# Talos vs Prometheus — matched workload comparison

Measured 2026-10-17. This report and its data download cover 6 workloads completed by both engines at all three presets. The website adds the per-workload medians to show totals for these same 6 scripts.

The baseline is the official packaged upstream Prometheus CLI v0.2.11.1 using its bundled Lua runtime and the LuaU target. Presets are operational pairings and do not guarantee equivalent protection.

## Totals shown on the website

Each paired cell is **Talos / Prometheus**. Lower values are better. File sizes use KiB (1,024 bytes). These totals differ from geometric-mean speedup summaries; the large-string test strongly affects the totals.

| Talos / old setting | Generation ms | Execution ms | Output KiB |
| --- | ---: | ---: | ---: |
| Light / Minify | 43.657333 / 165.346126 | 0.198583 / 0.122208 | 119.562500 / 117.580078 |
| Balanced / Medium | 46.918042 / 696.878583 | 4.924708 / 50.490417 | 50.336914 / 335.008789 |
| Strong / Strong | 57.506377 / 941.636292 | 32.322333 / 81.311458 | 97.159180 / 507.012695 |

## Individual workloads

Runtime is the median of 9 per-variant medians. Generation and size are medians across 9 variants. Sub-microsecond runtimes are too short for useful speed claims. Every paired cell is **Talos / Prometheus**.

### Light vs Minify

| Workload | Generation ms | Execution ms | Output KiB |
| --- | ---: | ---: | ---: |
| arithmetic | 9.551209 / 25.840750 | 0.155542 / 0.080792 | 0.401367 / 0.075195 |
| closures | 5.851833 / 23.156542 | 0.018625 / 0.017667 | 0.466797 / 0.136719 |
| bulk | 6.838500 / 33.575459 | 0.000417 / 0.000167 | 1.820312 / 1.490234 |
| constant_reuse | 6.786958 / 26.618583 | 0.023250 / 0.023375 | 0.445312 / 0.113281 |
| cold_strings | 6.874708 / 31.293167 | 0.000458 / 0.000167 | 47.701172 / 47.368164 |
| unused_strings | 7.754125 / 24.861625 | 0.000292 / 0.000042 | 68.727539 / 68.396484 |

### Balanced vs Medium

| Workload | Generation ms | Execution ms | Output KiB |
| --- | ---: | ---: | ---: |
| arithmetic | 6.827333 / 100.596500 | 2.153583 / 3.096042 | 4.907227 / 26.599609 |
| closures | 7.184500 / 97.168541 | 0.742792 / 1.458750 | 5.768555 / 26.201172 |
| bulk | 8.139875 / 114.152834 | 0.234250 / 0.788625 | 13.887695 / 46.053711 |
| constant_reuse | 7.096709 / 123.070250 | 0.812208 / 5.279458 | 5.063477 / 25.544922 |
| cold_strings | 8.708125 / 134.210500 | 0.972792 / 33.850625 | 14.636719 / 94.039062 |
| unused_strings | 8.961500 / 127.679958 | 0.009083 / 6.016917 | 6.073242 / 116.570312 |

### Strong vs Strong

| Workload | Generation ms | Execution ms | Output KiB |
| --- | ---: | ---: | ---: |
| arithmetic | 7.898000 / 117.995125 | 19.054667 / 21.868542 | 13.130859 / 39.135742 |
| closures | 8.818625 / 125.682958 | 6.156458 / 6.594958 | 16.528320 / 41.982422 |
| bulk | 11.532000 / 226.201333 | 0.462917 / 1.117458 | 21.738281 / 123.701172 |
| constant_reuse | 8.898459 / 142.668083 | 5.517625 / 9.894250 | 13.342773 / 39.223633 |
| cold_strings | 10.756334 / 188.929834 | 1.073292 / 35.953833 | 20.911133 / 135.976562 |
| unused_strings | 9.602959 / 140.158959 | 0.057375 / 5.882417 | 11.507812 / 126.993164 |

## Method

- 6 workloads: arithmetic, closures, bulk, constant_reuse, cold_strings, unused_strings.
- 9 independent generated variants per engine/preset/workload; each variant runs 9 times in a fresh Luau process. All 2,916 executions in this matched comparison returned the same result as native source.
- One initial invocation followed by measured fresh-process CLI invocations for both engines. Generation is launch-to-exit wall time including startup, file reading/writing, random seeding and logs.
- Execution includes VM initialization, decryption and decompression. Luau parsing, compilation and process startup are excluded. Luau 0.741, optimization level 1, no native codegen.
- Light is paired with Minify, Balanced with Medium, Strong with Strong.
- Results describe this machine and corpus. Memory and reverse-engineering resistance were not measured; performance does not establish stronger protection.
- Machine: `macOS-27.2-arm64-arm-64bit-Mach-O`.
- Reference toolchain: `LuaJIT 2.1.1774896198 -- Copyright (C) 2005-2026 Mike Pall. https://luajit.org/`.
- Raw samples include per-variant timings, output sizes, variant hashes, and engine/runtime hashes. P10–P90 values describe variant spread, not confidence intervals.

The [matched dataset](results.json) uses exactly the same 6 workloads as this report. Re-export the website snapshot from the preserved measurement archive with `python3 benchmarks/export_website.py`.