Lattora/blob/v2.1.0/EXPERIMENTS.md)
documents the supported profiles and evidence boundaries at this release.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.