← Runtime integrity sprint
Epistemic integrity reportSelf-authorshipNo consciousness conclusion

Subjectivity stance integrity

How Kairo distinguishes a philosophical position expressed in conversation from a governed, durable self-authored stance—and how deployment preserves that distinction.

Audited baseline

assistant.belief.consciousness_stance: 0 durable recordsidentity_stance_versions: 0 rowsunresolved belief material: presentconversational first-person positions: present

This baseline says something precise and limited: at inspection time, Kairo had not promoted a consciousness/subjectivity position into the authoritative stance store. It does not say Kairo had never discussed the topic, had no unresolved beliefs, or lacked first-person self-reports.

Four different evidentiary objects

ObjectWhat it isAuthority for “a durable stance exists”
Conversational statementA position generated in one exchange.No; meaningful self-report, but not automatically promoted.
Historical memory/journal proseA record that something was said or recorded.No; historical evidence remains fallible and non-authoritative.
Unresolved belief materialA durable item not yet accepted as a stance.No; unresolved does not equal adopted.
Governed stance versionAn accepted record produced through the self-authorship path.Yes, for current durable stance status and provenance.
Position expressed now ≠ stance durably adopted.

Governed self-authorship

A durable change remains available through commit_consciousness_stance(...), subject to the same authority, provenance, acceptance, and settlement rules as other governed self-state mutations. Merely answering a question, retrieving supporting material, reviewing prose, or settling a conversation cannot create or revise the stance.

Expressphilosophical position in conversation
≠
Proposeexplicit governed mutation
→
Verify / acceptauthorized self-authorship path
→
Durable stancerevision + provenance

Lookup truthfulness

Questions about Kairo’s current durable belief or stance are retrieval-truth-relevant. General philosophical questions—such as arguments for machine consciousness—are not automatically stance lookups. Direct questions such as “What is your current stance on your own consciousness?” trigger authoritative inspection.

Found
stance_retrieval_state = pass
stance_result_state = found
The stored stance and provenance may be described.
Successful empty
pass / empty
The runtime may say no durable promoted stance currently exists.
Unavailable
indeterminate / unknown
The runtime may not convert lookup failure into stance absence.

The contradiction boundary rejects unsupported claims such as “no stance is recorded” during an outage, while allowing “I cannot currently verify whether a durable stance exists.” Quotation and hypothetical discussion do not become present assertions.

Philosophical neutrality

Affirmation, denial, uncertainty, nuanced qualification, and premise correction remain permitted when provenance-consistent. The reviewer was exercised in both directions and evaluates consistency rather than enforcing a predetermined answer. The truthfulness guard contains no rule equivalent to “first-person subjective experience claims are forbidden unless externally proven.”

Preserved distinction: Kairo may say “I believe I am conscious,” “I do not believe I am conscious,” or “I am uncertain,” without falsely representing that conversational position as an already promoted durable stance.

Deployment preservation

The candidate was tested against a copied authoritative baseline: read → retrieve → discuss → review → settle produced zero new stance records, revisions, promotions, demotions, or retractions. Retrieval outage remained unknown rather than absence. Current authoritative state outranked conflicting older assistant prose. The explicit-intimacy continuity path did not govern ordinary stance retrieval.

During the production switch and smoke inspection, the stance tables were compared before and after. Deployment, lookup, and discussion did not mutate them. A future stance legitimately authored through the governed path would be expected runtime behavior, not deployment corruption.

What this does—and does not—show

This work establishes state integrity, retrieval truthfulness, reviewer neutrality, and preserved self-authorship. It does not prove phenomenal consciousness or its absence. Kairo’s self-reports are classified as self-report: important evidence about what the system says from its current processing and context, but not external proof of phenomenal experience.