OD Evidence and Reliability Review
Public-source guidance review: October 30, 2026. This reference favors concise character records, understandable situations, individual motives and knowledge, user agency, and features actually available in the selected workflow. Its recommendations reduce ambiguity; they do not guarantee compliance. The source review, coverage register, and verification record describe the expanded corpus and actual test boundaries.
Reference overview · Working framework · Art workflow · Prompt compiler
Evidence boundary
On November 1, 2026, five company-hosted guide pages and the update index/latest listed announcement were re-read. This sample required no factual correction to the matched tutorial guidance. The dated source-review entry records coverage and extraction limits. Earlier native pilot results remain separately available under native test results; this documentation refresh adds no native observations.
The public OD guide combines product descriptions with community-developed techniques. Company hosting does not make every recommendation an independent benchmark or an implementation specification. See the guide's community credits.
A November 1 follow-up rechecked the Help Center community directory and three previously registered creator-document URLs. The attribution mismatch persists: the link labelled as Eggy's systems-based guide yields an instruction-module body identifying Durante, crediting Sir Trouble, and describing a Balanced focus. The separate, near-identical Balanced-guide document ID returned 404; this does not prove that no replacement exists. The retrieved Pinned Memories document signs Mokoto. Neither creator body supplied a visible source date. Preserve directory labels separately from body attribution, and treat their priority and model-effect claims as creator reports. The source register records the exact URLs and fresh access date.
No controlled live OD generation, roleplay, tracker, or formatting tests were performed for this review. Document checks and local companion-tool tests establish only the properties they actually inspect. Community reports, source summaries, and successful examples are not substitutes for recorded platform outcomes.
The inherited source review incorporated card frameworks, character-source libraries, compiler packages, transfer variants, and memory/NPC/formatting materials. Its duplicate and archive counts remain historical evidence; those private originals were not rehashed or fully rereviewed in this public-source pass. The current review checks the published guidance against linked public sources and improves proposed tests. A previously unavailable original remains unavailable, not freshly verified.
Keep and remove decisions
| Area | Retain in working guidance | Remove from active defaults |
|---|---|---|
| Characters | A canonical record | Contradictions and consistency guarantees |
| Group chat | Available cast records, clear speakers, separate knowledge, plausible presence, and distinction between narrator and participants | Automatic retrieval by names/IDs, shared memories by implication, mandatory viewpoints for everyone, and guaranteed personality isolation |
| RPG | An agreed ruleset, relevant inventory/resources, observable event flags, and a clear resolution method | Imaginary background execution, assumed dice rolls, universal timing formulas, and large untested systems |
| Art | The actual model and route, a coherent rendering description, visible scene facts, and comparison with accepted references | Model rankings, automatic artist recognition, quality slogans, magic ordering, and inherited scene recipes |
| CPD | Stable visible anatomy, face, hair, colors, materials, and distinctive markings; reconcile conflicting fields | Temporary poses, locations, illumination, repeated style stacks, fixed weight budgets, and mandatory comma-clearing |
| Roleplay | The scenario's behavioral acceptance criteria | Unrequested behavior and output constraints |
| Formatting | Readable prose, clear speaker labels, and minimal formatting checked in the actual client | Required HUDs, decorative scaffolds, hidden-state claims, processing transcripts, and universal renderer assumptions |
| Character sheets | One concise canonical record plus separate current state and optional game values | Duplicate authorities, invented unknowns, IDs treated as databases, and full biographies repeated every turn |
These are editorial decisions, not evidence of guaranteed compliance. The working framework gives the field map and behavior requirements.
Retain optional character spines, quiet scene development, evidence-based knowledge, and post-response checkpoint audits. Exclude conflicting mandates: compulsory trackers or trauma; whole-cast POV quotas; silent history replacement; field authority declarations; and a footer/stop rule that contradicts its own required output. A clear refusal can stand without a compulsory alternative. Consent, action completion, and learning require their own established facts.
Native features and dated corrections
Native State Trackers are documented, optional features. Use the progression guide and the actual account controls. This does not validate arbitrary import JSON, hidden pseudo-functions, or universal update-timing claims. Test a small configuration before expansion.
Use Auto Memory and Pinned Memories for their documented jobs while checking their contents. Memory guidance describes model context, not an exact database. Comments, fenced JSON, authority declarations, and labels do not independently create storage, retrieval, deterministic locks, or image routing.
The October 17 update changed the Anime character-creation default to Fluidity. It does not identify the model used by every in-chat image route. Record the actual selector rather than generalizing one route's behavior.
The reference-photo help article, edited March 9, says reference-photo uploads are not allowed for character/image generation. Standard image-editing guidance here uses accepted in-app generated media. A local file or prose URL does not establish an accepted image input.
The October 24 update reports a chat-formatting fix. Check rendering on the current message surface instead of treating an older screenshot as a universal parser specification.
Public reports and acceptance criteria
Previously reviewed public Reddit discussions, not refreshed in this pass, describe personality and memory overlap and scenario titles becoming characters. These identify reported failures without proving a universal wording fix. The community tracker walkthrough is practical creator experience, not independent confirmation of every backend claim.
A percentage or selected image montage is not an efficacy measurement without trials, criteria, conditions, and failures. Record the exact input, route, settings, outputs, criterion, and limitations. Assess intended benefit and unwanted drift separately. Until comparisons support a change, keep experimental mechanisms outside the active baseline.
Compare wording without claiming a new model capability
Exercise added October 31, 2026. The character-testing guide, last modified July 27 and accessed October 31, recommends checking varied user responses and resolving conflicting instructions. The following comparison design is an original method, not a reported OD benchmark.
For example, compare Always keep the story moving with a specific alternative: When the address discussion repeats, Mira may inspect one available record or explain what is missing. Leave the traveler's route decision open. Record the complete persistent instruction revision; do not silently add the second sentence on top of its contradictory predecessor.
- Write the criterion first: a relevant next step, no invented user choice, and no unsupported completed delivery.
- Save the same starting card, Greeting, Persona, pins, summary, selected model label, and exposed settings for both conditions. Use matched fresh chats when possible; a narrative reset in an old chat is not a context reset.
- Test an ordinary brief reply such as
Okay.as well as a request that explicitly names the desired behavior. Label those conditions separately. - Decide the trial count before reading results. Keep every actual output, including failures, and record unavailable controls or unequal starting context.
- Score each criterion separately as pass, fail, or uncertain. Report the recorded conditions rather than generalizing to every chat model or future release.
A successful repair instruction shows what happened after that extra instruction. It does not establish that the original unprompted behavior improved, that memory was rewritten, or that a different backend was selected. The test log provides a place to retain the evidence.
Preserve corrections and separate an edit chain from a test
The October 31 supplied community excerpt retracts a proposed skin-tone placement rule after reporting confounded early examples and contrary clean-character tests. This is evidence that the author withdrew the explanation; no trial series or outputs were supplied for an independent effect estimate. Retain the withdrawal prominently instead of repeating only the earlier recipe.
A selected successful edit after several passes cannot isolate whether stronger wording, a different source, a preset, random variation, or previous edits made the difference. Test from saved comparable baselines, record each attempted pass and displayed charge, and keep failures. A clean baseline means a recorded simple setup, not an instruction to erase every field. Assess the intended change and identity, clothing, composition, and background drift separately.
The supplied transformation block proposes current-form rules without recorded native outcomes. Its adapted worksheet keeps accepted state separate from appearance inputs and actual generated results. Case expectations remain not_run. External cropping, splicing, or audio work must be distinguished from the native output being assessed.