RPG OS v0.9.1

RPG OS v0.9.1 — universal agent basis and prospective deepening

This experimental release refines independent-agent establishment so an NPC, creature, machine or collective is portrayed from its own supported nature, source of control, applicable drives/directives, information and circumstances rather than from the player character's interpretation or a universal human personality template.

It also closes a residual late-deepening loophole found during a synthetic playtest: once outward behavior has been portrayed, a later private detail must not be invented as the hidden pre-existing cause of that behavior merely because it fits the player's interpretation, or because it contradicts that interpretation and therefore appears "independent." Unexplained prior causes remain unresolved unless recovered from an existing authority or independently established through an accepted procedure using admissible inputs.

Added and changed

• Entity-appropriate behavioral basis: independence from the player is distinct from individual autonomy. People, creatures, machines, externally controlled bodies and collectives use only dimensions that actually apply to them.
• Control and information separation: shared control does not imply shared knowledge. Local perception, controller knowledge, communication channels, delays and local discretion remain distinct.
• Interpretation boundary: the PC's unverified interpretation, an unperceived player wish, a proposed narrative role and a desired success or refusal are not evidence of private nature. Actual observations constrain possible private facts without automatically establishing their perceived meaning.
• Inapplicable dimensions stay inapplicable: a relationship, emotion, social capability or other dimension is not treated as an unknown or zero score merely because a human character would have it.
• Prospective deepening: new evidence may change current beliefs, aims, orders or orientation from that point forward. It does not rewrite that change backward into an earlier motive or hidden history.
• Past-cause protection: an earlier private cause may be established later only from an already authoritative source or an accepted generation method whose admissible inputs do not include the player's interpretation or a need to explain the old portrayal. If several causes remain compatible and no independent determination is required, the cause stays unresolved.
• No contrarian laundering: choosing the opposite of the player's interpretation is not proof of independence and cannot be used as a hidden-cause generator.
• Sparse ownership preserved: no personality spreadsheet, controller schema, cast-wide simulation, mandatory startup read, private scratch layer or automatic save is introduced. Shared state keeps one existing owner.
• Compatibility preserved: existing characters, controllers, relationships, history and generation results are not regenerated or relabeled when adopting v0.9.1.

Regression coverage

TOOLS/test_agent_state.py now checks the universal input boundary, control/information separation, inapplicable dimensions, sparse ownership, persistence limits and the prospective-deepening rule. ADMIN/TEST_AGENT_STATE.md contains fifteen synthetic behavioral cases, including the new U15 case for an observed presentation followed by a player interpretation and a later attempted backstory explanation.

The behavioral cases remain evidence protocols rather than automatic LLM simulations. Passing text/route checks does not prove unbiased generation or hidden pre-establishment. A claimed private determination is not proof that it existed earlier.

The existing structural, handover, packaging, exact-source, scoped-search and evidence suites remain part of CI, together with the agent-instruction suite.

Preserved

v0.9.1 keeps the v0.9.0 save, agreement, archive, person/system and handover formats. The standing d6 fallback remains optional and unchanged. Existing source-access, search and evidence-audit tools remain available. Ordinary PLAY remains read-only; unsaved private state is best-effort until an authorized save or checkpoint succeeds.

Compatible v0.7.3, v0.8.0, v0.8.1, v8.1.1 and v0.9.0 campaigns can adopt the program update in a protected separate copy through ADMIN/UPGRADE_V08.md. The update itself does not change an existing agreement, engine selection, campaign facts or generation permissions.

Use it

Quick start · Installation · Independent agent state · Targeted agent cases · Verification

This remains an experimental release. Structural and instruction checks do not establish long-campaign, cross-model or hidden-state reliability.