# Evolution 01 — From personal decision support to private Universes 18 September 2026 · CURRENT planning synthesis of new founder direction, with EXPERIMENTAL designs, HYPOTHESES and LONG-TERM research separately identified. This is not Round 31 and has not received the completed baseline's 30-round peer review. Read the [living master plan](/review/simsim-astro-master-plan/), [unchanged Round 30 baseline](/review/master-plan-round30-baseline/) and [proposed architecture](/documents/universe-architecture/). The original [first public edition](/editions/2026-09-15-first-public-archive/) remains available at its existing URLs. ## Sources and editorial treatment The founder supplied a conversation transcript and an accompanying integration brief on 18 September 2026. That is the integration date, not an inferred date for every conversation event. The brief ends mid-sentence after its request to challenge the architecture; no missing continuation is presumed. The transcript mixes founder ideas, private life reports and assistant suggestions. Only product-level concepts are synthesized publicly. Personal names, intimate accounts, copied messages and unsupported judgments about real individuals are excluded. Assistant-generated percentages, multiplier values, emotional scores and rollout dates are illustrations, not observations, commitments or validated science. The source identifiers and file fingerprints are recorded in the [source crosswalk](/review/source-crosswalk/); the raw private transcript is not published or used for training. ## What exists and what changes The public archive, reviewed documents, concept art, asset candidates and editing kits exist. This update changes documentation, discovery and archive versioning. No playable First Light release, personal world-model service, Room Mode, private-Universe bridge or learned predictor is established by this work. The 30-round baseline already contains weighted goals, distinct source status, scoped projections, simulated counterparts, optional symbolic lenses and recommendation-only experiments. New direction supplies a clearer responsibility split, explicit local-first ownership, event/outcome learning, multidimensional resonance metaphors, rooms, shadow-agent identity, temporary bridges and eventual joint simulation. It broadens the architecture without retrospectively claiming those features passed the earlier reviews. ## Dated decision register Each E01 entry belongs to this 18 September 2026 integration. Earlier A:/B:/reviewer: identifiers retain their original meanings. CURRENT means the present planning direction, not shipped software. ### E01-01 · Responsibility split — CURRENT direction, LONG-TERM learning **Before and reason:** R24–26 and baseline §§1/3 described Astro as a personal world model and simsim as its simulation laboratory, while keeping real and fictional histories separate. The compact framing connected the projects without assuming a validated common causal model. **New insight and reason for evolution:** The founder asks for a learned transition model underneath the conversational companion and a coordinator connecting private models. Separating these responsibilities makes future learning and provider replacement clearer. **Current:** LLM interprets and communicates; Astro is the companion/interface; the person's Universe holds permitted evidence and derived state; SimSim compares transitions; Master Brain coordinates permitted joint requests. This refines terminology rather than transferring ownership. Authored fictional transitions remain separate from empirical personal models. **Rejected / unresolved:** Reject a single omniscient memory or an LLM's narration as ground truth. Component boundaries, model accuracy and actual implementation stack remain unproved. ### E01-02 · Events and deletion — CURRENT foundation, EXPERIMENTAL schema **Before and reason:** R06/24/28 preserved causal save boundaries, source-attributed reports, corrections and deletion. G1 needed small reliable state, not a general event platform. **New insight and reason for evolution:** The transcript proposes immutable observations that survive model upgrades. Reinterpretable contextual events avoid hard-coding today's personality theories into tomorrow's model. **Current:** Retained events preserve source, time, context and uncertainty across model versions; views and hypotheses can change. Corrections are linked, and decision-time rationale remains distinguishable. Sensitive payloads and revealing derived copies remain deletable under the established controls. Public intellectual history is versioned separately from personal records. **Rejected / unresolved:** Reject “immutable means retained forever,” including retention disguised as provenance. Backup deletion, provider retention and training withdrawal need explicit designs. No universal event-store implementation is added to First Light. ### E01-03 · Storage and sovereignty — CURRENT product direction **Before and reason:** R08/10/20 protected selected-life exports, local usefulness and independent operation, but did not select a three-option storage hierarchy. **New insight and reason for evolution:** The founder explicitly chooses My Device, then My Cloud, then Astro Cloud. Ownership should remain understandable regardless of provider. **Current:** My Device is the default direction. Location, authorized recipient, reasoning depth, purpose and duration remain independent. Broad access by one's own Astro does not imply remote processing, third-party access, pooled training or publication. **Rejected / unresolved:** Reject mandatory Astro Cloud or bundled training consent. Encryption, key recovery, sync conflicts and support costs remain unimplemented; no cloud vendor is selected. ### E01-04 · Voice and Room Mode — EXPERIMENTAL, priority unresolved **Before and reason:** First Light's native visual experience was protected, with scripted assistance and R11–12 voice/provider boundaries. The personal pilot could remain manual and text-based to avoid infrastructure. **New insight and reason for evolution:** The founder questions whether the personal experience needs conventional graphics and proposes Astro becoming quieter and more useful around other humans. **Current:** Explore a voice-first personal surface with text/captions and useful optional visuals. Sleeping / Observing / Invited are explicit permission states, with a manual stop alongside “Astro, private.” Neither invitation nor one person's permission authorizes recording everyone nearby. **Rejected / unresolved:** Reject always-on capture by default and silent replacement of the authored First Light commission. The founder must choose resource priority before substantial parallel production. Low-graphics Astro can be tested independently; it does not require discarding existing artwork. ### E01-05 · Multiplayer foundations and shadows — CURRENT boundary, EXPERIMENTAL protocol **Before and reason:** R17/24–28 kept shared Circle authority, private projections and real-person consent separate. Multiplayer delivery remained later to protect G1. **New insight and reason for evolution:** “SimSim us” needs multiple private Astros and a coordinator from the conceptual outset. A single-player subjective model of someone must not later become that person's account. **Current:** Personal schemas distinguish owner, participating human, shadow and fictional actor. Each contribution carries scope and source status. A shadow is its owner's perspective; a real participant remains separate. Establish actor and permission boundaries from the first personal-model design, while deferring a network service until usefulness and consent checks justify it. **Rejected / unresolved:** Reject silent shadow-to-human merges, inferred identity and treating a shadow as another person's vote. G4's fictional Circle and a real-world planning room are distinct products. Do not force their participation rules to be identical. ### E01-06 · Temporary bridges — EXPERIMENTAL **Before and reason:** R24's Private / Sim-safe / Shareable / Public scopes prevented automatic disclosure across uses. **New insight and reason for evolution:** Particle / Signal / Constellation / World / Universe makes scope emotionally legible and supports limited-time coordination without handing over a database. **Current:** These are descriptive depth labels, not a replacement for exact permissions. A bridge names what, who, purpose, duration, processing location and allowed outputs. Prefer local queries and minimal derived answers. Expiry is enforced on controllable use; already disclosed information has no guaranteed recall. **Rejected / unresolved:** Reject “raw content hidden, therefore private” and a universal share-ALL checkbox. AI recipients and coordinators are processing parties too. Repeated-query inference, broad scopes, bystander information and consent under social pressure require tests. The one-hour example is not an installed timer or default. ### E01-07 · Frequency, rooms and recovery — HYPOTHESIS **Before and reason:** R25 separated symbolism from empirical claims and protected meaningful fiction. R26 prohibited invented confidence from imagined trajectories. **New insight and reason for evolution:** The transcript describes social context, authenticity, masking, residual effects, interaction depth and recovery. A room can model bidirectional influence rather than label one person as the source of all change. **Current:** Frequency and resonance are metaphors for revisable multidimensional states. Context, activity, duration and subjective experience matter. Separate immediate relief from delayed costs; unknowns stay unknown. Interaction depth and resonance cascades are candidate mechanisms, not fixed coefficients. Symbolic interpretations can remain valuable without predictive success. **Rejected / unresolved:** Reject literal human Hz, good/evil rankings, universal intimacy multipliers and meditation as the only recovery route. Do not convert a reported coping behavior into a recommendation. Measurement burden, confounding, emotional harm and the value of these metaphors remain open. ### E01-08 · Prediction, outcomes and model progression — LONG-TERM; small probe CURRENT **Before and reason:** R26 compared simple coaching with scenarios on one voluntary decision and distinguished qualitative accuracy from numerical calibration. A tiny trial avoids spending on scale before usefulness. **New insight and reason for evolution:** Preserve delayed outcomes and errors so later models can learn transitions, rather than indefinitely elaborate scripts. **Current:** Link state/context, decision, prospective prediction, actual action, immediate/delayed outcomes and error with goal/model versions. V0 rules/event graph → V1 statistical personalization → V2 outcome predictor → V3 sequence model → V4 multi-agent model → V5 foundation-model research is a possible evidence-gated progression, not a timetable. **Rejected / unresolved:** Reject automatic confidence gains from branch count, synthetic counterfactuals as observed training labels, and contextual probabilities unsupported by data. A permanent global trust score is rejected; even narrow follow-through estimates need appropriate observations, scope and evaluation. Sparse data, missingness, advice-induced behavior, drift and causal identification remain central limits. ### E01-09 · Communication and population learning — EXPERIMENTAL / LONG-TERM **Before and reason:** R24/27 required approved sources, minimal data and explicit execution authority. Personal reflection was allowed without giving the system ownership of other people. **New insight and reason for evolution:** Selected communication sources could create contextual events, while personal adaptation and shared priors might improve the model over time. **Current:** Start with voluntary conversation and low-burden follow-up. Future share/import mechanisms and supported connectors need actual provider/OS verification and scoped permission. Personal storage, inference, adaptation, shared simulation and population training are distinct grants. Data minimization remains preferable to collecting everything “just in case.” **Rejected / unresolved:** Reject covert cross-app access, treating user life as studio-owned training inventory, or assuming de-identification removes all risk. The claimed valuable dataset is a hypothesis about utility, not a present asset or training license. Training withdrawal and third-party rights may constrain which records can be used at all. ### E01-10 · Hubs, business and consent — LONG-TERM **Before and reason:** R17–22 separated gameplay, publishing, real action, truthful invitations and measured demand. **New insight and reason for evolution:** A shared simulation can help people after matching: evenings, trips, collaboration, teams and companies, not merely game encounters. **Current:** Hubs can connect permitted goals and offer introductions. Compare strengths, tensions, constraints and uncertainties, rather than rank human worth. Recommendations remain proposals. Each person accepts a real plan; contact, spending and booking need independent authority. **Rejected / unresolved:** Reject predicted attraction as consent or simulated compatibility as a hiring verdict. Whether a joint model helps more than a simple group discussion is untested. Silence and refusal cannot be interpreted as hidden agreement. ### E01-11 · Referral and network value — HYPOTHESIS, not an offer **Before and reason:** R19/22/23 separated subsidies, meaningful play, causal growth and recipient freedom; senders received no usage dashboard. **New insight and reason for evolution:** The founder suggests gifting three months of Advanced and receiving three months after a friend activates and meaningfully uses it. More participating friends might improve permitted joint planning. **Current:** Preserve this exact concept as a proposed experiment, without activating a promotion, billing or entitlement. Any offer needs funded limits, actual terms, abuse controls and separate voluntary measurement. The claim that each new participant improves value must be tested. **Rejected / unresolved:** Usage-conditioned rewards can disclose recipient behavior even without a dashboard. Do not promise both perfect silence and a reward that reveals activation. Before launch, choose an explicitly consented minimal completion signal or a non-usage-conditioned gift; the latter is the privacy-preserving recommendation, not an adopted replacement. No loss of promised recipient access for declining feedback, publicity or sharing. ### E01-12 · Current synthesis over preserved history — CURRENT archive behavior **Before and reason:** Published editions and 30 numbered rounds provided stable references and retained disagreement. The builder assumed a single first edition. **New insight and reason for evolution:** A living model direction needs a distinct current plan without rewriting the reviewed baseline or its snapshots. **Current:** Keep the existing master-plan URL as the living synthesis. Preserve the original baseline byte-for-byte at a named historical route and retain first-edition snapshots. Add this dated decision log and a new architecture proposal. Machine-readable discovery states which revision is current and which is historically reviewed. **Rejected / unresolved:** Reject calling this update “30-round reviewed,” rewriting earlier research as current fact, or equating successful website publication with game implementation. Further changes append a new evolution record and preserve the revision they supersede. ## Next bounded decisions and tests **NOW:** Keep First Light's W0–W6 and provisional human gate intact. Offer the separately authorized manual personal decision probe with only the minimum event/outcome fields. No new recording routine, multiplayer backend, billing or ML training is authorized by publishing the plan. **NEXT:** If personal scenarios earn their burden, test a small local event/projection prototype and a synthetic two-Astro permission exchange. An actual joint-planning study requires willing participants and its own narrow data scope. Preserve G2's independent evidence gate. **LATER:** Voice Room Mode, temporary bridges, hubs and business simulations require demonstrated coordination value, intelligible permissions, funded compute and realistic provider behavior. Chapter / Circle / creator / hosted-world gates remain distinct. **RESEARCH:** Predictive and causal validity; leakage through derived answers; personalized latent states; future training rights; privacy-preserving computation; referral economics; attention cost. Research stages may stop without invalidating private artistic play. Unrun acceptance scenarios: correction and deletion across views/backups; replay without duplicate events; a revoked bridge with delayed output; repeated-query disclosure; shadow/participant identity collision; Room Mode with a bystander and failed voice stop; conditional referral privacy; simple-coaching comparison with immediate and delayed outcomes. Synthetic success cannot establish human usefulness. The next founder decision is a bounded effort allocation between First Light and the voice-first personal probe. No requirement to build two full products follows from sharing conceptual architecture. R04 emotional payoff, R13 paid-assistance dissent, Astro's predictive limits and the NOW/NEXT resource disagreement remain unresolved.