# Selected campaign audit — publication edition v3

Authored 2026-09-15 for `technical-preview-residency39-v3`, applying the
owner-approved policy in the Horizon repository at
`docs/superpowers/specs/2026-09-15-contract-valid-publication-policy-design.md`.

The selected `residency-executors-39-20260911-01/consolidation-01` campaign
contains thirteen configurations, 39 workers, 468 measured 128-token responses,
and 39 excluded 64-token warmups. Each configuration has twelve prompt IDs
repeated across three blocks: 36 responses, not 36 independent prompts.
Sources and SHA-256 identities remain in `campaign.json` and each extract.
Every measured value is retained in the v3 projection.

## Independent state and public admission

Validation describes contract satisfaction (`valid`). Evidence level describes
claim maturity (`E3`, controlled research). Publication describes admission under
that contract (`public`). Reproduction records the separate independent attempt
outcome (`not_attempted`). Valid and verified E2/E3 measurements are publicly
admissible under their exact contracts. E4 describes paper-grade evidence, not
ordinary public eligibility. `not_attempted` is reproduction provenance, not
invalidation. Independent reproduction is not inferred from checksums or E4;
`reproduced` requires its own recorded outcome and an E4 claim contract.

## Aggregation and limits

Decode is pooled sum(N−1)/sum(internal decode seconds), numerator 4,572 per
configuration, not a mean of rates. Overall output rate is
sum(N)/sum(internal response runtime), numerator 4,608, including first-token
wait. TTFT and complete response duration are arithmetic request means using
internal boundaries. Driver VRAM and process RSS are distinct whole-worker
peaks including preparation and warmup; VRAM also includes the shared desktop.
H2D is a sum over the 36 measured requests excluding warmup, not a transfer rate
or per-response quantity. Three block rates each pool twelve responses;
spread is 100 × (max−min block rate) / pooled configuration rate, a descriptive
range rather than a confidence interval. Controller residual time is a
subtraction rather than directly instrumented overhead.

OLMoE c30–c60 uses PRODUCT-core scalar; c64 uses Grouped T3. Qwen uses
top4-fused-v1. The c60-to-c64 difference cannot isolate residency from executor.
Models and executors remain separate groups; no cross-model speed ranking is
admitted. INT4 storage, 256-value blocks, FP16 scales, reconstructed FP16
execution values, and original checkpoint dtype retain distinct meanings.
Quality remains unevaluated. Isolated prefill is UNKNOWN. Common HTTP timing
for OLMoE is UNSUPPORTED. VRAM is not process-exclusive or decode-only.
The frozen GPU label is a campaign declaration; same-campaign telemetry records
12820938752 total driver VRAM bytes and 68472115200 system RAM bytes.

## Verification boundary

The local source gate checks the pinned summary, contract, report, per-prompt,
per-worker, source-hash manifest, frozen model preflight, catalog.py and worker-01
telemetry hashes. Exact numerical fields match summary cells, model identities
match frozen inputs, and counters/times/temperatures bind to the pinned summary.
RAM/VRAM totals bind to same-run telemetry. Metric, expert denominator, date,
executor, precision, and limitations are independently bound. The preceding
source-binding audit reconciled 667 declared hashes and aggregate arithmetic.
Token parity and the original full frozen-input validation are attributed to
the original consolidation, not rerun by the browser or download verifier.

The standalone v3 checker validates published hashes, thirteen aggregate copies,
and their four state dimensions. It neither repeats the experiment nor
authenticates the publisher. The unchanged original English report remains in
the edition; JSON downloads are curated aggregates. Runtime source is private
by author decision; the complete raw bundle is not included. See the v3 guide
served under compatibility filename `audit-method-v1.md` for the exact offline
command, all formulas, and the mutable publication-index boundary.


## Version 3 directory and publication preflight

This edition preserves the v1 and v2 releases unchanged and introduces no new
measurement or policy state. The v3 verifier enumerates the physical release
directory and requires exactly the nineteen manifest files plus
`manifest.json`, all regular files. Extra files, subdirectories, missing
entries, symlinks and other file types are rejected.

Before any release write, the v3 builder checks both the complete existing
destination inventory (when present) and the pending index candidate
`publication-index.json.v3-next`. A conflicting candidate or an incomplete,
extra or wrong-type destination fails without creating or completing a
release. Byte-identical complete releases remain idempotent.

The sole mutable discovery index advances from the exact admitted v2 baseline
to v3. It retains explicit historical manifest references to both v1 and v2,
as well as the eighteen existing flat URL records and classifications.