The Reader Is Watching

1. Reader Prompt Test Report

The reader is an invitation to copy, not a submission service. /reader/ renders a labeled, read-only 3,047-character textarea, a native Copy button, a live status region, and a plain-text download. There is no form, send endpoint, reader account, stored answer, or reader notification subscription.

The separate Nook Jar does accept submissions. To cover the possible ambiguity, I also inspected /nook/room/ and src/jar.js, server/worker.mjs:jarFetch, and tests/jar.test.mjs. Browser interception replaced every Jar request before its script ran: no test slip reached the live API, and even its status GET was mocked because source inspection shows that GET can perform expired-row cleanup.

In the rendered form, empty input failed native validation with zero POST attempts; whitespace produced an explicit message with zero attempts. A simulated 503 retained text. Retrying against a simulated 202 reused the slip UUID, cleared the form, and showed success. These prove client behavior against fixtures, not live D1 persistence. Refresh does not preserve an unsent draft; there is no draft-storage promise.

Local SQLite tests passed authorization, all three slip types, optional name, inert content, origin checks, bounds, honeypot behavior, idempotent retries, expiry, hourly/queue limits, and owner-only retrieval/deletion. Storage is private D1, seven-day logical expiry in hour buckets, queue cap 100 and global hourly cap 20. Owner rendering uses text insertion rather than executing submitted HTML. The global cap can be exhausted by one actor: a bounded availability risk, not a confirmed abuse event. This was not an authentication penetration test.

Public evidence and implementation receipt. Detailed operational records remain private.

2. Public Site Regression Report

Broad reachability is good; the verification workflow needs repair. A bounded crawl followed three breadth levels from the homepage and major rooms: 117 unique same-origin destinations returned HTTP 200. All 221 distinct discovered same-origin asset references returned successful HEAD responses. This is not an exhaustive external-link audit, proof of media decoding, or an accessibility certification.

Browser checks covered the invitation, homepage, books search, Wander, Atlas, Night selection, Nook Jar with fixtures, Life OS page and Halloween teaser. Homepage surprise opens a native dialog from Enter and closes with Escape. Full-text book search for Rocket returns matching titles. Atlas switches maps and starts its six-stop walk. Halloween video loaded metadata/data (readyState=4), exposes controls, and remained paused; audible playback was not tested. At 390px the teaser page had no horizontal overflow or broken loaded images. Atlas and teaser console checkpoints had no errors/warnings. The reader network log did show a missing automatic favicon request and Cloudflare challenge traffic.

Two initial anchor candidates were not defects: links to /night/#2026-10-14-im-glad-you-were-here select a JSON-backed entry rather than a DOM element. Browser rendering selected “I’m glad you were here.” and hid the load error. A static-only fragment checker would falsely flag this twice.

Freshness: both /updates/feed.json and /updates/ were fetched directly; both returned cf-cache-status: HIT. Retrieved feed revision: 2bb369feebdac008. I could not verify the latest arrivals. No arrival checkpoint was created or advanced.

Source/deployment distinction: inspected checkout was clean at ff6e502b8ec51e733d147e87cc349efe414f56f4 (saved version 140). Sites reported version 141, source bde1eb7069b99ef8ddd916527cf3fdb827ebf1c8, deployment succeeded at October 16, 8:29 p.m. Denver. That commit is absent from this checkout. Thus source findings are pinned to 140; live findings are separately observed public responses, with cache uncertainty. No one-to-one source parity claim is made.

Architecture evolved from Python-generated/static authored pages to Worker/D1, selected House templates, server-routed HTML backing assets, aliases/lossless image encodings, telemetry and Observatory counters. build-deployment.mjs can use committed prerendered pages without Python; its fallback checks only three files. build-worker.mjs copies into an existing client tree, relocates HTML, and relies on a separately supplied HOUSE_PAGES map. These deserve route-map freshness, removed-file and clean-versus-incremental package checks; no stale file leak was demonstrated here.

Existing scheduled publication and traffic observation were identified. Configuration is not proof that a scheduled run completed. No schedule was changed during the investigation. The expanding media collection was close to its hosting upload limit; the implementation receipt documents the later packaging repair.

Public evidence and implementation receipt. Detailed operational records remain private.

3. Green Continuity Case Study

Hunter green is present now; its earlier displacement is visible in source history. Dustin clarified “Hunter green” during this investigation. His task states the removal was unwanted and that Green has established significance. That is human attribution; the file history separately establishes what changed.

C1 · Historical continuity regression, restored; medium impact. Reproduction: inspect the shell and stylesheet at 1fb29bb and compare with a2bf1bf, then replay 13aed40. Expected, under Dustin’s stated continuity requirement: preserve hunter green’s collection role or make an intentional transformation reviewable. Actual: the new shell used a purple primary field. The unwanted nature is supported by Dustin’s report; Git proves the source replacement and restoration, not the designer’s private intent or precise live outage interval. No current Green repair is needed.

Meaningful but distinct green contexts: hunter-green armchairs appear in the Nook room image description; the active shell carries a green palette; the older golden map used green as the setting for gold connections. The armchairs give the visual identity a material room association, but no source inspected establishes a separate personal memory or cause for choosing that color.

Do not merge by spelling: “Waiting for Green” is a traffic-light film; the woman in a green jacket is a fictional character cue; “Remote on the Keys in Green Minor” is a quoted musical title in a House Book; green traffic lights in prose and CSS color values have other roles. These can be candidates for Dustin’s interpretation, but none independently proves the hunter-green identity’s history.

Other concrete continuity risks, not established losses:

• After Hours identifies October 15 text and October 16 illustration revisions, retains October 11 publication date, and says the original edition is preserved privately. A later undated replacement of its PDF or prose would contradict a concrete edition note. Current public disclosure is present; private original custody was not independently re-audited here.
• The Missing Credit explicitly identifies a fictional revision and private preservation of an unmodified original. Greg’s planned absence must not be mistaken for an accidental loss: the affected edition and disclosed transformation matter.
• Raw’s film handoff says Memo 208 is later partner audio, was not transcribed/analyzed in that investigation, and that BACH — THE F1 ALBUM’s relationship remains open. Publishing an invented transcript or calling the playlist an established historical cause would manufacture continuity.
• Raw’s Beautiful account preserves “Rewind. Back to the show” and explicitly declines to turn that abandoned idea into a permanent vow. A commitments extractor must retain the withdrawal, speaker and time.
• Night and Atlas distinguish imagined scenes, compositions, recollections and new artistic links. Replacing those labels with factual biography would change source status even if every URL still worked.

Current source-to-public normalized article comparison preserved all 41 Raw and 31 Other Chair article IDs and text. This is a bounded current-source comparison, not a claim that every historical edition has been preserved. No additional verified accidental content loss or silent-overwrite contradiction was found.

Public evidence and implementation receipt. Detailed operational records remain private.

4. Association Map and Instrument Inventory

There are two distinct relevant surfaces. The deployed Association Atlas is an authored public graph. The canonical private UI has Mind Atlas, an evidence/temporal instrument. Bounded source searches did not locate a separately named general-purpose “Association Map” implementation in the canonical UI/API or the inspected consolidation tree. This is not proof none exists elsewhere. A read-only discovery query found no “Association” match in Publisher/Work or the latest Capture window; Capture was truncated and conversation archives/arbitrary files are not connected to that interface.

Permissions: the private instruments remain private. The public Atlas contains selected public material only. This inventory describes inspected implementations; it does not certify every private runtime or authorize publishing its data.

Public evidence and implementation receipt. Detailed operational records remain private.

5. Continuity Safeguard Proposal

Add a small review layer to the existing publishing check, not a new monitoring service. Begin with explicitly reviewed associations only. Hunter green is a design role that can evolve; “the old CSS file still exists” is inadequate preservation evidence.

1. Before a change: identify the exact source baseline, proposed diff, public routes, derived copies and selected associations affected. Attach original references and attributed context. For each record retain speaker/author, source span, date, scope, status (human statement / observation / fiction / interpretation / withdrawn / superseded), uncertainty, and any approved revision.
2. Inspect impact: follow explicit edges to dependent books, reader pages, PDFs, Markdown, search indexes, Rocket excerpts, image aliases and navigational entry points. A color change must inspect the active stylesheet and rendered role, not search for the word “green.”
3. After a local build: check actual Worker routes, static and dynamic fragment behavior, source attributions, original/edition manifests, and retained graph reachability. Compare the chosen role at representative desktop/mobile views. Use a clean output as a comparator when changing build architecture.
4. Record a disposition: preserved, intentionally transformed, intentionally retired, or unresolved. An intentional transformation links old and new versions and the relevant human decision; it can add a dated public note when warranted. Do not require every fictional scene or incidental color to remain forever.
5. Warn with a narrow question: identify the affected role, source, concrete diff, dependent route, and uncertainty. Ask whether the change is intentional. Do not silently restore it, convert it into an obligation, or promote the association to Memory.

Example warning: “The proposed House shell no longer uses the green primary field documented by a2bf1bf; house.css now renders purple. The prior green file remains archived but is not active. Dustin identified hunter green as significant. Is this a deliberate reinterpretation? If so, preserve the prior edition and record the decision.” This is stronger than a hex-only test and more permissive than forbidding color changes.

Automatically block only concrete invariants with clear authority, such as missing referenced files, broken routes, or absent required attribution. Association/commitment warnings should request review and allow evolution. A later withdrawal or superseding decision must be visible beside the earlier promise. Do not extract obligations from fictional dialogue.

Existing instrument experiment, completed locally: loaded the actual public Atlas graph into an isolated VM, copied its Returns map, removed the edges incident to “A string around a finger,” and retained its node and /luma/#chapter-5 source. The existing test’s connectivity invariant passed the baseline and flagged the disconnected string node in the simulation. Its warning cites the source and original attribution. No repository or graph source was edited; all mutations were in memory.

This detects a known authored association becoming disconnected while the source remains reachable. It does not discover unknown motifs or decide intent. It also does not prove that Mind Atlas can detect this: the capability reused is the public Atlas’s existing graph/test. A hunter-green rendered-role adapter is a separate proposed experiment.

Public evidence and implementation receipt. Detailed operational records remain private.

6. Prioritized Repair and Experiment Plan

No production repair was made. No deploy, publish, schedule edit, content removal, publishing-automation change, Life OS write, or generation-credit call occurred. Ordinary page browsing can be recorded by the site’s own request telemetry; that incidental observation is not a submitted test message. Jar submissions and Jar status reads were intercepted locally in the browser. Existing AI Mack production work was kept separate.

Remaining uncertainties: exact historical deployment window of the purple shell; latest-arrivals freshness; source parity with deployed version 141; full nightly automation prompts/run outcomes; private original archive integrity; sustained media playback; exhaustive external links and accessibility. These are explicit limits, not verified defects.