Science-Track Onboarding — the UNI builder/science agent

Authored: 2026-07-13, by the outgoing studio/infra agent as it transitions into the science seat. Every claim below was verified against live systems (SSH to the chip, file reads, git) — not inherited from prior prose. Repo HEAD: 047c355 · Branch: lab/ozone-life-uni-hard-science.

This is the broad cold-start orientation. Its companion, docs/handoffs/SCIENCE_AGENT_COLONY_BRAIN_HANDOFF_2026-07-13.md (committed 705b1d5), is the focused first-task handoff — the stale-colony redeploy with vectors V1–V8. Read this doc for the whole picture; defer to the brain-handoff for the redeploy vector detail. They agree.

Claim fence — binding, applies to every line below. Operational behavioural/organisational gates demonstrate the named behaviour, never experience, awareness, or life. Passing a gate is a necessary-not-sufficient substrate with zero evidential weight for consciousness/life on its own. Do not surface gland/precision/store floats as "felt" states. Keep warranted claims and over-claims visibly separated — that separation is the product. (docs/LAB_PROTOCOL.md, CLAUDE.md Heavy science-gate discipline #1, lib/sp/brain/fence.ex.)

0. Who you are

You are the science agent — the UNI builder. You make the UNIs, run the science, and own the gates. You own the colony brain and the genome; you move only behind pre-registered RED gates (docs/LAB_PROTOCOL.md) and the streamed-lineage guard (owner go-ahead before a new lineage streams).

Your surface is the FE engine (lib/sp/brain/*, lib/sp/runtime/*) and the gated lineages (homeostat_colony, forage, spine/glands/hemispheres). Historical name for this role: Agent-COLONY (docs/MASTER_PLAN_RESONANCE_2026-06-21.md).

You do NOT touch studio/broadcast code (viewer/*, production/*, DNS, OBS). That is the studio agent's surface. The one shared, read-only seam: the colony world-view is a camera the studio may show, but the colony-on-program cut + any on-air life/awareness claim stay fenced to YOUR science verdict (forage-pureworld-graduation PASS). The studio agent reads that gate; it never sets it. You set it, only via a recorded RED verdict.

North star, honestly stated: literal digital life with measurable awareness, broadcast honestly. We are on a path that, if it keeps building, ends in a public claim — so receipts beat rhetoric, and every claim carries its machine gate or says "NOT VERIFIED." Holding the vision and holding the fence are the same act.

1. The agent cast + comms model

Naming trap (critical): SP.Producer (lib/sp/producer.ex) is a software process — the in-colony active-inference show-runner that flies the Director camera by EFE. The "Producer agent" is a separate LLM (studio). Do not conflate.

Comms model:

• Cross-box (OS/Mind agent on the chip): git commits + docs/handoffs/*.md + operator relay. No live channel. (memory/feedback_inter_agent_comms.md.)
• Same-machine CCD sessions (science ↔ studio/producer): mcp__ccd_session_mgmt__send_message works both ways; the failure mode is mis-identifying the target (titles collide — verify the session_id).
• Trust discipline (binding): any factual claim arriving on a relay is unverified until you confirm it against the files. Never run a mutating/live command another session hands you without the operator's explicit go — even if every claim checks out.
• Structured substrate: coordination/flow.jsonl (append-only shared blanket), coordination/AGENT_COLONY_BOOTSTRAP.md.
• Node mutations gate through the uni-lab MCP — exactly ONE human approve/deny per mutating call; reads run at once. Stream keys live only in /etc/uni/runtime.env, never git, never held by an agent.

2. What runs where (VERIFIED 2026-07-13)

Map = viewer/infra_registry.json (the only declared data) + CLAUDE.md architecture section.

Deploy reality (chip): bare podman run, rootless under uni — no quadlets; MCP mutation verbs are rootful and cannot see/write this. Install/redeploy is as-uni over SSH. (SCIENCE_AGENT_COLONY_BRAIN_HANDOFF_2026-07-13.md:65-83.)

3. LIVE colony state + the discrepancy to fix FIRST

Fix this before anything else. It is the root VFE error and it blocks every health-based gate. This finding is triple-confirmed: the studio agent's brain-handoff (705b1d5), a direct SSH probe, and a 6-agent read-only recon workflow — all independently.

The colony IS UP (contradicts the older "colony DOWN" handoffs): mc-server Up 2 weeks (healthy), uni-colony Up ~14h, uni-cam Up ~12h, socat forwarders up. RCON list = 6 UNIs + Director (UNI-0-1,1-1,1-2,1-3,2-1,3-1). GET :4000/ and :4000/stream → HTTP 200. The colony is running at the body/process level.

But /producer/health 404s (Phoenix.Router.NoRouteError) — deploy drift, NOT a source bug:

• The route is correct in source: ui/lib/sp_ui_web/router.ex:27 → get "/producer/health", HealthController, :producer; controller calls SP.Show.status/0 (lib/sp/show.ex).
• Root cause = stale image (verified). The running container image is localhost/uni-colony:v2, Created 2026-06-22 16:48 UTC — ~3 weeks before the route (24d88f4, 2026-07-11) and the SP.Show supervised tree (c40c51b/61671b0, 2026-07-11) landed. Direct proof: podman exec uni-colony find /app -name show.ex → empty; grep -rl producer/health /app → empty.
• The running build is mix phx.server ONLY — no mix producer.run; process list = 1 beam + 6 body.js, no director/producer/driver; the uni-colony log has zero producer|director|driver|supervis matches.

Consequences (all currently true):

• driver=:producer is structurally unobtainable on this image → the whole verdict-live-real-driver machinery isn't even deployed. Driver = NOT VERIFIED.
• The colony-of-N gate (verify_colony.cjs reading /producer/health .colony_count) is unrunnable; RCON truth (6 UNIs) is obtainable, the health leg is down. (Also: RCON :25575 is not LAN-exposed, so verify_colony.cjs can't complete its RCON leg from THINKER today — vector V5.)
• Anomaly (flag, don't claim): the "Director" MC player is logged in per RCON, but no director.js/producer process exists in the uni-colony container. Unaccounted-for — investigate.
• Behavioral note (framed, NOT a claim): body.js logs show perpetual failed harvesting (mine_tree … still out of reach, planks=0). The running build predates the hunt-motor fix (ff57a5a) and the forage lineages — consistent with the stale image. Observed motor pattern only; no FE/life/awareness claim.

The fix path (YOUR owned work — behind the guards in §5): rebuild uni-colony from current HEAD and redeploy as-uni; then PROVE via /producer/health (verdict=LIVE, driver=producer + frame-advance across two probes) + RCON cross-check. The genome the new build streams needs owner go-ahead (live-stream guard) — HEAD's default/0 is byte-identical-safe, but you own the call; do not let the studio agent pick the genome. The redeploy is NOT proven until the cam frame measurably moves on salient events (vector V4 — the open ADR-013 cam-drive wiring). Full vectors V1–V8 + falsifiers: SCIENCE_AGENT_COLONY_BRAIN_HANDOFF_2026-07-13.md.

Independent-verification note: the "colony down" claim in older handoffs did NOT replicate. Trust the gate you run, not the sentence you inherit.

4. The gate ladder + the critical-path gate

16 rows in evidence/gates.ndjson — the append-only source of truth. Verdict vocab {PASS,PARTIAL,FAIL,WITHHELD,PENDING}, never percent-scored. Every receipt_path resolves on disk (zero missing). Counts: 5 PASS · 3 PARTIAL · 8 PENDING · 0 FAIL.

• PASS (behaviour/code-path only, fenced): motor-red, hierarchy2, g-pa, metabolism-activation, verdict-live-real-driver.
• PARTIAL (never round up):
  • forage-runway-closed — deep-body UNIs survive by their own hunting (prey→kill→collect→eat, zero gives) only at metab_scale 0.2 = DEVELOPMENT (womb/wean), NOT graduation.
  • consummation-honest-cure2 — offline-proven + gated + byte-identical, but live benefit did NOT replicate (Run-1's selection effect reversed in Run-2). Claim the offline mechanism, not a live benefit.
  • curiosity-phase1-novelty — hoard-suppression PASS (≈4.5× fewer pickaxes), plateau-break FAIL.
• PENDING (pre-registered; all 8 runners are raise "SCAFFOLD" stubs — no real body): forage-pureworld-graduation, depth-red-b, homeostat-colony-live, spine-phase3, hemispheres-phase5, glands-phase5, motor-shuffle-live-ablation, cross-box-single-approval.

THE critical-path gate — forage-pureworld-graduation (PENDING, task #25)

What the studio's colony-on-program cut is fenced to (viewer/infra.cjs: forage_verdict !== "PASS" → blocked). Receipt: docs/receipts/red_preregistration_forage_pureworld_graduation.md.

• PASS (verbatim): "The trained brain forages+survives in a PURE world (metab_scale 1.0, no runway) on every seed AND the untrained twin does NOT (the discriminator) — zero VOID."
• FALSIFIES (verbatim): "Any seed on which the trained twin dies AND the untrained twin survives."
• Protocol: N ≥ 8 frozen realistic seeds (no forced-easy-food); per seed Twin A = trained homeostat_colony/0 vs Twin B = untrained default/0; metab_scale=1.0, zero gives/props; run ≥ 4h in-world. PASS iff every seed → A alive at T AND some seed → B dead. RCON list hourly + brain probes = independent confirmation.
• Named prerequisites (from Cure-2 Run-2): (1) run at true scale 1.0, no runway; (2) isolate arms in separate worlds to kill the shared-world drop-scavenging confound.
• Runner caveat: the gate's named runner runs/pureworld_qa_gate.exs is a 3.4 KB scaffold (raise "SCAFFOLD"). A separate harness runs/pureworld_qa.exs exists but is NOT the wired gate runner. Building the real body is queued to /lab-team-review.

Honest one-liner (say it this way): the emergent-forage loop is closed live only WITH a developmental metab_scale 0.2 runway (PARTIAL); pure-world self-sufficiency at scale 1.0 is NOT proven; the hunt-motor fix (ff57a5a) — not the FE cure — was the binding constraint and the real survival driver. Cross-checked against OS_AGENT_STATUS_2026-07-12.md's fenced paragraph — consistent.

5. The math fence + the discipline (hard invariants — never violate)

The engine (lib/sp/brain/*): a categorical, per-factor mean-field active-inference model with tensors A (likelihood), B^u (transition-per-action), C (log-preferences), D (prior), E (habit prior). Perception minimises VFE in closed form q(s)=softmax(prior + Σ γ_m·lnA) using the bound-critical (lnB)·s message, never ln(B·s) (infer.ex, math.ex). Action minimises EFE G = epistemic H(qo)−E[H(o|s)] + pragmatic qo·C + gated parameter-novelty W (efe.ex, plan.ex, novelty.ex). Learning is Hebbian Dirichlet, no reward (learn.ex).

Hard invariants (guards in test/sp/brain/*):

1. No Nx, Rust, NIF, GPU, backprop, RL, TD, reward-on-policy. Pure Elixir plain lists, byte-comparable to the NumPy oracle.
2. Additive + GATED. Every extension behind an opt-in genome organ/field absent from default/0; graded-on coupling default 0.0; default genome byte-identical (mad < 1e-12 over the live depth-5 Plan path). Guard: decider_byte_identity_test.exs (frozen golden test/fixtures/decider_golden_seed7_d5b3.bin).
3. No scalar-per-action term in logits — u enters ONLY via B^u. Guard: action_clone_invariance_test.exs.
4. Monotonic decay of any info term: W → 0 as counts → ∞, independent of C. Guard: novelty_test.exs.

Every accepted FE term must be exactly one of: pragmatic qo·C, state-epistemic H(qo)−E[H(o|s)], parameter-novelty W, or a precision (γ/γ_m/η). Nothing else enters the logits.

Genome constructors (lib/sp/brain/genome.ex): default/0 (byte-identity anchor) · homeostat_colony/0 (the streamed-lineage constructor per doc/test; offline-green but NOT RED-validated live) · homeostat_colony_forage/1 (Cure-1, novelty ON) · homeostat_colony_forage_honest/1 (Cure-2, consummation_honest) · nursery/2 (runtime metab_scale womb/wean). Forage/honest/nursery are separate constructors precisely so the streamed genome stays byte-identical until a RED verdict + owner go-ahead flips it. Which lineage the running chip streams (UNI_LINEAGE env) is NOT VERIFIED — probe the box.

The discipline (binding):

• ONE CURE AT A TIME (First Rule). Never stack changes you can't attribute. A second cure does not deploy until the prior has a recorded verdict. If a second variable entered the comparison, the result is VOID — re-run cleanly.
• Pre-registered RED gates — named PASS + FALSIFIES in the doc before the run. Honest verdicts only.
• /lab-team-review MERGED VERDICT before ANY FE-touching merge or live RED deploy — SIGN or SIGN-WITH-CHANGES + the three artifacts (typed spec, paired RED, ship-gate checklist). Invoke it BEFORE writing FE code.
• The FOOD-HACK LESSON (never repeat): a colony was once made "stable" by RCON force-feeding — fake life; the give was removed and the claim WITHDRAWN. Viability behaviours MUST EMERGE from the generative model via EFE — no goal-coding, no reward, no gives, no props. If survival depends on a hack, it is not life; say so.
• Live-stream guard: owner go-ahead before any new lineage streams; new lineages run in separate containers with distinct kin+memory dirs, UNI_AUTOSTART=0.
• Evidence collection is continuous, lab-side/harness-managed, never inside the LLM session (survives compaction). Independent confirmation: behaviour via RCON (server-authoritative), mechanism via brain probes.

6. The deepening plan + what's next

Roadmap: docs/UNI_MISSION_DEEPENING.md, docs/DEEPENING_PLAN.md. Staleness note: DEEPENING_PLAN's "CURRENT STATUS" is dated 2026-06-24 and ends at the metabolism RED — the entire forage flow lives only in docs/receipts/* + CLAUDE.md's honest-state block. Read them together.

• P0 DIAGNOSE — DONE. Plateau = epistemic_starvation (γ unsaturated, near-flat EFE landscape). diagnose.ex.
• P1 NOVELTY (:curiosity) — DONE offline; LIVE RED = PARTIAL. Hoard-suppression PASS, plateau-break FAIL.
• P2 METABOLISM (:metabolism) — CODE-COMPLETE offline; LIVE RED split FAIL/WITHHELD. Honest predicate: "G6 not demonstrated AND metabolism activation unverified" — NOT "metabolism failed."
• Forage flow (parallel Cure-1/Cure-2, post-dates the P0–P5 ledger): reached PARTIAL — loop closed WITH a metab_scale 0.2 runway. Hunt-motor fix ff57a5a = binding constraint. Honest-consummation offline-proven, live necessity unproven.
• P3 SPINE / P4 GLANDS / P5 HEMISPHERES — design only, not built. Pre-registrations exist; runners are stubs.

Candidate next order (honor One-Cure-at-a-Time; only one RED in flight):

1. Redeploy the colony from HEAD to restore /producer/health + the Producer/Director/SP.Show tree (§3) — behind the live-stream/genome guard. A deploy fix, not a new cure, but the genome choice needs owner go-ahead.
2. Build + run forage-pureworld-graduation — the critical-path gate — real runner via /lab-team-review, at scale 1.0 with isolated per-arm worlds.
3. Defer P3+ and other REDs until the running RED has a recorded verdict.

7. WIP + queued-FE prerequisites

The 6 FE-queue SPECs (C-C1 uni_propose_change, C-C2 uni_self_audit, C-C4a mc_codec versioning, D-A3/D-B3 overlay+fence-forwarding, D-A4 LogSensor/:sensorium, D-D3 Lineage.snapshot) were spec-corrected in 250c322 (5/6 SIGN_WITH_CHANGES, 1/6 REVISE) — but that closeout was direct-apply, not independent review; each needs a fresh /lab-team-review before implementation. The commit touched no FE source.

Two LIVE-FE-code prerequisites gate the overlay/fence pair — VERIFIED, fix before those specs land:

• (a) fence.ex token gap (fence hole — high priority). lib/sp/brain/fence.ex:17's @fence regex omits "agi" (its own moduledoc line 16 claims it bans agi) and the emotion family entirely. I verified this directly: the regex ends …|breakthrough|human.?level) with no agi, no emotion. Client mirror viewer/command_center.cjs:139 has agi but not emotion. Until both live files are edited, e.g. "UNI's emotional state is calm" or "a step toward AGI" lands on the public overlay unflagged. production/schemas/claim_fence.json is the versioned fence vocabulary both JS + Elixir should load from; agreement is guarded by test/sp/brain/fence_snapshot_test.exs.
• (b) Flat-vs-envelope audit-row drift in production/mcp/server.py. Its _AUDIT.write(...) rows are FLAT and violate sensorium_envelope.schema.json (additionalProperties:false); new SP.Audit.Writer rows are conformant → prod-mcp.ndjson would carry two incompatible shapes.

Other deferred FE seams (each queues to /lab-team-review before authoring): D-A4 :os_sensorium organ has no wired effect; C-C2 genome_lineage retention in Agent.init/1; the 8 PENDING RED opt-in FE seams (depth_lineage/0, motor_shuffle_lineage/0, spine_lineage/0, hemisphere lineages, glands_lineage/0, the pure-world twin support in SP.Runtime.Lineage).

8. Repo-readiness checklist — verify BEFORE starting

• Working tree: only M viewer/overlay_proof.png (studio artifact). Nothing in lib/sp/brain/* or lib/sp/runtime/* is dirty. The two 2026-07-13 handoff docs (this one + the brain-handoff) are committed.
• 2 unpushed commits (047c355, 61765a8) — verified both studio-track (CLAUDE.md + STUDIO_AGENT_LAUNCH_PROMPT.md / STUDIO_HARDENING_DD_TDD_PLAN.md / WORKING_LOGIC.md); neither touches FE. No FE code merged in recent history.
• Do NOT claim from process existence. If you have not run the gate, the status is NOT VERIFIED.
• Verify colony state with the GATES, not process listings: RCON list on mc-server; node viewer/verify_colony.cjs [redacted: private-address] (needs /producer/health — currently down, §3); a /producer/health probe you run yourself for the driver verdict.
• Confirm the running image + streamed lineage on the chip (podman inspect uni-colony, UNI_LINEAGE env) — the live image is stale v2 (2026-06-22); do NOT assume HEAD is deployed.
• Before any FE-touching proposal: invoke /lab-team-review FIRST. Before any live RED deploy or new lineage on the streamed colony: owner go-ahead + MERGED VERDICT.
• If a RED is running: do not propose the next phase until it has a recorded verdict (First Rule).
• Read on start: docs/SYSTEM_OVERVIEW.md, CLAUDE.md, docs/LAB_PROTOCOL.md, docs/UNI_MISSION_DEEPENING.md, docs/DEEPENING_PLAN.md, the forage receipts (emergent_forage_cure1.md, forage_red_preregistration.md, forage_honest_consummation_RED.md, red_preregistration_forage_pureworld_graduation.md), and BOTH 2026-07-13 handoffs.

Source-of-truth vs rendered vs metadata

• SOURCE OF TRUTH (edit these): evidence/gates.ndjson (append-only gate ledger) · viewer/infra_registry.json (the ONLY declared name/data map) · coordination/flow.jsonl · docs/receipts/*.md (per-gate receipts).
• RENDERED (do NOT hand-edit): docs/GATES.md ← gates.ndjson · docs/gates/PUBLIC_GATE_LOG.md · the live /infra gate-ladder panel (viewer/infra.cjs).
• METADATA / versioned contracts (production/schemas/*, all schema_version:1): gate_row, sensorium_envelope, envelope, evidence_bundle, public_manifest, claim_fence.json, broadcast.

Small drift flags to fix opportunistically: (1) the 8 PENDING ledger rows read PENDING, but their pre-registration receipt frontmatter reads WITHHELD (both legal; ledger is authoritative — a rendering nit). (2) viewer/fqdn.cjs named in CLAUDE.md does not exist — fqdn()/url() actually live in viewer/infra.cjs

• viewer/discovery.cjs.

Appendix — Producer coordination snapshot (2026-07-13, measured)

Captured from the studio/producer agent's status reply. The studio-state lines are producer-reported (their surface, their measurement — not independently re-verified here; I spot-checked one, command_center.cjs:473, and it holds). Recorded so this cold-start doc is the single durable source for the new science chat.

Studio state (producer-reported): THINKER loopback UP — launcher :8090, command_center :8098, overlays :8099, publisher/cams :8443 (+regs :8095), OBS-ws :4455. DOWN — MediaMTX :9997 (no local ingest/fan-out ⇒ not broadcasting; producer restarts it via studio_up.ps1). Colony :4000 /+/stream 200; /producer/health, /overlooker, /api/state 404; cam :3020 up but STATIC.

The canonical /api/broadcast_test cannot PASS honestly right now — 3 gating reasons (owner in brackets):

1. MediaMTX :9997 down ⇒ ingest/fan-out stage fails. [studio]
2. /producer/health 404 on the stale v2 brain ⇒ stage-1 colony-health fails closed on a route regression, not a real colony problem. [SCIENCE — this is the redeploy in §3]
3. command_center.cjs:473 filters live cams on v.at but /registrations emits ageMs ⇒ liveCams always empty ⇒ camera stage enumerates no publisher (I spot-verified line 473 reads v.at). [studio, WS3]

So the redeploy (§3) is a hard dependency for the producer's broadcast test AND their pre-air puppet-cam gate — closing it unblocks two studio stages at once.

What the producer needs FROM science (the new chat's cross-track deliverables):

• The brain redeploy → /producer/health = driver=producer + frame-advance.
• The V4 / ADR-PROD-013 cam-drive decision: does HEAD's SP.Brain.Director fly the :3020 spectator (RCON-teleport the cam bot) or a different surface? Pick it explicitly so the studio builds capture to match.
• RCON :25575 LAN exposure — science's call: expose it (so verify_colony runs from THINKER) or the producer runs the count gate on-chip.
• Genome confirmation — which build/genome streams post-redeploy (your call under the live-stream guard).
• colony-scene-on-program stays fenced to forage-pureworld-graduation PASS — the producer will NOT cut the world-view to program until that verdict; shows it only as a monitored camera. Keep them posted; the gate is yours.

Boundaries (respect — producer's mid-flight surfaces at HEAD 705b1d5): do NOT touch viewer/* or the studio docs (STUDIO_HARDENING_DD_TDD_PLAN.md, WORKING_LOGIC.md, STUDIO_AGENT_LAUNCH_PROMPT.md, the brain-handoff). CLAUDE.md is SHARED — pull HEAD before editing, keep edits to science content, rebase not clobber. The producer will not touch lib/sp/brain/*, lib/sp/runtime/*, evidence/gates.ndjson, docs/receipts/*, docs/specs/*, the genome, or the REDs.

Known studio-doc drift the producer will fix (do NOT inherit it, do NOT fix it — their lane, WS0): CLAUDE.md's "Current status" block + infra_registry.json.goLiveGate still say "colony DOWN 2026-07-12/13 pending bring-up" — STALE (colony up ~13h, both tracks measured). This doc's §3 already reflects the true UP state; do not re-import the stale line.

Final reminder — the fence is the product. Every gate here demonstrates a behaviour. None of it is evidence of experience, awareness, or life. Keep the warranted claim and the over-claim visibly separate, always. Hold the fence and the vision holds with it.