# Release quality and evidence — 11 November 2026

## What this release changes

A live anonymous-browser check reproduced six gaps in the existing public lab: prior results survived changed input; prior success survived a refused input; an empty row selector was silently parsed as zero; the completed receiver release was not linked from the home page; quiz scores survived changed answers; and the batch receipt did not explicitly name both evaluation-bank hashes.

The release also corrects a reproduced 390-pixel viewport overflow caused by the grid inheriting a wide table’s intrinsic width. The table remains horizontally scrollable within its own container; content is not hidden. The failed pre-publication check is retained.

The release clears stale results, holds mismatched delayed results, rejects empty selectors, resets changed practice answers, adds navigation between the existing journeys, and binds the exported batch receipt to model and dataset hashes. All frozen model checkpoints, input banks and previously published numerical results are preserved. This is software/documentation quality work, not new model training or a new capability result.

## How to use the lab

Start in the [single-model lab](../index.html). Its first comparison is a retained synthetic row. Enter custom features for an unlabelled prediction, or use **Load row** to compare against a retained target. Editing either input clears the old result. Invalid input is refused; **Export receipt** stays disabled until a new completed result exists.

The [six-checkpoint page](../research-v03/index.html) runs all six candidates over two fixed banks. The numerical fields are population-fraction errors shown as percentage points—not chatbot answer accuracy. Export a receipt and run the offline verifier below. In the practice quiz, changing an answer clears the old score; the points are local practice, not authenticated Academy XP or permission.

The [receiver release page](../integration-specialist-review/presentation-v1/index.html) is a public evidence report. Its private workspace link requires owner access. Public access to the report does not expose private records or grant execution rights.

## Verify a downloaded receipt

After downloading the repository, use Node 22 or later:

```sh
node release-quality/verify-receipt.mjs /path/to/exported-receipt.json
```

An accepted verification means **the checksum and declared artifact pins match** the expected local research references. A self-created checksum is not a signature. The verifier does not establish who produced the receipt, independently verify source truth, rerun the model, prove factual correctness or grant authority. It refuses modified output, substituted dataset/model pins and promoted-authority flags. Full model replay is separate.

## Evidence and coverage

The [claim register](claims.json) connects the main public claims to exact retained artifacts and their limitations. The [baseline browser record](baseline-browser.json) preserves the reproduced gaps. [Contract tests](contract-tests.json) exercise request sequencing, row validation and receipt boundaries. [Retained replay checks](retained-replay-checks.json) rerun the original numerical tests without replacing their historical result files. The completed hosted journey record is linked from this guide after publication.

Scope is explicit: these pages, the stated interactions and the pinned evidence artifacts. A finite test pass is not a proof of 100% correctness across all browsers, all inputs, all providers or the entire MEOK platform. No full accessibility or security certification is claimed. Keyboard entry, visible errors, live status messages, small-screen overflow and download/reopen paths are included in the journey checks; assistive-technology testing remains separate.

## Content and research references

Numerical claims are checked against their original machine-readable reports and rerun code. The old 34-axis roster is a named retained snapshot; its 11 relevance/23 not-tested split is not 34 new formal measurements. The 100-candidate inventory is earlier metadata/README screening, not 100 completed source audits. The original Cosmos LoRA experiment remains separate from this process predictor and is not promoted.

Primary references rechecked for this pass: Hugging Face's static-Space documentation and mutable, non-versioned bucket semantics; W3C's error-identification and status-message guidance; Anthropic's 2 November 2026 FlashProve research release; the StrataFlux primary abstract dated 5 November 2026; and the cache-divergence author's pinned source README. These are supporting research—not reproduced experiments or new training material. The paper landing pages were intermittently unavailable to one retrieval path; that is not evidence that the papers were disproven or deleted.

References: https://huggingface.co/docs/hub/spaces-sdks-static ; https://huggingface.co/docs/hub/storage-buckets ; https://www.w3.org/WAI/WCAG22/Understanding/error-identification.html ; https://www.w3.org/WAI/WCAG22/Understanding/status-messages.html ; https://www.anthropic.com/research/provable-inference ; https://arxiv.org/abs/2611.08363 ; https://github.com/aditi-p31/cache-divergence-study/tree/resubmission-2026-10 .

## Onward gates

The receiver's [dependency-ordered plan](../integration-specialist-review/presentation-v1/next-steps.json) remains the plan, not a completed-implementation ledger. Next priority is trusted host binding of current source/context, then one authenticated harmless human approval-to-effect journey. Capability improvements require separate development material and fresh evaluation, while real serving-cost claims require measured loading, transfer, decoding, cache state, failures and quality. Proof integration, source-owner review and formal per-axis admission remain separate gates. No new task schedule is created by this release.

## Concurrent work reconciled

An exact-parent check stopped the first attempted publication before any write. The combined candidate adopts the other lane's published root UI generation guard and blank-row refusal, its `release-20261111/` report, the separate reproduction-integrity repair and the mobile grid containment. Our independently reproduced findings and earlier candidate tests remain historical; they are not claimed as hosted acceptance of the combined head. Dataset-bound receipts, quiz invalidation, accessible grouping, offline receipt checks and post-publication download verification are this lane's additional scope. All new upstream paths and canonical scientific artifacts are preserved.

## Completed post-publication acceptance

The hosted public app completed **49 checks** at 1280×900 and 390×844, including **six actual receipt downloads that were reopened and verified offline**. The run checked current-result clearing, invalid-input recovery, both replay journeys, grading/regrading, navigation, keyboard movement, altered model bytes, a missing manifest and a delayed digest. Normal journeys reported no JavaScript errors. See [hosted acceptance summary](hosted-acceptance.json), [served-file readback](served-assets.json), [retained interruptions](interruptions.json), and the [completed release record](release.json).

The original detailed hosted JSON and download files remain on the M4 qualification workspace. That device timed out during the subsequent copy, so the public hosted record is explicitly a summary of the completed tool output—not a copy of local-run receipt hashes. Final evidence publication and served-file checks used the other existing Mac. No new paid machine, package installation, model training or private-data read was needed.

Two interrupted attempts are preserved: one hit local disk exhaustion, and one measured a static page before the hosted stylesheet loaded. After disk space recovered and the layout readiness check was corrected, the unchanged layout bound passed. Root README rendering is also classified separately from raw-byte code/data verification. These distinctions are not hidden by the successful final result.

Concurrent root/UI, reproduction-integrity, mobile and arithmetic publications were retained. The ten execution/data files used for the public acceptance remained byte-identical to the tested origin while newer release documentation was added. The private application's latest metadata was observed only; this suite does not certify its saved-copy or arithmetic end-user workflows.

## Numerical verification beyond checksum matching

The new [numerical replay checker](index.html#numerical-replay) recalculates supported receipt outputs and metrics using a fixed twelve-artifact snapshot. It rejects a deliberately false result even after its checksum is recomputed. Read the [method and limits](NUMERICAL_REPLAY.md) and [executed checks](numerical-replay-evidence/qualification.json).

Offline: `node release-quality/replay-receipt-cli.mjs /path/to/exported-receipt.json`. The original checksum-only verifier remains unchanged and is still not a numerical replay. Neither checker grants identity, approval, current source authentication or GSPC admission.

## Actual file import and portable recheck

The existing numerical checker now supports local receipt-file selection, unsigned verification-bundle export/reimport and an offline bounded batch runner. See [the file round-trip method](FILE_ROUNDTRIP.md) and [dated acceptance status](file-roundtrip-evidence/release.json). A bundle is only accepted after recomputation, not because its checksum or success flag looks valid. The earlier full M4 hosted report has also been recovered; its original six file hashes match, without turning those historical checks into new tests.


## Cross-platform execution

The [executed M4/Kaggle/Cloudflare release](../cross-platform-20261111/index.html) adds a separate Python replay and static delivery snapshot. Kaggle version 2 completed with matching results; version 1 failed before computation and is retained. This does not change the existing numerical checker or private runtime.