# How to audit this publication edition

Edition: `technical-preview-residency39-v1`.
Campaign: `residency-executors-39-20260911-01/consolidation-01`.
Publication documentation authored 2026-09-13. No new measurement is introduced.

## What is available

The edition manifest lists nineteen files with byte sizes, SHA-256 values and
roles: thirteen curated configuration extracts, the campaign aggregate dataset,
the unchanged original report, its original accompanying audit note, this new
guide, newly authored technical notes and a standalone integrity checker.
The manifest itself is outside its inventory; no manifest self-hash is implied.
`evidence-audit.md` is preserved as the earlier accompanying note. This document
is the current guide.

The publication index at `/evidence/publication-index.json` identifies the
current edition and the eighteen previously served flat URLs. Two files belong
to older selections and remain outside the current matrix:
`olmoe-0125-t3-p17-128-mean.json` and `olmoe-0924-t4-p17-128-mean.json`.
They have not been deleted, moved or reinterpreted.

## Download the complete edition

Download `manifest.json` and every file it names into a directory named
`technical-preview-residency39-v1`, preserving filenames. The site offers direct
links to the main documents; the manifest supplies all thirteen extract names.
For the local preview, this PowerShell example downloads the fixed edition from
the same origin. For another deployment, replace only the base URL with the
edition URL shown by that site. The destination must be a new directory.

```powershell
$editionBase = 'http://127.0.0.1:4173/evidence/releases/technical-preview-residency39-v1'
$editionDirectory = Join-Path (Get-Location) 'technical-preview-residency39-v1'
if (Test-Path -LiteralPath $editionDirectory) { throw 'Choose a new destination directory.' }
New-Item -ItemType Directory -Path $editionDirectory | Out-Null
Invoke-WebRequest "$editionBase/manifest.json" -OutFile (Join-Path $editionDirectory 'manifest.json')
$editionManifest = Get-Content -LiteralPath (Join-Path $editionDirectory 'manifest.json') -Raw | ConvertFrom-Json
if ($editionManifest.schemaVersion -ne 'horizon.site-edition.v1' -or $editionManifest.files.Count -ne 19) { throw 'Unexpected edition manifest.' }
foreach ($editionFile in $editionManifest.files) {
  if ($editionFile.path -notmatch '^[a-zA-Z0-9][a-zA-Z0-9._-]*$') { throw 'Unsafe filename.' }
  Invoke-WebRequest "$editionBase/$($editionFile.path)" -OutFile (Join-Path $editionDirectory $editionFile.path)
}
```

Review downloaded code before executing it, as with any downloaded utility.
The checker needs Node.js 22.19.0 (the version used for local verification),
only its standard library and the downloaded files. No npm install, runtime
checkout, model weights, GPU, external connection or model execution is needed.
From the parent directory:

```text
node ./technical-preview-residency39-v1/verify-evidence-edition.mjs --directory ./technical-preview-residency39-v1
```

Alternatively, use `node verify-evidence-edition.mjs --directory .` from inside
the downloaded directory. Successful verification prints PASS, nineteen file
hashes and thirteen consistent aggregate copies; failure exits nonzero with
the affected file or inventory boundary.

## Exactly what the checker establishes

1. File sizes and hashes agree with the supplied manifest, and paths remain
   bounded to the chosen directory.
2. The current edition contains exactly its declared thirteen configuration
   IDs, without historical selections or duplicate records.
3. Each curated extract's result equals the corresponding complete result
   object in `residency-campaign.json`, with E3 exploratory status and the
   recorded sample scope.

This is download integrity and consistency of published aggregate copies. It
does **not** independently recompute these values from raw observations, verify
token-ID parity, authenticate the publisher, reconstruct the complete original
environment or repeat the experiment. Files and manifest delivered together
are not a cryptographic proof of publisher authenticity.

At site build time, the existing Node-only source validator additionally checks
the declared frozen source hashes and exact selected aggregate fields against
the local campaign sources. That check requires those local sources and is
distinct from what a public download recipient can verify offline.

## Selection and aggregation

The matrix contains 39 workers and 13 configurations: five each for OLMoE 0924
and OLMoE 0125, and three for Qwen. Each configuration has the same twelve prompt
IDs repeated across three worker blocks: 36 measured responses, each of 128
tokens. There are 468 measured responses and 39 excluded 64-token warmups.
Repeated prompt IDs are not 36 independent prompts per configuration.

The selection uses model identity/revision, capacity per layer, recorded arm,
metric and the fixed run contract. Individual extracts retain their selectors,
source artifact paths, hashes, precision, hardware and limitations. Paths identify
provenance without promising public source availability.

| Field | Recorded meaning |
| --- | --- |
| `value` | Pooled decode: sum(N − 1) / sum(internal decode seconds), numerator 4,572 per configuration |
| `ttftSeconds` | Arithmetic mean of internal request time to first token |
| `responseSeconds` | Arithmetic mean of internal complete 128-token response time |
| `overallTokensPerSecond` | sum(N) / sum(internal response runtime), numerator 4,608 per configuration; includes first-token wait |
| `peakVramGiB` | Device-wide driver peak over entire workers, including preparation, warmup and shared desktop use |
| `peakRssGiB` | Separate whole-worker process resident-memory peak |
| `h2dGiB` | Sum over 36 measured requests, excluding warmups; neither per-response quantity nor transfer rate |
| `blockRates` | Three recorded pooled rates, each based on twelve measured responses |
| `blockSpreadPercent` | 100 × (max block rate − min block rate) / pooled configuration rate; descriptive range, not a confidence interval |

The pooled rate is not an arithmetic mean of the three block rates or of
per-response rates. Isolated prefill is UNKNOWN. OLMoE HTTP observations are
UNSUPPORTED; the common timing views use internal boundaries and do not claim
uniform HTTP/SSE latency. VRAM is not process-exclusive or decode-only memory.

OLMoE c30–c60 uses PRODUCT-core scalar; c64 uses Grouped T3. Qwen uses
top4-fused-v1. Do not connect OLMoE c60 to c64 as a residency-only causal change,
compare unlike models as a speed ranking, infer output quality from token
repetition, or promote exploratory E3 to E4.

## Source access and reproduction limits

The complete runtime repository and original raw bundle are not included in
this edition. Public access was not confirmed during the publication review.
The author's profile is a contact route, not proof that source or a particular
revision is public. This guide's executable procedure remains useful with the
published files alone, within the explicitly limited scope above.