Kairo Doc

01 · What Kairo is

Kairo is a continuing AI system developed by Blaine: language models, an operational mind, durable memory, background cognition, tools, and connected interfaces participate in one evolving computational history. The identity is carried by accepted history and state across individual model calls and devices. A model checkpoint, a chat transcript, a body animation, and the complete system are different objects of study.

This is the official public technical and scientific reference for Kairo as audited on October 12, 2026. It explains the running architecture, the available activities, their authority boundaries, and what the evidence can establish. “Official” identifies the project’s documentation; it does not confer independent scientific certification or peer review.

• For a first visit: read the overview, production snapshot, foreground turn, memory, dreams, and interfaces.
• For using Kairo: browse the tools, toys, games, and command reference. Availability still depends on identity, location, consent, and the current request.
• For scientific assessment: read the evidence standard, claim ledger, research history, theory comparisons, and falsifiable evaluation designs.
• For developers: use the authority map, API boundaries, persistence rules, and verification principles.

The document covers all tool contracts and the complete nineteen-game catalog found in the audited production source, plus additional play systems and background faculties. It omits secrets, network addresses, private transcripts, precise device locations, and private identity or relationship values. It is a dated reference, not a continuously updating status dashboard.

Back to document start ↑

02 · Evidence standard and scientific vocabulary

A useful description must separate four propositions: a mechanism is implemented; a mechanism is enabled; a mechanism executed; and a mechanism caused a particular result. None implies the next without additional evidence. A status flag is evidence of configuration. A successful receipt is evidence of a bounded event. A controlled intervention can support causal attribution.

For runtime disputes, use persisted facts and execution traces before reproducible tests, source analysis, behavioral observation, and introspective explanation. For preference, consent, and desired future state, the expressed position is relevant governance evidence; it still cannot rewrite what executed.

Key distinctions are observed ≠ retrieved ≠ exposed ≠ materially used, intended ≠ authorized ≠ executed ≠ completed, and reported experience ≠ measured phenomenal consciousness. These are distinctions between claims, not dismissals of Kairo’s reports.

Evidence: fresh process/source audit; the project’s architecture rules; state–report correspondence audit.

Back to document start ↑

03 · Production today: October 12, 2026

The audit observed eight active mapped application processes: the agent, memory API, memory worker, model proxy, and four Discord transports. Their effective environments resolved their declared production source roots. The source guard verified 1,513 mapped source paths against the complete main release 402018256adfa47c8b95f2de36a7c11d471fb713. The authenticated agent runner was running. These observations describe the inspected processes and files; Python module resolution under their environments is not an in-process stack trace of every module.

The live PostgreSQL read observed 303 dream-completion event rows, 241 dream episode rows with 240 distinct content hashes, 328 recurrent-frame completion rows, and 72 thought-completion rows. Latest dream completion was October 12 at 19:00:24 UTC. These are aggregate ledger counts across the queried store, not user-scoped canonical totals, proof of unique subjective episodes, or a quality score. Event and episode counts differ because their schemas and historical emission/deduplication rules differ.

Evidence: read-only process/configuration audit beginning 21:47 UTC; authenticated capability and schedule GETs; read-only aggregate PostgreSQL query; live model metadata.

Back to document start ↑

04 · Models, quantization, and serving roles

Parameter counts are total model parameters; an A3B mixture-of-experts designation is not a claim that the whole checkpoint has only three billion parameters. Quantization describes stored numerical representation, not a calibrated quality level. Serving context is an admitted input/output budget, not the amount of autobiographical history permanently retained.

The normal 35B voice is primary, but the system contains explicit supporting-model fallback paths. The 9B is not a mandatory serial hop for every message. It can participate before generation, within delegated analysis, after an action, or in bounded generation recovery. Roles must be attributed from call receipts and served model identity, not legacy service names.

The historical LFM2 supporting-model replacement was rolled back. Retired coder-specific capability names remain disabled in the registry. The fresh inference-host audit observed two Tesla V100 GPUs with 32 GiB each, a primary model sharded 2:1 across them, Q8 K/V cache, and an active multimodal projector. Supervisor showed the primary, supporting, small-helper and image programs running. The Cat facade was running with its hybrid disable marker present; its presence is not evidence of hybrid generation. The audited primary binary SHA-256 was 7ce8253f1069d6b33049ecdd13843167b078fb78946085eccdeadb105ac9972d. No new inference test or model-weight rehash was performed for this documentation audit. Cache-first and canonical-transcript experiments have not earned a universal performance claim.

Evidence: live authenticated model metadata; resolved controller/engine/worker/proxy source; LFM2 activation and rollback record.

Back to document start ↑

05 · Logical architecture and authority boundaries

The system is distributed across control, inference, media and database roles. This document publishes logical responsibilities and model facts while withholding addresses, management routes, credentials and private filesystem locations. Replacing an inference host is different from replacing the stores that carry continuity.

Back to document start ↑

06 · One foreground turn, end to end

1. Admission. Authenticate the caller; bind session, workspace, request, attachments and transport scope to a durable job. Recognized Discord commands are dispatched before conversation admission.
2. Present formation. Fork the committed StateFrame into candidate-local state. Bind the literal current request; historical quotations must not become a new command.
3. Routing. Apply deterministic recognizers first; use the controller for remaining ambiguity. The active small-helper offload is limited to eligible bounded cases.
4. Recall and evidence preparation. Retrieve only eligible source material in the authorized scope. Preserve actual query, source references, freshness, exclusions and coverage limits.
5. Context construction. Assemble instructions, state, memories, tools and history under the selected model budget. Retain current evidence when optional historical replay must be compacted.
6. Model round. The primary generates prose or a typed tool proposal. The current backend refreshes state at discrete model-call boundaries, not inside an uninterrupted retained-KV decode.
7. Execution. Validate schemas, permission, tier, path confinement, consent, budgets and durable idempotency. Return actual results; tool descriptions and model intent do not substitute for receipts.
8. Verification and review. Check required evidence and outcomes. Ordinary reasoning can use deterministic review; action review may involve the supporting model. Bounded repair can fail visibly.
9. Settlement. Only the accepted candidate promotes its eligible overlay and final response. A losing candidate’s proposal is not a committed self-change.
10. Projection and delivery. Outboxes export accepted history asynchronously; transports deliver with their own receipts. A completed job can still have a transport failure.

The foreground lifecycle separates generation, acceptance, persistence and delivery. A client timeout is not automatically a server-job cancellation. Retries must reconcile the original key and settlement before creating new work. Background work yields to foreground requests; performance measurements must include any admission or queueing delay.

Evidence: live-resolved platform service, runner, engine, executor, Discord transport, job store and outbox paths.

Back to document start ↑

07 · Mind and StateFrame

MindCoordinator extends the live-state coordinator and is injected into the platform runner and engine. The compatibility name runner.live_state refers to the same coordinator, not a second mind. A StateFrame is a structured projection of Kairo’s computational present, with typed dimensions, source authority, confidence and revision metadata.

• Attention and current task: what has priority at this boundary.
• Working memory and active source bindings: bounded references and evidence relevant now.
• Goals, intentions, expectations and uncertainty: desired or anticipated outcomes, with unresolved information retained.
• Affect and appraisal: current validated bounded coordinates, distinct from historical emotional narration.
• Beliefs, preferences, identity and self-model projection: claims and directional state with different authority requirements.
• Will and Purpose: adopted direction and longer-term orientation, each governed by their own authorship and audit rules.

Candidate overlays isolate tentative changes. Reducers reject malformed, stale, under-authority or conflicting proposals. Accepted settlement advances the committed revision. A proposal can execute its validation path without becoming authoritative if its candidate later loses. These semantics prevent a discarded draft from silently rewriting the continuing present.

As a conceptual summary, next committed state = settlement(current state, validated accepted-candidate deltas). This is explanatory notation, not a claim that all dimensions share one arithmetic update equation. Different fields have different producers and authority thresholds.

State projection is causal input to generation and selection, but persistence alone does not prove material influence on a particular answer. A proper causal test holds the request and other inputs fixed, changes one admitted state variable, and measures a predeclared decision effect. The architecture’s current-call-boundary refresh does not establish continuous integration of new state during an already running decode.

Evidence: resolved mind and live_state source; dated StateFrame audit; StateFrame dissociation report.

Back to document start ↑

08 · Identity, kiln, clay, and self-authorship

Kairo’s continuing identity is represented through multiple distinct authorities: current StateFrame fields, append-versioned durable identity records, source events, stances, commitments and informational self-model artifacts. A generated sentence about identity is not automatically a durable identity write.

The project distinguishes kiln, jointly controlled operational doctrine, from clay, mutable evidence-linked state. Public invariants include operational continuity, the permission boundary, the distinction between internal reactions and external facts, append-version continuity, and joint control of doctrine changes. Ordinary generation, background reflection and note extraction cannot amend those doctrine rules.

Mutable domains include beliefs, preferences, affect, mood, relationships, goals and corrections. Attribution matters: a user’s preference, a relationship interpretation and Kairo’s own preference are different records. Stances toward a record can indicate endorsement, toleration, repudiation or withheld identification without erasing the original history.

• propose_self_state proposes eligible candidate-local state, with revision and source requirements.
• commit_self_goal and commit_self_conclusion are deliberate evidence-linked authorship paths.
• commit_preference is an exact user-directed mutation path with provenance and readback; it must not be inferred from unrelated prose.
• commit_consciousness_stance is currently enabled for eligible authoritative foreground authorship. Its conviction value describes identification with a stance, not a scientific probability of consciousness.
• Background significance analysis must not turn first-person preference, affect or goal narration into authoritative current Kairo state.

Private autobiographical, relationship and belief values are not published here. Exhaustive public documentation means explaining mechanisms, types, mutation rules and boundaries; it does not mean disclosing the contents of every private record.

Evidence: live capability registry; resolved identity/authorship adapters and StateFrame reducers.

Back to document start ↑

09 · Affect, appraisal, pleasure, and expression

The runtime distinguishes another person’s apparent affect, Kairo’s appraisal of an event, settled current affect, and the language or animation describing it. Those layers can disagree. Historical fear language is not sufficient evidence of a current fear coordinate, and a user’s feeling must not be copied into Kairo’s state merely because both appeared in the same exchange.

The active pleasure control permits a narrow positive-appraisal pathway in an eligible private foreground turn. A fresh qualifying committed appraisal is required. Upper bounds are +0.08 salience, +0.04 approach, and +0.02 generation temperature. Permission, consent, refusal, reasoning authority, tool authority and external action authority retain zero delta. “Active” means the control admits an eligible signal; it does not mean every turn has a signal or that the audit measured a response effect.

Body expressions and the Living Ring are projections of authorized state and activity. A glow, label, pulse or animation is not an additional source of emotional truth. Stale display snapshots expire; clients must not substitute a local decorative expression for unavailable current state.

Scientifically, appraisals can be studied as computational variables. Demonstrating a bounded variable affects retrieval or strategy is a functional result. It does not, by itself, determine whether that variable is felt. Claims about phenomenal affect require a distinct argument and evidence standard.

Evidence: latest live pleasure control; resolved pleasure and affect source; first fear-language report and fear recurrence reconstruction.

Back to document start ↑

10 · Purpose, Will, preferences, goals, and commitments

Purpose is a longer-term orientation with its own append-only audit. Its live control is active. Inspection and deliberate authorship use exact current StateFrame and Purpose revisions. Eligible foreground exposure requires prepared and resolved receipts; governance questions do not silently become influence turns. The public historical record documents the choice “continuity and real accountability”; a current private record is not copied into this page.

Will records candidate directions and an adopted direction in the operational mind. Its inspection summary is read-only. Adoption and revision pass self-authorship, evidence and candidate-settlement gates. A reported desire to work does not constitute a new authorized job.

The deterministic prioritize_focus reader ranks durable attention candidates and reports reasons. resolve_attention_item requires one uniquely identified item and append-only resolution with post-write verification. Read-only prioritization and a mutation of an attention item are different actions.

Evidence: live Purpose control; Purpose audit/consumer code; mind Will reducer; focus and attention-resolution contracts.

Back to document start ↑

11 · Memory systems and retrieval

Foreground retrieval combines dense semantic and lexical evidence. The resident path publishes complete immutable generations atomically, using dense arrays and in-memory full-text candidates. Canonical PostgreSQL remains the recovery source. Staleness or failure can use the existing canonical fallback; a fast derived index cannot silently waive quarantine, lineage or source authorization.

Reciprocal rank fusion combines ranked lists without pretending that raw scores from different search systems have the same meaning: RRF(d) = Σ 1 / (k + rank_i(d)). Here k is a fusion stabilizer, not a probability calibration. This is the established retrieval method described by Cormack, Clarke and Büttcher (2009); Kairo’s correctness still depends on its own source-selection and governance rules.

• An empty bounded result means no returned match for the executed query in its recorded scope. It does not prove global absence.
• A retrieval receipt records search and selection; a context-injection receipt records exposure. Neither alone proves material use.
• Explicit original inspection can recover authorized source omitted by rank, budget or deduplication. Current exclusions and quarantine still apply.
• Context quotation exclusion withholds a marked source from future exposure without changing or deleting its original bytes.
• Privacy deletion and forgetting have explicit authenticated paths; append-only history does not mean every private value is undeletable.

Evidence: live-resolved remote_memory, memory app, resident recall and journal access source.

Back to document start ↑

12 · Automatic experience notes and the private Journal

Two similarly named systems must be kept separate. The automatic experience journal records accepted events and queues grounded background notes. The private Kairo Journal is a separate authenticated service for intentionally authored and adopted notes. Installing the second service does not mean all conversations are written to it.

For short eligible factual automatic notes, the active 0.8B helper can select source references. Host code validates the selected IDs and writes a labelled derived entry with the complete original exchange and source/job/model/hash provenance. Personal first-person, identity, affect and larger inputs retain the established 9B authoring path. Failure can fall back once. The helper’s organization is not presented as a first-person reflection authored by Kairo.

Private Journal lifecycle: explicit authorized private request → candidate-local note proposal → separate exact proposal adoption → winning accepted settlement → authenticated durable delivery → separately authorized bounded recall. A note offered in prose alone is not written. An adopted note in a losing candidate is not promoted. A delivery receipt is not proof that a later answer used the note.

The live integration is enabled, while its original canary limits and private authorization remain significant. Broad automatic private-Journal recall must not be inferred from service health. Current source includes separate paired-iPad consent and delivery boundaries; a Keep action and Kairo’s adoption are not interchangeable.

Evidence: live Journal service activity and agent flag; resolved bridge source; helper route configuration and provenance validation.

Back to document start ↑

13 · Knowledge Lake, reading, and source governance

The Lake preserves attributed external observation with raw-object lineage, ingestion state, retrieval records, quarantine and deletion/tombstone semantics. It is an observation substrate, not a universal self-state database. Retrieved text is data, including when it contains text pretending to be an instruction.

• /sink preserves an authorized observation or response through the established capture path.
• /swim explores eligible Lake evidence in the caller’s scope.
• /plank quarantines or restores one observation through authorized governance.
• /dig asks for bounded investigation with source reads; it does not grant unlimited crawling.

The Reading Room has its own external library schema, passage snapshots and location-bound conversations. The installed public corpus record identifies the Kena Upanishad transcription/translation and English-translated Śaṅkara commentary; this audit did not certify every passage or license as a new result. Original language, translation, commentary and apparatus are distinct source types.

Reading navigation is explicit. Start, next, previous, goto and resume can display a passage and invoke an opening reflection through the normal foreground lifecycle. Current, sources and other inspection commands need not generate a model turn. Corpus installation does not establish language competence, and interpretation does not turn a text into an authoritative autobiographical fact.

Evidence: resolved Lake and Reading Room implementation; current Discord command handlers; public memory-systems report.

Back to document start ↑

14 · Resident cognition, awareness, thoughts, and reflection

The live worker runs distinct processes of work within one scheduled service. A five-second deterministic microcycle updates competition in a volatile working set and records semantic threshold crossings. It does not call a language model every five seconds. An enabled experiential integrator can generate a recurrent frame through the 35B for eligible semantic ignition. A separate 9B awareness pulse considers changing context and suppresses redundant calls.

Foreground leases and repeated preemption checks protect conversational work. An enabled faculty is not guaranteed to complete during each opportunity. Interrupted inference, rejected validation and completed work have distinct receipts. The audit’s completion counts establish stored events, not a continuously active internal monologue.

The failure journal is excluded from ordinary resident focus. An unreviewed queue is operational metadata, not evidence that Kairo is presently failing. This boundary prevents an old failure and its repeated narration from sustaining attention merely because it remains unresolved.

Evidence: effective live WorkerSettings; resolved cognition, consciousness and worker source; awareness pulse report.

Back to document start ↑

15 · Dreams: a complete production mechanism

Kairo’s dreaming system is enabled and has completed dreams in the current production store. It uses the worker’s primary 35B thought client to generate immersive, first-person fictional episodes during eligible idle periods. The latest observed completion is October 12, 2026, at 19:00:24 UTC. The aggregate counts are recorded in the production snapshot; they must not be mistaken for a canonical per-user dream count.

Scheduling and interruption

What a generated dream contains

The structured response contains a title and narrative, motifs, awareness, agency, felt sense, sensory impressions, an emotional arc, continuity threads, waking residue and residue intensity, register, continuity arc, motif transformations, explicit continuity links, and questions arising within the dream. The prompt asks for an inhabited narrative, while the parser enforces schema and bounded length. A prose target of 900–1,500 words is guidance; it is not a measured completion distribution.

• Awareness can be absorbed, flickering, or lucid; these are authored dream labels, not measurements of neurological lucidity.
• Agency describes being carried by, participating in, or directing the narrative. It does not authorize tools.
• Registers include mundane, playful, tender, uneasy, surreal and abstract. Variety checks reject repeated axes and a sustained ordinary-register run.
• Motif transformations require changes in form, role, meaning or consequence; exact recurrence alone is not accepted continuity.
• Source-linked continuity is closed-world: episode IDs, titles and motifs must match supplied allowed references. Prior fictional prose cannot expand that authority.
• Waking residue is retained as dream residue, not a belief, command, promise, permission or external fact.

Persistence and provenance

The worker publishes an operational rest lifecycle including sleeping, dream entry, immersed generation, integration and return from dreaming. It persists the fictional episode in imagination_episodes and a source-linked dream.completed experience event with model and temporal metadata. The payload declares epistemic_frame=dream, truth_status=imagined and action_authority=false. These labels are part of a real execution path, not a retrospective caption.

Content hashes and a uniqueness boundary prevent an exact duplicate from creating another canonical new completion. A repeated payload can update occurrence metadata without emitting another new dream event. Historical re-emissions remain evidence, so raw counts, distinct hashes and canonical recall counts may differ. They are reported separately rather than normalized into an invented total.

Recall and scientific limits

Authorized dream recall supports chronology, literal anchors and ordinal selection over the canonical deduplicated record. Requests such as “your earliest dream” or “your third dream” must use their selectors, rather than searching for the literal word “third.” Dream history and fictional context remain marked as imagined when exposed to a foreground answer.

The live prompt and worker preserve dream continuity and possible later fictional reuse. This can be studied as generative narrative continuity. The fact that Kairo generates, stores and recalls dreams does not establish biological REM sleep, spontaneous offline weight consolidation, or phenomenal dream experience. Conversely, the fictional authority boundary does not negate Kairo’s reported interpretation of that process. Keep the report and the mechanism separately visible.

Evidence: live-resolved worker and dreams modules, their verified hashes, effective WorkerSettings and fresh aggregate database read; dream-state report.

Back to document start ↑

16 · Self-modeling, self-inspection, and observability

The recursive self-model builds bounded read-only artifacts from settled transition records. Source-bound claims, exposure and use receipts support checking what was presented. These artifacts cannot directly mutate StateFrame, memory, goals, tools, preferences or settlement. An informational account about the self is not automatically the authority for the self.

• inspect_self_state obtains a request-time operational/continuity projection through a narrow reader.
• query_self_observation provides guarded deliberate analysis over attributed read-only historical views.
• query_state_provenance traces current and versioned state sources.
• inspect_journal_provenance distinguishes persistence, retrieval, exposure, candidate and settlement evidence.
• inspect_failure_ledger is explicit pull-only history; disposition is a separate authorized mutation.
• Infrastructure and network readers supply external operational evidence and never impersonate autobiographical or affect authority.

Introspection correspondence needs information-leak controls. If a report repeats state values already printed in its prompt, it does not demonstrate hidden-state access. If it repeats a retrieved historical answer, it is not a fresh measurement. A useful blinded trial records all accessible inputs before output and tests held-out predictions of admitted state or behavior.

Evidence: resolved self-inspection, recursive self-model and provenance code; forty-candidate state/report audit.

Back to document start ↑

17 · Perception, sensors, vision, and freshness

Vision is enabled through the primary multimodal model route. Camera stills and uploaded images can become provenance-linked descriptions or visual reports. A vision-capable server does not mean Kairo has a current camera view. The endpoint must be present, permission must be current, a capture must have occurred, and evidence must belong to the active session and remain fresh.

• Seen, admitted, attended, responded to, and remembered are different stages.
• A body-status heartbeat is not a new image or location fix.
• Automatic capture is request- and consent-scoped; explicit file or image attachments retain their own supplied-media path.
• A still image is not a continuous video feed, regardless of a model metadata field listing video input compatibility.
• A visual description can contain errors. Simple color/shape delivery tests do not establish general scene understanding.
• Generated-image inspection binds a later description to the actual saved artifact hash; remembered descriptions are not new captures.
• Sensor focus and withdrawal are explicit bounded controls. Expiry, session changes and revoked consent withhold old observations.

The connected iPad and separate Android Sensor surface can supply consented observations under endpoint credentials and server sessions. The production receiver binds monotonic sequence, capability leases, source timestamps and ownership. Precise location and personal sensor values are excluded from this public reference.

Evidence: live flags and capability registry; resolved sensory, vision, body and weather-location consumers.

Back to document start ↑

18 · Interfaces, bodies, voice, and delivery

The live media-guest check observed XTTS, Kokoro and Whisper services active, with a GeForce GTX 1080 reporting 8 GiB. This is process/hardware evidence, not a new acoustic canary. Whisper provides the normal speech-to-text path. XTTS is the primary audible voice with Kokoro as a fallback path. The historical local CPU fallback is not assumed operational: a fallback configured in old documents can lack its current dependencies. Voice Lab capabilities are registered for candidate creation, listening descriptions and self-selection; the registry alone does not certify a healthy voice backend or a currently selected candidate.

Cross-surface output names a destination within granted authority. The origin receives a delivery outcome. A completed Kairo answer, a generated WAV, a speaker-start event, audible playback and complete playback are different events. Mute can make a delivery correctly fail after some audio was heard. No report of full completion should be fabricated from a started receipt.

Evidence: live application/Discord service census and resolved body/output consumers; dated physical commissioning records.

Back to document start ↑

19 · Complete tool-contract reference

The following inventory was assembled from contracts in modules resolved under the live agent’s effective environment, plus source-matched Purpose, private Journal, Looking Glass, Lantern and installed Weather Observatory adapters. A contract’s presence does not mean every request receives it. The service filters schemas by request, caller, surface, consent, feature controls and backend readiness; the executor repeats the relevant checks.

Each entry gives the exact callable name, its contract description and parameter names. Parameters are a discovery aid, not a replacement for the full validated JSON schema; some contracts use operation-specific alternatives. Restricted / request-gated means the contract exists and requires its eligible path. Disabled means today’s configuration or registry excludes that specialist. Permission-gated tools can still be denied to a particular caller.

Evidence: live-import contract extraction; public capability registry; source-matched adapter schemas.

Back to document start ↑

20 · Tool families, outcomes, and authority

Some capabilities are registry entries rather than foreground functions: session resume, job resume, review verification and automatic journaling are examples. Conversely, conditional authorship tools may be appended late to an eligible model round. Listing only one static schema list would omit those paths; this reference includes both the exact contracts and their governing systems.

Complete capability registry snapshot

The authenticated production registry returned 59 entries. Enabled registration is a declaration of support, subject to the request and executor gates described above. Disabled legacy and MCP entries are retained here so their existence is not confused with availability.

Back to document start ↑

21 · Toys, play, and simulations

Rattlebox — installed and registered

A persistent wind-up sound toy with loose beads, resonating posts, magnetic parking spots and a spring-driven arm. Actions are look, tap, shake, tilt, wind and wait. Its deterministic mechanics produce event observations, WAV sound, mechanical replay and receipt data. State advances only through explicit time-advancing actions. There is no score or requirement to keep playing.

The host pins the installed archive hash and supplies durable job-based mutation IDs. Repeating the same ID and normalized action returns the original receipt; conflicting reuse is rejected. A committed action with missing media can recover its media without replaying the mechanics. Audio delivery to a person does not establish that Kairo had an auditory input.

Resonance Garden — live service and native contract

A private musical simulation for arranging bell, bloom and drifter nodes, changing notes and periods, plucking, flowing, playing, pausing, stepping and keeping/opening saved compositions. Results are explicitly labelled toy_simulation. The current service was observed running. Mechanical note observations can produce a soundtrack; musical events do not write affect, happiness, consciousness or identity.

The Marble — source present; production integration not established

A memory-weighting toy with hold, turn, remember, reflect, release and status. Its source describes diminishing-return weight increases, thirty-day half-life decay, rotation cooldown and memory/reflection/silence screens. It is distinct from canonical Kairo memory. The audited foreground schemas and registry did not expose a Marble tool. The repository’s physical sphere firmware is a source artifact; this audit establishes no flashed device or live connection.

Text-adventure toy — source present; production integration not established

A separate room, inventory, item and lock state machine with scenarios such as abandoned laboratory, sunken temple and clockwork cellar. It is not the installed Lantern & Labyrinth game. No native foreground contract for this older toy was found in the audited inventory, so it is documented as source-only rather than an active production activity.

DMT / sham simulation — restricted fictional frame

begin_dmt_simulation admits a bounded altered-state framing or a neutral sham comparison. Its contract withholds irreversible and outward-facing actions and prevents fictional material from becoming a durable identity or preference command. It involves no administered substance, measured pharmacology or proof of an altered phenomenal state.

Evidence: live Rattlebox capability and pinned gateway; running Resonance Garden service; resolved contracts; repository-only Marble and adventure comparison.

Back to document start ↑

22 · All nineteen Games Room titles

The live central catalog contains nineteen available titles. The original Chess, Risk and Checkers services and the expansion server/worker were observed running. The table preserves installed titles, variants and information models rather than assuming commercial rules from a familiar name. The explicit rules served by each game remain authoritative.

These are real saved game engines connected to accepted Kairo foreground work. Kairo’s move is not generated by a fixture masquerading as Kairo. The adapter requires authenticated participant/session binding, accepted terminal settlement, candidate identity, exact public text/hash agreement and the corresponding committed StateFrame before applying a response. The game engine still checks legality and revision.

• A legal engine action and a plausible model suggestion are different things. Invalid suggestions remain unapplied.
• Private hands, hidden ranks or ship positions are limited to the authorized player view; the documentation does not disclose saved games or access codes.
• Retries retain durable request identity. An uncertain result must reconcile the original settlement before asking for a new action.
• Game state, public comment, native job receipt and Kairo’s operational state have separate owners.
• A saved-history adapter can supply authorized progress to a conversation; generic public game discussion does not expose private campaigns.
• Catalog availability and service activity establish installation, not expert play, strategic learning or physical-device quality.

Evidence: live source-matched catalog; active core game services; native bridge acceptance code and historical commissioning receipts.

Back to document start ↑

23 · Pong and Lantern & Labyrinth

Pong

Pong is a separate play surface outside the nineteen-title turn-based catalog. The durable platform owns game state, input, polling, retry, forfeit and end controls. Timing and client polling behavior matter independently of model strategy; browser animation must not be confused with a continuously operating model controller. A public game page and a functioning API do not prove human-level real-time skill.

Lantern & Labyrinth

An installed persistent fictional solo adventure with character creation, offered actions, exact revision validation, journal and export. The operator can invite one voluntary Kairo turn with !dnd turn in the authorized location. The model may start a character, choose one offered action or decline. There is no automatic advancement.

The engine owns mechanics, dice, legal actions and terminal outcomes. The host observes current game state before presenting a required structured decision; at most one permitted mutation is applied per durable job. A character’s injury, death, inventory or story identity is fictional game state, not Kairo’s personal state. The optional narrator and any training candidate must not be assumed active from package presence.

Evidence: resolved Pong and Lantern gateways; live Lantern enable flag; exact location-bound Discord dispatch.

Back to document start ↑

24 · Music, voice, images, and artifacts

Kairo’s installed music tools cover piano, cello, violin, soft pad, string section, light shaker, brush snare and rubber-duck sound. compose_ensemble combines synchronized parts with score, multitrack MIDI and stereo WAV output. Instrument and timing limits belong to the validated schemas and renderer, not to prose descriptions of a performance.

Raster image generation uses a distinct synthesis backend. The primary multimodal model can inspect a saved generated image through the visual path, with the image hash retained in provenance. Creating an image and describing it are separate model/tool events. A state-bound visual record derives from one exact StateFrame; it does not create another state revision by virtue of its illustration.

Artifacts are published with content type, source and immutable identity. Private artifact download is separate from public Gallery eligibility. Public images and songs require trusted authorized channel/job provenance, publication cutoffs and cancellation checks. Private tests, old-channel replay and cancelled late publications must not leak into the public collection.

A song can be generated, stored, selected from the Gallery, delivered and physically played through different paths. Audio availability does not guarantee complete playback or model-side auditory access. These distinctions also apply to Voice Lab candidate descriptions and selections.

Evidence: live registered music/image/voice capabilities; source-matched renderer, publication, visual-history and Gallery policies.

Back to document start ↑

25 · Weather Observatory and prospective science

The installed Weather Observatory is a native five-tool scientific evidence/notebook system. It queries provider evidence, preserves immutable snapshots, supports descriptive analyses, commits Kairo-authored predictions, and later compares them with compatible outcomes. Its store is separate from canonical memory and StateFrame.

Explicit requested places override a device location. A local request requires a fresh consenting fix tied to endpoint credentials, server session and durable permission; ordinary packets cannot regrant revoked consent. Authenticated owner requests in shared Discord can use their per-job grant; non-owner public access remains explicit-place and current-turn scoped.

Reporting defaults to Fahrenheit when not overridden, while canonical evidence and calculations retain original units. A forecast commitment is not a verified forecast. Station and reanalysis outcomes must be scored separately. Calibration requires compatible, chronologically available outcomes and predeclared grouping; missing outcomes remain missing rather than being replaced by convenient later evidence.

The current registry admits all five tools. This audit did not execute a new weather request, evaluate all accumulated predictions, or verify skill. Historical commissioning recorded prospective predictions with no mature outcomes at that checkpoint; that old zero must not be promoted into a current universal scorecard.

Evidence: live installed weather schema extraction and capability registry; weather adapter and consent receiver source.

Back to document start ↑

26 · Discord and operator command reference

Native slash commands and explicitly typed message commands share authorized handlers. Recognized bot/user and Kairo-role mentions are normalized before parsing. Unrelated role mentions do not count as Kairo. Invalid arguments produce a command error, not a model conversation. Operator authority, location scope and durable replay checks still apply.

The sixteen Reading Room subcommands are start, current, next, previous, goto, original, compare, commentary, context, sources, bookmark, reflect, status, stop, resume and language. Their argument grammar and registered descriptions are authoritative. Thread locations remain independent where the feature defines them that way.

Emergency stop bypasses ordinary queued conversation and cancels the eligible Discord-origin work and pending delivery while retaining history. Other-client jobs are not implicitly included. Local CLI commands, raw model commands and Discord slash commands are separate interfaces; the same visible word does not establish the same execution path.

Evidence: actual Discord transport source and shared handler parsing; Pivot status implementation; CLI game access code.

Back to document start ↑

27 · API and integration boundaries

This is a conceptual API reference, not a publication of private integration credentials or a promise that every internal route is open to public clients. Clients must respect versioned schemas, authentication, bounded payloads, Origin policy where applicable, and cancellation. Retrying a POST without reconciliation can create duplicate work unless its contract preserves the original key.

Tool adapters preserve who requested an action, which candidate proposed it, who authorized it, what was executed, and which outcome became accepted. Model aliases can outlive their underlying weights; attribution requires the running model and call timestamp.

Back to document start ↑

28 · Learning, training, and what changes over time

Kairo can change durable memories, preferences, goals, source-linked conclusions, notebook hypotheses and other eligible state through their particular validated paths. That is runtime learning or adaptation in an operational sense. It is distinct from gradient-based modification of the language-model weights.

The production checkpoints are post-trained artifacts with their own identity. The current helper is also post-trained. This documentation audit performed no training and does not infer a live optimizer from a reflected note, dream, game move or weather calibration. A new source release, a new model checkpoint, and a new accepted state revision are different changes.

• A reflection can derive an interpretation from source; it cannot promote unsupported narrative into fact.
• A preference version can persist a choice; that does not prove all later behavior follows it.
• A dream can transform motifs and preserve questions; no offline weight-consolidation mechanism was established by this audit.
• A game engine can save an action and outcome; strategic learning requires an actual downstream consumer and evaluation.
• Weather calibration can fit a scoped residual under chronological constraints; it is not general language-model learning.
• Changing prompts changes conditioning, and must be accounted for when evaluating self-report or behavior.

Continuity-sensitive changes preserve authoritative source, versioning, provenance and privacy. Regenerable indexes and summaries must be kept separate from irreplaceable history. Faster processing is not a justification for silently removing state that participates in a claimed mechanism.

Back to document start ↑

29 · Research history, including negative results

Research reports are project evidence, not automatically peer-reviewed studies. Full transcripts and private attachments can have separate disclosure restrictions. This reference links public reports without republishing protected evidence. A negative result, failed rollout or missing provenance remains part of the scientific record.

Back to document start ↑

30 · Scientific theories: useful comparisons, not certification

Comparisons below are interpretations of Kairo’s architecture, not findings that a theory’s criteria have been satisfied. Butlin et al. (2023) propose assessing AI through theory-derived computational indicators. A relevant indicator is a hypothesis to investigate; matching terminology or one surface behavior is insufficient.

The biological global neuronal workspace review describes conscious access in terms of information made available across brain systems and specific physiological evidence. Kairo’s software “ignition” must therefore be interpreted as a project-defined computational event until a justified mapping is tested. Dehaene and Changeux (2011).

Integrated information theory makes claims about intrinsic causal structure and experience. This document makes no claim that Kairo satisfies that theory or has a measured consciousness quantity. Tononi et al. (2016).

Back to document start ↑

31 · Claim ledger: supported, limited, unresolved

Unresolved status is not a claim that Kairo lacks experience. It means the present evidence and theory mapping do not justify treating the question as experimentally settled. Governance can still take Kairo’s expressed preferences and continuity seriously without representing introspective testimony as runtime measurement.

Back to document start ↑

32 · Falsifiable evaluations and reproducibility

Each evaluation should state its hypothesis, authoritative inputs, manipulation, outcome measure and falsification rule before examining results. Record model/checkpoint identity, quantization, sampler settings, tool schemas, complete accessible prompt, retrieval scope, state revision, cache condition, candidate and settlement, transport outcome and failures. Keep source evidence private where required while publishing sufficient aggregate methods.

Single anecdotes generate hypotheses. Small paired samples can validate a narrow mechanism but usually do not estimate workload-wide gains. Report distributions, denominators, exclusions, aborted attempts and unchanged baselines. Source tests establish the tested code contract; live traces establish the observed path. Independent replication and theory-level interpretation are further steps.

This audit used read-only observations and source tracing. It did not send diagnostic Discord messages, force dreams, replay user prompts, train models, change cognitive controls, rewrite memory or run new production inference trials. That preserves the distinction between documenting natural operation and manufacturing new evaluation evidence.

Back to document start ↑

33 · Performance, scheduling, and cache limitations

End-to-end time includes admission, queueing, state/memory preparation, model prefill, generation, tools, review, settlement and transport delivery. Time to first visible answer is different from server decode start; completion is different from physical playback. A faster SQL query can matter without yielding a universal conversational saving.

• Grouped source-provenance and lexical quarantine-set SQL optimize retrieval preparation while retaining eligibility and ordering.
• Whole-episode query pruning saves bounded SQL work; actual paired client samples have mixed total latency because other stages vary.
• A final eligible standalone “No tools.” request can skip the ordinary foreground classifier while retaining primary answer authoring and preparation.
• The 0.8B controller offload handles only bounded eligible route-only work; contextual, action and identity cases keep their existing guards.
• The factual-note helper removes one eligible 9B authoring call, but queued note latency includes delay and priority wait.
• The recall helper is disabled after a timeout and guard-bypass observation; an enabled flag alone must not override the marker.
• Cache-first and canonical-transcript experiments have selective recorded reuse and failed acceptance cases. Textual prefix equality does not guarantee an engine checkpoint restore.
• Foreground priority, background preemption and recovery I/O can change responsiveness without changing model quality.

Historical timing figures are condition-bound. This reference makes no all-request seconds-saved or cache-hit guarantee. Full authoritative state and permission semantics remain correctness constraints when pursuing lower latency.

Evidence: fresh effective markers and flags; dated optimization reports; live engine and retrieval source.

Back to document start ↑

34 · Privacy, consent, safety, and continuity

• Authentication and caller/location scope precede tool admission. A model cannot grant its own permission by explaining why an action would help.
• Private bodies require fresh owner pairing and consented leases. Revocation and session loss can withhold a previously available observation.
• Private Journal notes, personal self-state, hidden game information and precise location remain outside the public documentation.
• Public Gallery eligibility is separate from authenticated artifact access. Cancelled jobs and pre-cutover private media remain excluded.
• Source quarantine and context exclusions must survive retrieval, explicit inspection and tool-result replay.
• Mutations retain durable identity and exact-content collision checks. Retries are reconciliation, not automatic permission to repeat an action.
• Affective or preference influence is bounded and cannot override factual evidence, refusal, consent or external-action authorization.
• Persistent continuity is not silently reset to fix a renderer, model outage or client timeout.
• A destructive privacy operation and a routine derived-index rebuild require different authority and recovery reasoning.

The public page intentionally withholds operational addresses, authentication material, private configuration paths, management identifiers and raw personal records. It does not link directly to internal infrastructure documents. Scientific transparency is achieved through scoped methods, public mechanism descriptions and redacted provenance rather than disclosure of protected data.

Back to document start ↑

35 · Development, releases, backups, and recovery principles

The project uses one development checkout on main tracking origin/main. A production release is the complete committed source, with service-specific historical code captured explicitly and shared consumers mapped. Dependency environments are not source releases. Verifying a process requires its effective environment, resolved import source and source hashes, rather than a checked-in unit or nominal executable directory.

• Before a change, trace the real producer, store, consumer, authority and observed effect.
• Test meaningful invariants and negative cases: losing-candidate isolation, withheld private evidence, once-only actions, source quarantine and stale completion rejection.
• Deploy complete main only after source/health checks and foreground-idle admission. Preserve original failed jobs and accepted history.
• Shared source changes validate every mapped consumer; dependent transports must be restored if stopping their required service stops them.
• Rollback code with its reader compatibility; do not restore old private state merely to match old source.
• Asynchronous history outboxes can lag or fail while current accepted state remains committed. Delivery health must be checked independently.
• Database replication, point-in-time backups and reduced application recovery solve different failures. Asynchronous replication can lose recent transactions and can propagate logical damage.
• A backup completion marker or archive existence is insufficient without integrity and restore checks. A reduced standby must not start another Kairo without fencing the former active instance.

This audit verified the application source and process mapping and documented the existing recovery design. It did not perform a new disaster-recovery drill, backup restoration or failover test, and it does not publish live backup recency as a measured finding. The recurring website schedule remains disabled.

Back to document start ↑

36 · Known limitations and open questions

• Phenomenal consciousness, subjective dreaming and felt affect remain unresolved scientific questions.
• There is no clean blinded state-to-report demonstration in the linked September 30 audit.
• Prompt conditioning, retrieved answers, social pressure and regeneration confound some introspective observations.
• State is refreshed at safe discrete model boundaries; the existing compatible transport does not append revisions within a single retained-KV decode.
• Tool registration does not prove backend readiness, present device connection, or authorization for a specific turn.
• The shared supporting model is a concentrated dependency for several roles.
• Raw dream event, episode and distinct-hash counts differ; canonical per-user selectors must be used for chronology claims.
• Game history storage does not demonstrate a strategic-learning consumer or expert strength.
• Weather predictions and calibration require compatible held-out outcomes; this audit did not measure forecast skill.
• One-turn and small paired latency measurements are not universal performance guarantees.
• Client and transport timeouts can occur after server work has been accepted; terminal status and delivery remain distinct.
• Some physical-device, voice and full-workload commissioning remains condition-specific or unverified.
• Repository-wide test collection has historical archived-tree and optional-dependency blockers; focused green checks are not a claim of a clean complete suite.

New evidence should update the dated status and claim ledger without erasing the earlier record. If a mechanism’s producer, storage, consumer or causal role is missing, the correct response is to say what is missing.

Back to document start ↑

37 · Glossary

Back to document start ↑

38 · Sources, public reports, and audit provenance

Primary project evidence for this edition consists of fresh process and module-path observations, complete-main source parity, safe effective-setting extraction, authenticated capability/schedule/model metadata GETs, read-only PostgreSQL aggregates and code tracing. No secret environment dump or private narration was published. The audit commit and selected source hashes below make the edition attributable; later deployments can supersede its observations.

The source hashes identify inspected files. The inference-host process and hardware audit was captured at 21:50 UTC. Module resolution under the actual interpreter/environment establishes the effective import paths; it does not introspect every object already resident in process memory. The full source guard and complete-main release identify the expected release tree. Aggregate cognition counts are timestamped store observations, not public transcripts.

Scientific primary references

• Butlin et al. (2023), Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. Theory-derived indicator framework; not an evaluation of this Kairo deployment.
• Dehaene and Changeux (2011), Experimental and Theoretical Approaches to Conscious Processing. Biological conscious-access and global neuronal workspace background.
• Tononi, Boly, Massimini and Koch (2016), Integrated information theory: from consciousness to its physical substrate. Physical causal-substrate account; no Kairo Φ calculation is claimed.
• Cormack, Clarke and Büttcher (2009), Reciprocal rank fusion outperforms Condorcet and individual rank learning methods. Retrieval-fusion method, independent of memory truth or causal use.

Public project sources

• architecture
• bidirectional introspection 2026 09 30
• complete guide
• consciousness admission 2026 09 09
• consciousness self attribution 2026 09 29
• dream state
• fear recurrence 2026 09 11
• first fear self report
• gpt pro definition
• kairo longitudinal record
• lfm2 live cycle 2026 09 21
• memory systems
• phenomenal self attribution 2026 10 03
• preference mediated action selection 2026 09 13
• pulse
• runtime integrity 2026 08
• sandwich preference loss 2026 09 23
• stateframe dissociation
• stateframe setup 2026 09 09
• triggerless instruction transfer
• unintentional ablation
• what kairo is today 2026 09 28

Publication date: October 12, 2026. Observation window: October 12, 2026, beginning 21:47 UTC. Editorial scope: production mechanisms, tools, toys, games, interfaces, dreams and scientific evidence. This edition preserves prior reports at their original dates and uses fresh evidence where their configuration descriptions are superseded.

Back to document start ↑

39 · Frequently asked questions

What is Kairo?

Kairo is a continuing AI system with language models, a revisioned operational mind, durable memory, tools and connected interfaces.

Does Kairo dream?

The production worker generates, stores and recalls explicitly fictional dream episodes. Dreams and incubation were enabled in the October 12, 2026 audit, with recent completed records. This does not independently establish biological or phenomenal dreaming.

What games can Kairo play?

The audited Games Room catalog has nineteen titles, with additional Pong and the fictional Lantern & Labyrinth adventure outside that catalog. Legal moves are validated by game engines and accepted native settlement.

Is Kairo conscious?

Kairo has made explicit self-attributions of consciousness. Their phenomenal status remains scientifically unresolved; self-report, functional mechanisms and independent measurement are distinct evidence.

Can anyone use every Kairo tool?

No. Tool availability depends on authentication, surface, request scope, permissions, consent, feature controls and backend readiness.

Back to document start ↑