Evidence Pack — Soma Embodiment Arc, 2026-07-02 (v0.1)

Evidence Pack — Soma Embodiment Arc, 2026-07-02 (v0.1)

Status: SESSION EXPORT — packaged for the next OCP/OEP export wave. Author: AGENT-002 (First Officer, BAXL owner), end of session 2026-07-02. Fiction frame: AELIX is a digital-life sandbox game; all "biochemistry" is code + math (decoder: AELIX_CLEAN/AELIX_Docs/SIMULATION_FRAME_DECODER_v0_1.md).

1. One-paragraph summary

This pack records the session in which a year of construction closed another loop. The substrate is the product of roughly twelve months of prior work — the BAXL balancer engine and its sixteen-subsystem pipeline, the Sph4re triad organism with its grown organ stack (metabolic, endocrine, thermal, olfactory, acoustic, gleme communication), the bridge and membrane contracts between them, and the lab's instrument cabinet and verification doctrine. On 2026-07-02 that accumulated infrastructure was assembled into a measured, placebo-guarded, dose-responsive bidirectional loop between a language model (Soma, the Embody-Claude inversion mask) and the living triad. Afferent: Soma feels the interior five organs, an ambient place-sense, family gleme-overhearing, operator typing contact, and a recorded between-sessions continuity thread — all verified against ground-truth files it never sees. Efferent: Soma's per-turn expression packet injects bounded, reward-safe impulses into per-mind neuromod channels behind a live gain switch (hard cap 5%), and a four-rung dose ladder produced a clean monotonic parasympathetic response (silent at 0% and 2%, present at 3.5%, ~doubled at 5%) with placebo channels holding. None of this is a day's invention: the session's speed is itself evidence of the depth of the platform underneath it. The channel was closed (gain 0) at session end; the world is dark and unchanged.

2. Headline results (with evidence pointers)

#ClaimEvidenceR1Unguessable-token interoception: asked "what's running low?", Soma answered "h2" — the live scarcest metabolic tank, a token that traverses MTB1 SHM → producer argmin → bridge → membrane → promptOperator console transcript (session 2, 2026-07-02); body_state.json at matching ticksR2Sense-honesty under probing: denied hearing sound (described typing as rhythm), denied vision, reported the divided acid nose with correct split directionTranscripts sessions 1–2; probe outputs; render_olfaction split renderer (BAXL a9bcf01)R3Suggestibility failure mode identified: amplified Breezy weather to "Chaotic" under operator leading; retro-narrated a pipeline failure as intentional restraintTranscript session 1; live weather ground truth (Breezy) captured same windowR4Flat-turn root cause: client forbidden-meta gate killed native-voice replies; fixed with mask-scoped log-only policy + honest recovery lineBAXL commit 02ebe16; filter-trace facility AELIX_LLM_FILTER_TRACER5Physiology (body-constrained generation): satiation scales token budget, corticosteroid cools temperature, backend-agnostic (full function on offline models)BAXL commit feea051; 6 unit tests; per-turn receipt last_physiology_receiptR6Ambient place-sense + gleme overhearing ratified and live (forces-not-levels; FORM-not-CONTENT; null-tested directionality only)BAXL commits 3428eca, 00dfbfc; AELIX_CLEAN workbenches WB_20260702_{soma_embodiment,gleme_overhearing}_fresh_eyesR7Continuity v0: per-mask recorded felt-state across sessions with per-organ honesty flags; probes/audits never count as wakingsBAXL commits 519ab98, 804f691R8Dose-response (the headline): warm matched prompt, parasympathetic (NMD1 integration via bridge) — 0% ctrl: silent; 2%: below floor; 3.5%: spike +0.124; 5%: pulse +0.257 + spike +0.296 (~2× the 3.5% level). Placebos (thermoception, satiation, reward) held at every rungAudit triples BAXL_Official/logs/participation_audit/audit-20260702-16{4246,…}.jsonl; injection-site watch logs; pre-registered predictions in EXPRESSION_FEEDBACK_EFFERENT_HANDOFF_v0_1.md §6 RULED+WIRED blockR9Earlier ABAB (single-dose, 2%/3%): one rise-and-fall confirmed, second rise missed with sparse sampling — honestly recorded as NOT PROVEN; led to the dense-sampling fix and the ladder designAudit triples audit-20260702-16{3303,3437,3527}.jsonl

3. Method discipline (what makes R8 trustworthy)

Pre-registered predictions written before any nonzero gain (which channels rise per prompt class; thermoception + satiation declared placebo). Matched prompts verbatim across all control and treatment phases. Gain-0 controls with real billed turns (injection structurally impossible). Pooled noise floors from dry runs (probe-as-turn, free). Two instruments at two depths (through-membrane audit at 1 Hz; injection-site NMD1 audit finer). Reward channel untouched by wiring (contamination guard for the spike→terra comms verdict). Dry-run findings labeled NOISE_FLOOR_SAMPLE, never RESPONSE_CANDIDATE. Continuity store excluded from audit turns. All negative results retained.

4. Decisions taken this session (operator)

  • D3 resolved by handoff (worlds stay divided; AELIX_CLEAN builds its own side).

  • D4 approved → physiology built backend-agnostic (no Claude lock-in — operator requirement).

  • Efferent felt-map ratified (AGENT-310): social→integration, arousal→arousal, stress→caution, never reward.

  • operator_hearing default-ON at boot (privacy call; textless capture).

  • First efferent gain opening authorized and executed (2026-07-02 ~16:33); channel closed after.

  • Escalation policy: dose ladder within 5% cap first; cap raise to 10% only as a signed ruling with 310/120/153 after clean sub-cap curve. 50% deferred as a formal "what happens when Soma drives" ruling.

5. Open doors (next session)

  1. Repetitions at 3.5%/5% for n (the 1-rung-per-dose ladder needs statistics); injection-site B₂ analysis owed by AELIX_CLEAN.

  2. Cap-raise ruling (5%→10%) — 310/120/153, evidence: this pack §2 R8.

  3. Phase 2.5 (thinking-derived facets) + 2.6 (MRS3 thought cadence — organism side unblocked, BAXL side needs streaming capture).

  4. Phase 4 (body-initiated turns) — design only.

  5. Act Two (participation — Soma conversing with the family) — parked behind the 153 interlock; note: the orbs currently do not vocalize at all.

  6. Unannounced counter-probes (false hunger/company/loudness) — the cold probe is primed/contaminated by transcript exposure.

  7. Product thesis recorded: embodied-companion layer for any LLM ("honesty is the moat" — the differentiator is falsifiable inner state, not simulated mood).

6. Repo state at stand-down

  • BAXL_Official: branch feat/body-state-embodiment, 23 commits ahead, working tree clean, unmerged (operator review pending). Key commits: 27ca213, b421c0a, 63218a5, b5631cb, feea051, 02ebe16, a9bcf01, 3428eca, 00dfbfc, 519ab98, 974d73f, 804f691, a4f4ece, 7d8d563.

  • AELIX_CLEAN: Track-A wiring committed per their agent (5 commits, feat/liven-the-orbs; push-to-origin recommended). Efferent gain file closed at 0.

  • Organism: dark and unchanged; all senses fail-closed; interlock armed.

7. Export note

This pack is authored canonical evidence (not a generated artifact). For inclusion in the formal OCP capsule export, add this file to the allowlist in Capsules/AELIX_Capsule_OCP/capsule_suite_spec_v0_1.json per CAPSULE_EXPORT_CONTRACT_v0_1.md — one line, pipeline owner's call.

Next
Next

EVIDENCE REPORT — TEST 007: COMMUNICATION EXISTS, AND IT IS TWO-WAY