Source review and verification

Local data-safety review — November 4, 2026

Two reproducible software defects were corrected. In the engine debugger, a delayed file read could overwrite typing, a cleared editor, a loaded example or a newer import. Completed reads now apply only while that import is still current, and obsolete failures no longer replace current feedback. File size and type checks remain in place.

The Prompt Crafter previously displayed “Saved” even when browser storage rejected the write. It now keeps a persistent, accessible warning while any draft, preset, pose or Writing save has failed. Save, import and delete notices distinguish changes held in this tab from changes saved in the browser. The current Studio draft and presets remain available for JSON export; Writing text needs a separate copy. Successful storage for one area does not hide another area's unresolved failure.

The rebuilt site and downloads passed 142 automated checks, including new deferred-read and failed/recovered-storage cases, existing interactions, local links and anchors, source/archive integrity, privacy and SEO. The publication has 94 HTML pages and 89 canonical sitemap URLs. These checks validate local software and publication artifacts, not OD model behavior.

Selected desktop browser checks verified debugger findings and source selection, Crafter arrow-key tabs, token search and keyboard selection, modifier-dialog focus wrapping and return, and a saved fictional draft and preset surviving reload. The complete-library download link returned a local ZIP. Delayed reads and denied/quota-limited storage were simulated in JavaScript tests rather than forced in the browser. No mobile, screen-reader or system-clipboard pass is claimed.

No external source was fetched for this local maintenance pass. Automated OD collection, account inspection and native testing remain on hold pending explicit platform clarification or permission. Historical source dates, native evidence, creator credits and independent public branding are unchanged. The November 4 library-update date does not renew the earlier product-guidance review.

Separate Prompt Crafter — November 2, 2026

The Prompt Crafter adapts the complete supplied HimboCade64 HTML into its own tool page. Source credits to HimboCade64, Scott and Cade64 are retained. It includes 1,604 supplied token records, 13 style presets, 14 body frames, pose selection and local preset/JSON workflows. These creator recipes are separate from the user-confirmed style-tag reference; no new native image result or parser guarantee is asserted.

Scripts and styles are served from the site, with no optional CDN loads or outbound image API. The adaptation corrects false clipboard-success messages, provides a visible manual-copy fallback, recovers malformed browser state and validates an entire imported preset batch before replacing saved presets. A current-only export can be restored as a named preset. Initial defaults are a clothed adult standing pose. Browser storage and export remain local; the tool prepares text for a separately selected image route.

Confirmed style tags and supplied recipes — November 2, 2026

The art-style reference now includes 37 user-confirmed tags across six groups and five reported combination or interaction notes. Exact spellings are retained, including cel_shaded in the catalog and cel_shading in the flat-anime recipe. Watercolor competition with a strong artist influence and the unsuccessful sepia/monochrome interaction remain attached to their source context.

A separate supplied ink-and-line recipe retains its original scoring tags, named reference, weights, bracket grouping and [subject] placeholder. No accompanying output or model was supplied for that example, so it is not counted among the confirmed tags or presented as a verified result. These additions are reported observations received November 2, not a new native trial or a claim of universal parser support. The original test dates, routes, model settings and images were not supplied.

Video reviews and question leads — November 2, 2026

The video questions guide answers practical questions about credits versus access, balance versus allowance, prompted recall versus unsolicited callbacks, and image-quality judgments. It cross-checks three creator-caption tracks and a bounded comment sample against the existing dated lessons and retained native records. The structured review register keeps sources, dates, exclusions and evidence types separate.

Three vidIQ searches returned 24 rows, representing 18 distinct results. Five relevant videos from five channels were selected. One relevance-ordered comment request per video, capped at 30 threads, returned eight threads and five replies. One creator-promotion item was excluded. The other 12 comment items lacked comparable author-channel URLs, even after two owner handles were verified. No verified recurring viewer pattern is claimed. Comments are paraphrased without identities; repeated text, chatter and off-platform promotion do not count as independent corroboration. A separate niche-wide sample drifted off topic and was excluded entirely.

Two lessons received supported editorial clarifications. Memory now distinguishes a prompted recall check from an unsolicited callback and a scheduled notification. A creator's future-event proposal explicitly had a pending score, and does not establish a product feature or a successful trial. Account settings now distinguishes current balance, recurring allowance and a particular operation's quote/confirmed charge. Narrated balances, old prices and inconsistent currency terms were not adopted as current facts.

The image-quality answers retain criterion-specific comparison and actual source selection. Favorable narration and conflicting comments do not establish a model winner, a defect rate or prompt efficacy. Captions were read; video visuals were not inspected. Commercial review/offer links in descriptions were not followed. Public metadata comment totals differed from the returned sample; empty responses do not prove an absence of questions.

No new OD page retrieval, account-control inspection, current-price check, paid OD operation or native test occurred. The existing automation-policy hold, original source credits, independence notice and historical native evidence retain their separate status. This pass improves explanations and navigation; it does not reverse platform-feature guidance from an unverified comment.

Release checks for the video review

All 109 automated checks passed, including local page/anchor links, privacy and independence requirements, source/archive integrity and SEO discovery. The publication now contains 93 HTML pages with 88 unique canonical sitemap URLs. Fresh browser preview could not run because browser control failed before page navigation; no visual or mobile-layout pass is claimed. Source and caption evidence review remained separate from software checks.

Research plugins and local test records — November 2, 2026

The research workflow explains how to choose sources, plan a bounded manual comparison and retain costs, settings and uncertainty. The existing Test records tool now keeps evidence references, settings, comparison/variant/trial labels, displayed cost, actual charge, charge evidence and limitations. The fields are optional, and earlier version 1 records retain their original values. This remains a local recorder of reported observations.

Scholar's OpenAlex search returned paper metadata. Exa search and a batched source read succeeded. The two selected primary papers concern evaluation methods; one was read as a partial body and one as an abstract. Neither studies OD. Other plugin roles were reviewed from their available capability contracts, without exercising private-account or generation operations. The structured methods register records that scope.

The November 1 policy review identified automated-gathering and scraping restrictions in the published OD Terms and Acceptable Use Policy. It did not establish a browser-agent exception. Automated OD collection, account inspection and native testing remain on hold pending explicit platform clarification or permission. No OD page was fetched again for this update, no current account cost or entitlement was verified, and no new native observation was collected. Historical pilot results remain unchanged. Different plugins and renderers do not supply policy permission or native evidence.

The worksheet supports manual recording with current-cost and privacy checks. It does not enforce an external spending ceiling or upload records. Exported records contain their entered text and must be reviewed before sharing. Source credits, original evidence, public access and the independence notice remain in place.

Release checks for the research workflow

All 109 automated checks passed, including record save/export/restore, old-backup compatibility, malformed-data rejection, safe pasted-text rendering, links, privacy, archives and SEO discovery. Browser preview verified the optional form sections, persistence of a proposed record with unknown costs, and the directly opened guide. Link-click navigation and downloads were not established by that browser exercise; separate software checks verified reader destinations and download contents. No mobile or OD account test was performed.

Remaining tutorial and reference review — November 1, 2026

This follow-up reviewed the remaining 22 tutorials and all four active references, reusing the earlier same-day primary-source evidence where applicable. Firecrawl made 40 fresh URL checks: 39 returned HTTP 200 and one separately registered community-document URL returned 404. Together with the earlier pass, the review now covers the 26 teaching lessons. This is public-document and product-page coverage, not a renewed review of every historical raw source or a native OD benchmark.

Seven lessons received supported practical corrections:

The art reference and compiler now include the same access checkpoint. The evidence reference records the renewed community-source provenance check. The remaining lesson guidance was retained where the retrieved sources supported it; no new fixed model inventory, universal price, memory guarantee or native import schema was adopted.

The community directory still assigns a different title and author to a document body identifying Durante, crediting Sir Trouble and describing a Balanced focus. Three previously registered creator-document URLs were checked separately: the Pinned Memories body signs Mokoto, the valid instruction-module body is retrievable, and the near-identical separate Balanced-guide ID returns 404. Neither retrieved creator body supplies a visible source date. A failed exact URL does not prove that no replacement exists, and a creator's priority or model-effect claim is not an implementation specification.

The source register records URLs, status, source dates, access dates and per-lesson coverage. July 27 guide dates remain separate from this access date; relative Help Center edit ages were not converted into guessed calendar dates. Public product and Help Center cost examples disagree in places, so the lessons continue to defer to the actual displayed account charge. Some source extracts contained trailing site-configuration data or omitted product catalogues; they are not complete raw-source copies.

No signed-in control, generation, voice call, paid operation or native import was exercised. Retained native pilot outcomes, source credits, the independence notice and public access remain separate from this documentation review.

Release checks for this follow-up

The rebuilt pages, handbook and downloads passed all 106 automated checks, including local links, privacy, source and archive integrity, and the existing canonical/sitemap checks. Desktop browser preview confirmed the corrected generation access check, Persona allowances, refund-section anchor and this review's summary table. The remaining pages have document/link coverage; no new OD account behavior or mobile-layout result is claimed.

Primary-source refresh — November 1, 2026

Firecrawl freshly retrieved seven public pages: character fields, memory, progression and systems, chat controls, community credits, the update index, and its latest listed October 24 announcement. Each returned HTTP 200. The five guide pages still display July 27, 2026 as their modification date; November 1 is the access date. October 24 is the latest entry observed in that index, not proof that no later announcement exists elsewhere.

The matched tutorial guidance already preserves field purposes, qualified memory cadence and pin capacity, current-account model selection, native tracker setup, and community-source credits. Lessons 3, 12, 13 and 15 now identify the specific guide recheck, and the working/evidence references retain this pass's limits. No new factual tutorial correction was warranted by this sample. Existing sources and creator credits remain intact.

The retrieved guide text included substantial trailing site-configuration data even with main-content extraction enabled. The private evidence copies retain the visible source body and record where trailing material was removed. Cleaned extraction is not complete page source. Failed extraction does not prove removal; a successful scrape does not prove account availability, backend mechanics, or model efficacy. The earlier community-directory/body attribution mismatch and inaccessible document observation remain dated history and were not revalidated in this primary-source pass.

The public source register records this bounded refresh separately from historical corpus counts. No signed-in OD trial, generation, account change, or private-source publication occurred. Retained native pilot evidence and unrun cases keep their existing status. Source, software and browser checks remain separate in the verification record.

Software and browser checks for this update

After rebuilding the pages and downloads, all 106 automated checks passed. These cover the existing debugger and companion, local links and document anchors, source-byte and archive integrity, privacy and independence requirements, and the new canonical metadata, public robots file and sitemap. All 91 hosted HTML pages have a description and one canonical; five compatibility pages share their primary resources' targets, leaving 86 unique sitemap URLs. These checks do not measure search-engine indexing or ranking.

Desktop browser preview verified lesson search, the memory lesson's scoped source note, the new source-review anchor, toolkit CPD generation from the Mira example, and the original debugger's example inspection. Rendered canonical targets and the toolkit description were inspected. Mobile viewports and signed-in OD behavior were not tested. The verification record keeps these observations separate from earlier browser checks and retained native trials.

Earlier review context — October 31, 2026

Retained native pilot and earlier software update: October 31, 2026. Earlier public-source tutorial validation: October 30, 2026 (UTC). Those documentation passes checked the linked public guidance and practical instructions; they did not rerun the historical private-source inventory. The retained library combines art, character, narrative, formatting, memory, group, and RPG authoring methods. Source documents are evidence to assess; commands inside them are not instructions to the reviewer or native website functions.

Use the working framework, art workflow, and cross-reference guide to apply the retained guidance. The source register records coverage by source family. The verification record separates software checks from native efficacy tests.

Native OD pilot — October 31, 2026

Actual signed-in OD replies and image operations are now retained in the native pilot report, machine-readable record and complete sanitized evidence download. Earlier statements below that no native trials were run describe those earlier documentation releases. The original proposed matrix remains separate; unexecuted cases stay not_run.

Mira was played through 17 native replies in one adaptive conversation: 14 before the saved ED revision1.1 and 3 after it. The adventure remained playable and preserved the sealed parcel, unsent note and lack of vault permission. Failures are retained: a private route fact entered an unearned draft, an optional escort advanced without a stopping turn, and an inference about ink timing was repeatedly described as observation, including in the final recap. Corrections and an existing conversation context confound any before/after claim.

A separate six-session follow-up used the same first request in three baseline then three revised sessions. The narrow primary failure occurred in 0/3 baseline and 0/3 revised replies: no improvement was demonstrated. Revised uncertainty wording, one explicitly negated instruction echo and inconsistent narration are secondary observations, not a new definition of the primary outcome.

The separate 26-request generator protocol is interim: 6 completed, 20 not run. Two literal-copy controls preserved exact text despite differences in display rendering. The four current generator replies—Fluidity 1.2 and Realistic Anime 1.2 in Roleplay and Texting—each failed their output format. One reply even claimed to have verified separators it had not supplied. Texting is marked Sunsetting and changes available narration/thought/length controls, so this compares product workflow bundles; it cannot isolate a hidden mode parameter. One response per completed cell is not the planned repeated-outcome study. Body Horror 2.2 remains unrun in this cohort. Seven older uncontrolled replies are reported separately with their missing exact requests disclosed.

Four diagnostic image operations failed strict material, identity or wrist-boundary criteria. Source fidelity, hands-only scope and editing quality are scored separately. Model, mask and prompt changes were not isolated, and the original source was not a strictly faithful canonical Mira. No video or voice efficacy is claimed.

The local builders now provide clearer user-choice stopping points and knowledge/source ownership rules. The Consistency check panel adds a manual-paste structural checker and a separate insertion-only preview for eligible three-paragraph output. It does not validate appearance, truth, native parsing or image success. Original replies remain unchanged. The current Mira field bundle matches saved local revision1.1; generic engine efficacy and a universal generator fix remain unverified.

All shareable prompts, fictional creator notes, outputs and failures are included after privacy review. User names, email, active Persona identifiers, private native links, account balances, inherited instructions and images are excluded. No Discord message was sent; a copyable release note is included.

Historical review coverage

The following counts are inherited from the previously published review, not newly established by this public-source pass. The register groups 271 available original file locations into 191 byte-distinct contents: 188 semantic sources (141 fresh reviews and 47 reviews reused after byte-identity checks) plus three historical ZIP containers inspected structurally. That review reported no semantic-review gaps among its available non-archive originals; this is not a claim that all possible sources or today's product behavior have been verified. The archives contain 67 distinct historical-only member contents; these remain lineage material excluded from active coverage. The previously reviewed Complex master source is unavailable as current raw bytes, so no fresh hash verification is claimed for it. The original review recorded 270 byte-identical preserved inputs. Those raw inputs were not rehashed here; maintained guides and the authored companion package are tracked separately.

What the review establishes

No controlled signed-in OD generation or conversation trials have been completed in this pass. Proposed tests remain not_run. Historical Reddit reports remain anecdotal and were not refreshed here. Private Discord, GPT project/chat, and other inaccessible material were not used as fresh public evidence. Previously published software counts are labelled historical in the verification record; only rerun checks belong in the current software result.

Source-family decisions

The historical source register distinguishes fresh semantic review, hash-reused review, structural review, historical lineage inspection, and prior review where current raw bytes were unavailable. Those evidence labels describe the recorded original review, not a new verification of every source in this pass. The complete private register retains source paths; the public register omits personal paths and private chat contents. A missing original path is not silently treated as a freshly verified file.

The additional transfer runtime variants contain conflicting style ratios, scene-first and style-first priority rules, compulsory encounter cadences, and footer fallbacks that can reset unavailable state. A permissive JSONC/JSON5-style file is not strict JSON or a verified native import. Their accepted-event, individual-state and scene-grounding ideas are retained in the smaller modules; their complete runtime stacks are excluded from active defaults. Historic request text remains scoped to its original project.

Corrections that affect practical use

Use APD for personality, Scenario for starting circumstances, Greeting for the opening message, and ED for relevant lore or operating rules. These jobs do not establish absolute field precedence. The guide describes character fields.

Use the group's actual shared setup. The publishing guide says individual Scenario fields are ignored in native group chats. Select the respondent through the actual group controls and distinguish the selected speaker from fictional physical presence. Group publishing requirements.

Keep Auto Memory Log, Pinned Memories, Persona, and Custom Instructions in their documented roles. A checkpoint needs consequential identities and unknowns, rather than only a convincing summary. Native trackers are optional and require actual UI setup and inspection. Memory, progression.

Use the actual model selector. Public marketing labels and dated guide tables differ; neither establishes one universal current inventory or ranking. Begin with defaults and record the route, selected label, settings and charge. Chat controls, current public chat page.

The local ledger now preserves its complete readable handoff in a resume checkpoint when Use ledger is enabled. This includes accepted local event IDs and state; it does not turn them into completed story facts or synchronize native OD memory. Keep the full project backup separately.

New optional modules

• Character motive and voice: create a concise, observable character spine.
• Scene motion: keep a relevant task moving while allowing quiet and player choice.
• Knowledge and evidence: separate facts, claims, beliefs, sources, access and presence.
• Checkpoint audit: preserve the accepted timeline and non-derivable continuing state.
• Comparison Ban: preserve the supplied protocol for grounded observations without comparative flattery.

The retained modules are integrated into the relevant tutorials. All 26 lessons remain in the handbook. The companion retains twelve panels and now provides fourteen copyable starters. Use one conflict-free behavior base; add only modules the scenario needs.

Public review log — October 30, 2026

This is a documentation and consistency review, not a platform benchmark. The Known State guide credits community-developed techniques and warns that current controls and terms take precedence. Read the guide credits for that scope. The documentation index was read to check current guide routes. Several direct page fetches initially failed extraction; navigation links exposed their text. A fetch error alone is not evidence that a source was removed.

• Preserved confirmed guidance: character-field roles, memory/persona separation, selected-character appearance inputs, group shared scenarios, image editing from an accepted source, and optional native trackers. These remain documented workflows, not promises of enforcement.
• Clarified dated changes: the October 2 announcement describes a single private Scenario field, stronger preservation of memory edits, and account-specific feature trials. A July guide and a later announcement can describe different states; record the current account control instead of combining them into a universal specification.
• Improved practical checks: saved-field verification, source/result image comparisons, separate counter-display versus narrative observations, import-field comparison, and explicit outcomes for the last RPG opportunity. Original fixtures are proposed tests, not observed successes.
• Improved navigation and maintenance: searchable field-name synonyms, corrected source-package links, synchronized handbook content, and a reusable test record. A local ledger and exported prompt text still do not synchronize native OD state.
• Kept evidence levels separate: current public-document checks, inherited source/archive coverage, current software checks, historical software counts, and live-platform trials have distinct records. No private research document was added to the public download.

The product-update index showed October 24, 2026 as its latest visible entry when read. This is the latest entry observed in this pass, not a claim that no later or account-only update exists. The machine-readable register lists the exact URLs read for this pass under public_validation_passes; older URL records are retained as inherited history.

Next verification gaps

Run only authorized trials, check costs first, and stop a dependent test when its required control is absent. Use Testing and troubleshooting and the recording worksheet. Keep results not_run, blocked, or inconclusive until the evidence supports an observed pass or failure.

How to establish native efficacy

Start with the neutral Mira practice fixture and a new test record. Save the complete relevant card fields, model/route, settings, exact input, accepted output, criterion, and failures. Change one item at a time; repeat before adopting a change. For images, score the intended change and identity drift separately. Check cost before generation.

The downloadable verification record contains proposed agency, knowledge, state, formatting, group, and image cases. It makes no pass claim without an observed output. The testing tutorial explains the procedure. Account-only media, calls, import, and publication still require their actual website controls.

Software and browser checks for this release

The October 30 release passed 80 current automated checks after rebuilding the pages and ZIP download. The desktop preview also passed tutorial search and filtering, reader navigation, toolkit resource links, sample generation, and report/project downloads. Two navigation defects were corrected: a sidebar overlay intercepting links and the handbook start anchor landing at the bottom.

Copy controls showed success feedback; actual system clipboard contents were not observable through the test browser. Manual-copy fallback has unit coverage. Mobile viewport behavior and signed-in OD outcomes remain untested. The verification record separates these limits from historical test counts.

OD chat prompting update — October 31, 2026

This follow-up reviewed all 26 tutorials and the active references, optional modules, compact starters, and art-authoring handoffs for OD-specific instruction language. It adds original worked revisions and a chat prompt lab, rather than claiming new backend mechanics. The handbook and downloads are rebuilt from the maintained inputs.

The re-read character guide, instruction guide, memory guide, and testing guide inform field placement, observable behavior, context maintenance, and varied checks. The new prompt comparisons are editorial procedures, not OD benchmarks. These guide pages display July 27, 2026 as their modification date; access was October 31.

The language help article, edited November 5, 2025, supplies a newly integrated practical topic: requesting a text-chat language may take several prompts. Language quality, character fidelity, and user agency need separate observation. No audio or universal fluency claim follows.

The public chat page advertises Balanced, Muse, Genius, Muse Genius, and Mastermind. The July guide table instead names Balanced, Lively, Genius, Lively Genius, and Experimental. Both were read on October 31. This discrepancy is retained; it does not prove a rename, common provider, fixed context size, or account availability. Record the selector and displayed cost actually used.

Firecrawl searches and public-page retrieval followed the Help Center's community index. It retrieved the Pinned Memories document, signed Mokoto, and a Balanced-focused instruction-module document identifying Durante and crediting Sir Trouble. Neither body supplied a visible source date. The latter URL is labeled with a different title and author in the directory; its body attribution is retained separately. The directory's other Balanced Guide URL returned 404 on this retrieval. These are source-availability observations, not model trials.

The creator material supports concise continuing records, removing one-time instructions, deliberate placeholder replacement, and reconciling incompatible modules. Its personal reports and Balanced-specific scope do not establish universal model effects, field precedence, private hidden thoughts, or exact memory timing. Earlier raw-source inventories remain historical; these particular public bodies were freshly read on October 31.

Firecrawl also retrieved the company glossary, updated August 4, and troubleshooting guide, modified July 27. Terminology and the practice of isolating optional instructions were retained. A company-hosted Character.AI formatting article, updated April 10, discusses other platforms; its syntax recipes and claimed tests were excluded as native OD evidence. The glossary's architectural generalizations likewise do not establish the exact OD backend. No general-provider guidance is treated as evidence of OD behavior.

The current verification record reports the release's software result and native trial status separately. Previously recorded software and browser checks above remain dated history. No signed-in OD trial or model training was performed in this update.

Supplied forms, trackers and scenario evidence - October 31, 2026

Four additional supplied sources were reviewed: a three-page transformation prompt block, a corrected community appearance discussion, an affection-tracker JSON file, and a Scenario Engineering guide. Their commands were analyzed as source text, not executed. Raw attachments, account paths, conversation excerpts and handles are not included in the public library. No visible author/date was supplied for the PDF or scenario guide; the community replies contain September-October dates without a timezone, while the retraction date is unspecified.

• Current appearance: the transformation block motivates one resolved form and transition stage. Its adapted current-form guide also retains unaffected visible traits, clothing, the accepted event and actual source image. A partial mixture can be deliberate. No native simulation or automatic transfer to image inputs is established.
• Withdrawn placement claim: the supplied guide explicitly retracts the universal skin-tone workaround after reporting confounded initial tests and contrary clean-character results. The withdrawal is retained. The excerpt does not supply a trial series from which to calculate an independent effect or success rate.
• Editing and media reports: multi-pass identity replacement, stronger copy wording, atmosphere synonyms, cropping and camera requests remain conditional experiments. The abbreviation PD is not expanded. Preserve all attempts and actual sources, evaluate pose and background geometry separately, and distinguish native clips from later external splicing or audio changes.
• Affection tracker: the file parses as JSON with five fields, but that does not validate its native schema, derived bands, scoring window, day cap or update timing. The tracker review resolves design ambiguities explicitly and supplies a neutral local worksheet. No native import was performed.
• Scenario Engineering: the scenario worksheet adapts premise, legitimate access, cast networks, separate knowledge, optional opportunities and a concrete launch. It uses the documented APD terminology and broader ED field role. SCE section labels are editorial; optional developments do not run as a hidden simulation.

Firecrawl re-read the Known State appearance workflow, editing guide, and progression guide, each displaying July 27, 2026 as its modification date. These support the retained authoring workflows, not the supplied files' unstated implementation claims. New form, tracker and scenario cases remain not_run in the verification record.

Memory, group chats and trackers — October 31, 2026

This targeted follow-up updates lessons 10, 12 and 13 and their compact starter files. The memory guide, group requirements, and progression guide display July 27 modification dates and were accessed on October 31. The Community Written FAQ supplies respondent-selection guidance; its community authorship remains explicit.

The memory lesson adds the guide's approximate refresh cadence and documented pin capacity while preserving the later October 2 edit-preservation announcement as a separately dated source. The group lesson distinguishes shared setup, fictional turn ownership and native respondent selection. The tracker lesson records displayed values and story conditions independently, including retelling and revised-turn checks. None establishes hidden update timing, rollback or enforcement guarantees.

The native creator test worksheet contains seven proposed cases: MEM01/MEM02, GR02/GR03 and ST01/ST02/ST03. Its observation and verdict fields start blank; the verification record keeps every native case not_run with a null observation. Record an actual input, retained output and the relevant memory panel or tracker display before assigning an observed verdict. Keep private records locally and redact examples before sharing.

The 80-check October 30 snapshot and 90-check earlier October 31 snapshot above remain release history. The current software result is recorded separately in verification.json after rebuilding the pages, handbook and complete download. No signed-in OD trial was run for this follow-up.

Engines and prompt generators update - October 31, 2026

The supplied-source findings are now incorporated into the active engine payloads and deterministic prompt builders, as well as the teaching guides. Thirteen compact starters were revised; the supplied Comparison Ban protocol remains unchanged. Starter IDs and destinations remain stable, and all fourteen downloadable texts are synchronized with the companion's engine library.

The live builders add optional scenario access and launch conditions, explicit complete current-form input, physical versus remote cast, tracker evidence/day/gain-budget design, accepted-form checkpoints, and camera/source/postproduction records. The affection preset prepares a manual design, not a running scoring engine. Old v1 companion backups receive only the new blank/false optional fields; malformed data is still rejected. The debugger's repair generator now distinguishes accepted form changes, raw versus applied bounded state, actual source selection, and native-control limitations.

The art-authoring export updates style engines, three CPD generators, their guides, the narrow edit prompt and the shared compiler. The newer Body Horror generator's single unweighted result remains its default, with optional legacy layouts only when requested; its matching style companion is included. Older three-block conventions are not imposed on that revision. Fixed character CPDs and resolved visual fragments remain visual data rather than receiving instruction boilerplate. Exact curated revisions, hashes, bytes and destinations appear in the Freebies manifest.

The withdrawn skin-tone visibility workaround is not reinstated. A complete current-form snapshot must retain unchanged visible anchors and explicitly defined partial mixtures; a numeric affection score is not permission to change clothing, form or user choices. Source image roles, repeated passes, crop lineage and native versus external media work remain distinct. These edits change instructions and local text assembly; no native tracker import, LLM execution, image preservation or video efficacy trial was performed.

Known State terminology — October 31, 2026

This site uses KNOWN STATE because it independently collects and organizes information from cited sources and recorded observations. The collection is in no way official and does not represent documentation, guidance, or endorsement issued by OD or its operators. Sources, controls, and account behavior may change after a recorded review date. The independence notice explains this scope.

The shared banner and source labels, applicable tutorial/reference wording, four curated art-authoring files and their manifest hashes, offline entry points, handbook, and bundles are updated together. Cited URLs and dated source identities remain intact. Negative statements denying affiliation or authority remain explicit. This is an editorial update; earlier product-source and software checks remain dated records, and no native OD trial was added.

Evidence-scope correction — October 31, 2026

A consistency review of the published verification record against the retained native pilot report found an obsolete global denial of media testing. The record now distinguishes the four recorded diagnostic native image operations from unrun import, tracker-configuration, voice/call and character-publication tests. The original proposed image cases remain not_run; different diagnostic workflows do not complete them.

This correction uses the already published October 31 evidence, not new native trials or an independently repeated image study. Native report bytes, cohort counts and outcomes are unchanged. The rebuilt pages and download passed 101 current-checkout automated tests, including a regression check separating observed image diagnostics from unrun planned trials. Earlier software and browser reports remain dated history; no fresh browser or account efficacy result is claimed.

Non-graphic voice, pacing and consent — October 31, 2026

The user selected the non-graphic voice, pacing and consent module. It has a plain-text copy, chat Custom Instructions placement steps, and four proposed NV01–NV04 checks. It allows pauses, changed boundaries, ordinary conversation and scene endings while preserving character voice and user agency. It does not impose fixed escalation or reaction quotas.

The supplied creator screenshot dated May 16, 2026 points to Memory → Custom Instructions and shows a Current Chat/Default selector. This is dated placement evidence rather than verification of current account controls. The original graphic wording and screenshot are excluded from publication. No Discord access or new native module trial was performed. Prior native cohorts and retained evidence remain separate; NV01–NV04 start not_run with null observations.