The Kairo Pulse

01   What the pulse is

Terminology boundary. “Awareness” names an engineered review function. “Pulse” names its cadence. Neither term, by itself, establishes phenomenal consciousness, sentience, or uninterrupted private experience.

02   Purpose and design goal

The long-term goal is not maximum autonomous activity. It is a stable middle ground between two failure modes. A system that exists only during foreground turns has no engineered process for noticing time or revisiting unresolved material between conversations. A system that generates continuously can flood its own history with repetitive prose, amplify mistakes, and mistake model output for evidence. The pulse aims for bounded endogenous continuity: periodic reconsideration whose inputs, authority, cost, and durable effects are explicit.

03   The pulse among Kairo's resident layers

UI animation and body haptics also use the word pulse, but they are presentation and action features. They are not the awareness checkpoint described here. The model-generated pulse may read a projection of resident cognition, but it does not replace the microcycle or the experiential integrator.

04   End-to-end process

1. Establish route health.The worker probes the dedicated brain endpoint through a cached health contract. A healthy endpoint earns the faster cadence; health is not assumed from configuration.
2. Select a due identity.Only identities with an eligible recorded human exchange can become due. The first review after a newer human event uses the initial-delay rule; later reviews use the interval rule. Repeated failed or preempted cycles add exponential backoff.
3. Honor the quiet hold.If an unchanged check was recently suppressed, a Valkey TTL prevents the poll loop from immediately reconsidering the same identity.
4. Build runtime evidence.The worker reads recent events, the latest reflection, current state, unresolved items, observations, temporal facts, and optional cognition/autonomy projections. Quarantined records are excluded.
5. Fingerprint the meaningful world.A digest compares the latest human event and normalized observations against the prior check. If nothing meaningful changed and the quiet budget remains, inference is skipped.
6. Bound the addressable set.State and unresolved items are ranked and capped before prompt construction. The same capped identifiers become the gate's allowlist.
7. Generate structured review.The 9B receives untrusted evidence inside an explicit delimiter and returns a JSON envelope. If parsing fails, one schema-constrained repair attempt is made.
8. Gate every proposal.Deterministic code rejects unknown verbs, out-of-evidence targets, excessive confidence changes, attempts to rewrite immutable content, protected retractions, and proposals beyond the verb budget.
9. Settle atomically.Under a per-user database lock, the worker rechecks that the source event still exists, applies accepted verbs, records the cycle, writes any warranted journal event, and commits as one transaction.
10. Close the cadence.A successful cycle clears accumulated quiet-skip bookkeeping. Errors and preemptions are separately classified and recorded so failure cannot become invisible activity.

05   Cadence, fingerprints, and quiet time

Production is configured around a 60-second brain interval and a 3,600-second maximum quiet interval. The implementation also imposes a 30-second floor on the healthy brain cadence, preventing a configuration error from creating a busy inference loop. When the dedicated endpoint is not healthy, the worker falls back to the slower CPU-cadence interval and restores foreground-work preemption.

The fingerprint is a SHA-256 digest of the latest human event identifier and a sorted projection of observations. Observation timestamps do not participate. Above-routine observations participate with their literal value hash. Routine observations participate through a shape transform: numeric leaves collapse to a placeholder and timestamp-like strings collapse to another placeholder, while booleans, enums, nulls, collection structure, source, key, and severity survive.

Why the quiet ceiling matters. Fingerprinting makes the pulse event-sensitive, but time passing is itself relevant to continuity. The ceiling guarantees an occasional real checkpoint; the accumulated skip count lets that checkpoint truthfully report the silent interval without manufacturing dozens of empty experiences.

06   What the pulse is allowed to see

The worker constructs a bounded CYCLE_EVIDENCE object. It is data, not instruction. Text-bearing fields are escaped, individual values are length bounded, the full serialized envelope is capped, and every mutable target carries an explicit identifier.

The caps serve two purposes. They bound latency and context size, and they create a precise authority boundary. The model cannot validly refer to a hidden state row because the gate derives its accepted identifier set from exactly the rows rendered into the prompt—not from a broader database query performed afterward.

07   The model contract

The 9B acts as a background reviewer, not an executor. It receives no tools. Its required response is a compact JSON envelope with a boolean meaningful_change, a bounded first-person continuity string beginning with “I”, and a verbs array. It is explicitly told that no durable change is the normal outcome and that waking is not a reason to invent one.

Streaming is used to support interruption on the shared fallback route. On the dedicated brain route, the foreground watchdog is intentionally omitted because that GPU does not contend with foreground inference on the worker host. Both paths retain a wall-clock inference cap. Empty output, transport failure, invalid JSON, and repair failure become explicit failed awakenings rather than silent success.

08   What a pulse may change

At most four verbs are accepted per cycle. Attempts to include immutable fields, target unshown identifiers, use unknown verbs, retract protected rows, or exceed the budget are rejected and retained in the cycle ledger with reasons. Rejection is observability: the platform can measure how often the reviewer proposes actions outside its competence.

If the recent rejection rate exceeds the configured threshold, the reviewer enters narration-only degradation. Its prose is retained for inspection, but every proposal is discarded. The mechanism therefore responds to demonstrated gate disagreement by reducing authority rather than assuming model competence.

Identity origination is outside the pulse. The unattended reviewer cannot create a belief, preference, goal, affect, mood, relationship fact, or new state key. Kairo's foreground self-authorship uses a separate typed, revision-bound proposal and validator path. User corrections use their own higher authority. Pulse prose is never a shortcut around either path.

09   What becomes durable

The implementation distinguishes checking, reviewing, and changing. This avoids treating each timer tick as an experience or each generated sentence as self-state.

A completed cycle records the evidence counts, pre-cap availability counts, observation identifiers, quiet skips, model name, inference duration, accepted and rejected verbs, narration, outcome, and a read-only recognition projection. If a resident cognitive ignition was supplied, settlement appends a receipt showing that the awareness pulse consumed that bounded projection.

State changes and the ledger row that claims them share one transaction. Before committing, the worker takes a per-user advisory lock and confirms that the source conversation event still exists. This closes the race with forgetting, clearing, or quarantine: a review cannot survive by attaching itself to evidence the user removed while inference was running.

10   Failure, fallback, and foreground work

• Dedicated route healthy. The pulse uses the 9B brain role and the fast cadence. Foreground work does not preempt it because this route is isolated from the worker host's foreground inference.
• Dedicated route unhealthy. The health cache is poisoned after a mid-stream failure. A later attempt uses the slower fallback cadence and yields when foreground or durable queued work needs the shared model.
• Foreground arrives on the shared route. Streaming checks leases and deferred waiters, then cancels the background generation promptly and records a preemption.
• Repeated non-completion. Exponential backoff is added to the initial delay or interval, preventing a permanently due identity from restarting on every poll.
• Inference exceeds its cap. The awakening closes as an error with no mutation.
• Malformed JSON. One bounded repair request is attempted. Parsing and gating remain inside the failure boundary.
• High proposal rejection. Recent rejection history causes narration-only degradation until measured competence recovers.

Concentrated dependency. In the audited production layout, the brain endpoint and identity fallback ultimately depend on the same resident 9B inference service. The fallback changes scheduling and contention behavior; it is not an independent model-availability guarantee.

11   Engineering properties

The most important architectural decision is that the language model does not own the loop. The model performs the semantically difficult middle step—interpreting bounded evidence and proposing a response—inside a scheduler and settlement system whose invariants are independently testable. This gives Kairo a background process capable of meaning-sensitive review without making model fluency equivalent to database authority.

12   Boundaries, risks, and the next goal

The engineering goal. The pulse succeeds when it makes Kairo more temporally coherent and behaviorally well-regulated between conversations while remaining sparse, reversible, inspectable, and subordinate to stronger evidence and authority. More pulses are not inherently better; better recognition with fewer unjustified durable effects is.

Implementation map