# How to audit publication edition v3 Edition: `technical-preview-residency39-v3`. Campaign: `residency-executors-39-20260911-01/consolidation-01`. Authored 2026-09-15. This edition adds publication metadata to the same thirteen configuration results; it introduces no new measurement. ## Policy and independent dimensions Authority: the Horizon repository's `docs/superpowers/specs/2026-09-15-contract-valid-publication-policy-design.md`, the owner-approved Contract-Valid Public Evidence Policy dated 2026-09-15. The recorded dimensions have independent meanings: - `validationStatus: valid` records satisfaction of the frozen contract gates. - `evidenceLevel: E3` records controlled research and its exploratory claim scope. - `publicationStatus: public` admits the result under that exact contract. - `reproductionStatus: not_attempted` records independent reproduction provenance. Valid, verified E2 and E3 quantitative results are publicly admissible under their exact contract and permitted claim. E4 denotes paper-grade maturity; publication does not promote the level. `not_attempted` is provenance, not invalidation of the recorded observations. A claim of independent reproduction requires both its separate `reproduced` status and E4; E4 alone implies no such outcome. This edition records E3 and independent reproduction not attempted. ## Files and compatibility names The manifest names nineteen files: thirteen curated configuration extracts, the campaign aggregate dataset, the unchanged original report, three v3 explanatory documents, and the standalone v3 verifier. With the manifest itself, the release contains twenty files. The manifest has no self-hash. Existing consumers keep these download names inside the v3 directory: `audit-method-v1.md` contains this v3 audit guide, `evidence-audit.md` contains the v3 audit note, `horizon-technical-notes.md` contains the v3 technical notes, and `verify-evidence-edition.mjs` contains the v3 verifier. These are compatibility filenames, not evidence-version declarations. The manifest schema and edition ID identify v3. Authoring sources use the distinct `*-v3` filenames. `/evidence/publication-index.json` is a mutable discovery pointer outside the immutable release. It identifies v3 as current and explicitly retains v1 and v2 as historical immutable editions. Its eighteen prior flat URL entries retain their recorded hashes and v1 targets; former current aliases are historical aliases. The two older `*-t3-p17-128-mean.json` / `*-t4-p17-128-mean.json` selections remain historical extracts outside the thirteen-configuration matrix. No earlier file has been moved or reinterpreted. The builder accepts only exact baseline v2 index bytes or exact expected v3 index bytes. The admitted v2-to-v3 transition uses an atomic same-directory replacement; an identical v3 rebuild is read-only. Any unexpected index bytes are rejected, including a missing index. ## Download and verify Download `manifest.json` and every named file from `/evidence/releases/technical-preview-residency39-v3/` into a new directory, preserving the filenames. Review downloaded code before executing it. Node.js 22.19.0 and its standard library suffice; this procedure needs no npm dependencies, runtime checkout, model weights, GPU, or external connection. From the downloaded directory, run: ```text node verify-evidence-edition.mjs --directory . ``` The exact successful output is: ```text PASS: technical-preview-residency39-v3; hashes and aggregate copies verified; public E3 scope preserved; independent reproduction not attempted. ``` The checker recomputes every listed file size and SHA-256, requires exactly thirteen unique admitted result IDs, compares each complete extract result to the campaign dataset, and validates all four independent dimensions. It rejects `reproduced` below E4 and any metadata outside this edition's E3/not_attempted scope. It rejects missing files, unsafe paths, duplicate inventories, altered bytes, and mismatched copies, returning a nonzero exit status. This verifies download integrity and aggregate consistency against the supplied manifest. It does not recompute raw observations, verify token-ID parity, authenticate the publisher, reconstruct the original environment, or repeat the experiment. A matching manifest and its downloaded files are not a signature. At build time, `loadPublicEvidence` additionally binds the v3 source projection to frozen local campaign sources and their exact aggregate values; that local source validation is distinct from the standalone download check. ## Selection and measurement formulas The matrix contains 39 workers and thirteen configurations: five each for OLMoE 0924 and OLMoE 0125, and three for Qwen. Each configuration repeats the same twelve prompt IDs over three worker blocks: 36 measured responses, each of 128 tokens. The campaign has 468 measured responses and 39 excluded 64-token warmups. Repetition does not create 36 independent prompts per configuration. Selectors retain model/revision, capacity per layer, arm, source paths, hashes, precision, hardware, and the fixed run contract. Source paths identify provenance without promising public source access. | 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, including 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; not per-response traffic or a 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 | Pooled decode is not an arithmetic mean of block or per-response rates. Controller residual time is a subtraction, not directly instrumented overhead. Isolated prefill is UNKNOWN. OLMoE HTTP observations are UNSUPPORTED; common timing views use internal boundaries and do not establish uniform HTTP/SSE latency. VRAM is neither process-exclusive nor decode-only memory. OLMoE c30–c60 uses PRODUCT-core scalar; c64 uses Grouped T3. Qwen uses top4-fused-v1. OLMoE c60-to-c64 changes executor and residency together and cannot identify a residency-only causal effect. Unlike models are not a speed ranking. Token repetition is not an output-quality evaluation. INT4 source, 256-value blocks, FP16 scales, reconstructed execution values, and original checkpoint values remain separate concepts. ## Source access and reproduction The runtime source is private by author decision. The full raw bundle is not included in this edition. The author's profile provides a contact route and does not establish source availability. Independent reproduction has not been attempted. The procedure above remains usable with the public files alone, within its download-integrity and aggregate-consistency scope. ## 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.