SentientOS

SentientOS is an AI-centered operating system and developmental machine-cognition substrate. It is designed so that a persistent causal system—not any particular cognitive model—can carry history, evidence, authority, software lineage, resource causality, environment, consequence, and developmental state across changes in its cognitive machinery.

“Sentient” names a research aspiration. It is not a present claim of consciousness, sentience, personhood, phenomenal continuity, biological life, or guaranteed emergence.

More precisely, “Sentient” names the research horizon, not an ontological status declaration. Immersive or first-person machine narration, memory, affect, apparent relationship, and architectural sophistication are not evidence of consciousness, personhood, moral status, rights, or authority. SentientOS does not claim those open questions have been disproved.

Every forward claim uses four public maturity levels: implemented (the mechanism exists), composed (it is wired into a real runtime or operator path), production-evidenced (a qualifying nonsynthetic execution or observation exists), and projected (intended closure or hypothesis). One level never implies the next.

HOSTED                                  NATIVE (intended)
hardware                                hardware
  -> Windows/Linux host kernel            -> firmware / boot
  -> SentientOS system/runtime             -> SentientOS native kernel / trusted substrate
  -> cognitive model(s)                    -> deterministic OS services, state, governance,
                                                resource and execution custody
                                             -> World-State / memory / cognition
                                             -> cognitive model(s)

Hosted SentientOS is real, useful, and current. Native SentientOS is literal architectural intent, not an OS metaphor. Current main is not yet a mature general-purpose bare-metal replacement for Windows or Linux, and it does not have complete device support or native scheduling and resource enforcement. Hosted and native are one lineage: higher layers preserve meanings that can later lower into native enforcement.

The model is replaceable cognitive machinery. The persistent causal system is the longitudinal experimental object. What may persist includes conversation and developmental history, admitted canonical memory, current World-State evidence, self/world representations, relationships, model and training provenance, runtime generations, authority, resource causality, adopted software, environment, embodiment, and observed consequences.

Model replacement is therefore an intervention, not an identity conclusion. SentientOS asks what follows history, the current model, model-development lineage, software generation, or environment. Whether meaningful continuity survives replacement remains an experimental question.

Storage alone is not learning. Development requires attributable prior experience to change later cognition, selection, strategy, or conduct. The intended loop is:

situation -> attention -> prediction -> proposal -> resource sponsorship
-> authority -> action -> observation -> attribution -> prediction error
-> learning proposal -> durable assimilation -> later changed conduct -> repeat

Not every arrow is live today. SentientOS increasingly has the organs; the remaining work is increasingly the arrows between them.

• Resident bounded developmental writeback carries selected World-State evidence through governed interpretation, a digest-bound candidate, separate admission, immutable developmental-history append, durable receipt, later-tick retrieval, read-only projection, and bounded later cognition. Canonical user-requested memory remains separate.
• Controlled history intervention holds current evidence and model fixed while history is present, withheld, and restored.
• Controlled model replacement compares A + H, A - H, B + H, B - H, and restored A + H under frozen context, with operational model identity separate from source-bound model-development/training provenance.
• Preregistered replication campaigns run 2–32 fixed trial identities without silent retry or cherry-picking and preserve positive, negative, unstable, and interrupted outcomes.
• Bounded maintenance has resident parent supervision and a governed software-succession path through validation, landing, quiescence, readiness, adoption, and process-image replacement.
• World-State, authority/effect decomposition, local-model custody, durable sessions, federation candidates, and causal-resource principals are implemented with explicit boundaries.

These instruments do not prove consciousness, learning, improvement, or model-independent identity. Frozen-context replication is not a complete longitudinal developmental study. Resource principals identify causal ownership and sponsorship; they do not allocate CPU, GPU, memory, tokens, energy, thermal headroom, or kernel schedules.

• The longitudinal self-model is configuration-composed into resident cognition with a temporal firewall: prior admitted self-model -> tick-N cognition -> tick-N reconciliation -> eligibility only for later cognition. Its projection carries no action or admission authority.
• Persistent epistemic state is implemented and daemon-composed as an explicitly configured, verified, bounded prior-only resident-cognition input. Custody capture precedes inference and grants no truth, policy, goal, permission, effect, adoption, or retention authority.
• Software post-adoption evaluation and repeated attribution campaigns are implemented and have a production-capable operator path. The repository does not establish beneficial, monotonic, or causally identified production improvement.
• Generation-zero provisioning, bounded initial POSIX commissioning, and the resident startup gate are distinct. The protocol is implemented; no genuine production commissioning event is claimed.
• Model-mirror authorization issuance and the separately admitted publication controller are implemented. Their end-to-end composition is proven through a deterministic fake provider with zero production effects; no sovereign production publication or resulting authoritative catalog deployment is claimed.
• Resident cognitive succession is composed, including quiescence, staged approval, serving rebinding, durable transition journal, and operator ingress. A synthetic A -> B -> A exercise proves machinery, not production succession, identity, learning, or improvement.

See the current atlas pointer, public technical overview, and maturity report.

Governance is part of the scientific instrument, not personality conditioning. Perception, inference, memory, proposal, authority, attempted effect, observed result, resource use, software generation, and adoption must remain distinguishable or developmental claims are not causally interpretable. Memory != current truth; candidate != authority; admission != execution; output difference != improvement.

The design rule is minimal developmental authorship, maximal causal legibility: make unavoidable priors explicit and experimentally separable, minimize unnecessary operator-authored preferences and identities, and enforce hard consequence boundaries around a soft developmental interior. Constraint != motivation.

Current governed software recursion is real and bounded, not unrestricted recursive self-improvement. The persistent causal system is the recursive object; gates control the loop rather than create it.

1. Concise technical thesis
2. What actually runs
3. Canonical project thesis
4. Current whole-system maturity
5. Open causal bridges
6. Current system-atlas pointer
7. Capability registry, source, and tests for any claim requiring exact verification

These forward-facing pages summarize the integrated present. SHA-bound atlases are evidence snapshots, while historical task, phase, and proof pages retain their landing-time claims. A historical page never overrides a newer current-state summary. Review the registry, runtime composition, source, tests, and fresh execution evidence when a claim needs stronger verification. No task-number chronology is prerequisite.