UNI Universal Natural Intelligence

Wiki · Evidence & Verdicts

Proposal packet — Producer remote colony-sense + observe-only fence (option A′, 2026-07-15)

Evidence & Verdicts · docs/receipts/producer_reattach_remote_sense_spec.md @ 44baf03d5041 (gen2-runtime) — opens the published snapshot ac338733bbba
2 values were removed from this page. Each one is marked in place as [redacted: category] 2 internal hostname. Nothing else was altered. The document is otherwise exactly as it is written in the repository, and the sha256 below is of the original, so what was ingested stays checkable.

How to read this page

Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.

Eighty-seven dated pages: receipts, pre-registrations, handoffs, validation records and review verdicts. A receipt is written at the moment a piece of work was checked. It names what was claimed, the commit and the seed, what was actually run, and the outcome in one of a small set of controlled words. Then it names what the work did not achieve. That last part is what makes it a receipt rather than an announcement. A pre-registration is the same discipline run in advance: the conditions that would count as a pass and the conditions that would falsify the claim are written down before the run, so neither can be adjusted once the numbers arrive.

That is why so many small dated stubs are an audit trail rather than noise. No one of them is meant to be a good read. The value is in the sequence and in the dates, because you can watch a prediction be registered, then the run happen, then the verdict land — sometimes against the prediction. Pages here record a falsified result, a rejected fix, a retracted overclaim, and a green receipt that turned out not to be reproducible from the commit that carried it. A record that carried only successes would be worth a good deal less than this one.

A gentle way in is to read a pre-registration first, so the shape becomes familiar, then a result page, then one of the corrections. This section sits off the main navigation on purpose: it is the record you check the rest of the site against, not the place to begin.

What it is not: documentation, and not a summary. Nothing here has been tidied in hindsight. Every entry reads as of its date, a later entry may overturn an earlier one, and the presence of a page is not a claim that its result stood.

Your browser cannot switch reading levels, so the document itself is shown.

Precise — the source document

This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.

Purpose: reattach a real, HEAD-code Producer UNI to the live camera of the untouchable uni-colony:v2 colony (owner directive 2026-07-14/15; handoff docs/handoffs/SCIENCE_AGENT_PRODUCER_CAMERA_REATTACH_2026-07-14.md; owner chose option A′ in chat 2026-07-15). Reviewed by the full lab team 2026-07-15 → MERGED VERDICT: SIGN-WITH-CHANGES (§3). This version integrates every required change; the draft's two falsified claims are corrected inline and noted. Canonical typed spec: docs/specs/producer_remote_sense_observe_only.md.

0. Measured context the proposal stands on (receipts in chat log, probed 2026-07-15 02:20–02:35 UTC)

  • Inside uni-colony (the ONLY --sname uni node; cookie sp): an 06-22-vintage SP.Producer (pid live) + SP.Brain.Director (pid live) run with driver: :producer and port: nil (UNI_CAM=0) — the living Producer's camera directives die in cam_write(nil).
  • uni-cam runs viewer/director.js (bot "Director", RCON pass sp verified) with no stdin writer and subject=null — a rendering, unflown camera.
  • The colony image cannot host director.js (no prismarine-viewer/canvas/gl).
  • RCON list = 6 UNIs + Director; RPC colony_count = 6 (count-consistency holds).
  • Read-only RPC into uni@uni-colony works.

1. The packet (Step-0 five elements, post-review)

Math object: None. No change to F, G (EFE), C, E, γ/γ_m/η, or learning. The change is confined to the generative process: (i) telemetry transport (where rows come from — same channels, remote source) and (ii) the Jido interpret boundary (which chosen actions may actuate). SP.Producer.plan/2, SP.Producer.Brain.*, SP.Brain.Plan/Infer/EFE/Learn/Novelty are byte-untouched. (Theorist's framing: the fence is a generative-PROCESS change — active states' world-coupling severed post-decision — never a model edit.)

Intended behavioural effect (one sentence): a new, non-uni node on the chip runs HEAD's Producer+Director sensing the REAL colony's board over a pure read-only RPC and flying its own director.js camera Port, so the broadcast camera cuts/follows the living UNIs under genuine EFE Producer control while the colony, world, and minds stay untouched.

No-go failure mode (one sentence): the new node mutates the world or colony in ANY way — a spawn/cull/kill @e reaching the live server, a body process started, a write into uni@uni-colony (including rpc to any function with side effects) — or its Producer stars phantom subjects (rows transport broken/shape-drifted).

Typed model diff: see the canonical table in docs/specs/producer_remote_sense_observe_only.md. Review corrections integrated there:

  • remote SP.Brain.Colony.snapshot/0:rpc.call(node, SP.Runtime.Board, :all, [], 500) (the draft's "pure ETS board read" claim was FALSE for snapshot/0 — its ensure_started fallback is a remote WRITE, colony.ex:65-68; falsified by the Math-Breaker, confirmed by the Theorist). Timeout/badrpc/exception → []. Rows normalised against HEAD consumer keys.
  • "the mind adapts to dead actuators"FALSE (producer genome: learn_a/learn_b/learn_e all false). Corrected: a standing, unlearnable prediction error on population/ server_health under the fence; senses honest; instrumented via per-fenced-action counters + the perseveration tripwire. At 6 UNIs population sits at its C-peak → the regime is LATENT in this RED (untested, not absent).
  • FORBIDDEN FIX (binding): no per-action scalar, no action-cost, no shipped-genome C edit to quiet fenced choices — an observer-role genome variant is lawful only as its own separate constructor through its own review.
  • Two-minds honesty line: the living v2 Producer keeps cast hands + legacy :4000/stream narration; the new fenced node is the broadcast show-runner. Only one mind has hands.

Code touch-points (exact, post-review):

  1. lib/sp/show/remote_rows.ex (new): fetch(nil) → local Colony.snapshot() under today's exception→[] semantics; fetch(node):rpc.call(node, SP.Runtime.Board, :all, [], 500), any failure → []; row normalisation (username required, kin/mode/senses/action put_new); colony_node/0 + observe_only?/0 env readers (UNI_COLONY_NODE, UNI_OBSERVE_ONLY).
  2. lib/sp/producer.ex: safe_snapshot(opts) through the seam; PURE public fence_directives(dirs, opts); interpret(state, {:fenced, dir}) → log + per-action counter (exposed in :status — the once-per-distinct dedup would hide rate).
  3. lib/sp/brain/director.ex: safe_snapshot(state.opts) through the seam (both beat clauses).
  4. lib/sp/show.ex: status/0 colony count through the seam (env-driven; local board alone reads 0 on the new node and makes gate 1 unreachable — Architect/Experimentalist).
  5. lib/sp/brain/colony.ex: spawn_agent/2 + stop_agent/1 refuse under UNI_OBSERVE_ONLY=1 (defense in depth; default unset = byte-identical).
  6. ui/lib/sp_ui/application.ex: UNI_POPULATE gate (default "1" = today) + show_opts/0 passing colony_node/observe_only/mc_host into SP.Show.Supervisor.
  7. viewer/director.js: stdin-EOF self-exit (single-camera invariant across supervisor restarts; studio-track edit under the owner's 2026-07-15 waiver).

Deployment shape: see canonical spec. Load-bearing env (attested in deploy receipt): UNI_OBSERVE_ONLY=1, UNI_POPULATE=0, UNI_COLONY_NODE=uni@uni-colony, --sname producer (one-uni-law), MC_HOST=mc-server, VIEWER_URL → the :3020 LAN path, host :4200→:4000 ([redacted: internal-hostname]; NOT :4100 = masterplan — registry collision caught 2026-07-15).

RED test design (paired, post-review):

  • Arm A (control) = the measured current state (§0): camera parked, wheel unplugged, 6 UNIs.
  • Arm B (cure) = uni-producer up + uni-cam stopped — ONE composite ownership variable (MC kicks duplicate "Director" logins; ownership physically cannot be shared). Nothing in uni-colony/mc-server/minds changes.
  • Window preconditions (gate 0): RCON list shows exactly one Director; podman ps shows uni-cam absent; env attestation captured (podman inspect uni-producer).
  • Collector (Lab Protocol III — harness-managed, no LLM-watched window): runs/red_producer_camera_collector.sh on THINKER (ssh, read-only), sampling every ≤5 s for the whole window, appending timestamped NDJSON: (a) new-node /producer/health INCLUDING the frame-stamped knowledge ring (instantaneous star aliases under a 5 s poll) + fenced counters; (b) RCON data get entity Director Pos + data get entity <current star> Pos; (c) RCON list (names + count — the abort tripwire); (d) legacy :4000 liveness.
  • PASS gate producer-camera-attached (ALL of):
    1. Fresh GET http://[redacted: internal-hostname]:4200/producer/health → 200, verdict=LIVE, driver=producer, seam-joined colony_count=6, frame advancing across two probes ≥60 s apart.
    2. Over a ≥10-minute window: ≥3 discriminating subject-REATTACHMENT events — a star/glide directive (from the knowledge ring) where old→new star separation ≥ 2× the reattach threshold, followed within ≤15 s by dist(Director, new star) ≤ shot r + margin (≤12 blocks) in the Pos series. Wide-shot fallback: one :overview altitude event (Director y ≥ subject y + 10) counts if the colony clusters below discrimination all window. (Orbital motion alone is V4-blind — glide() tps continuously once any subject was ever set.)
    3. RCON list UNI-count/names identical before/after (no entity added/removed by the new node); zero fenced directives EXECUTED (counters may increment; effects may not). ALLOWED RCON writes = the camera's proven verbs on Director only (tp, gamemode spectator, forceload); FORBIDDEN = anything touching UNI-* or entities/items.
    4. Legacy :4000 narration liveness across the window (colony untouched).
    5. Supervised-restart repetition: mid-window, kill SP.Producer (or restart the container once); rest_for_one recovers; ≥1 further reattachment event AND still exactly one Director in list (single-camera invariant live-tested).
  • FALSIFIES (ANY of): directives issued with no reattachment (the pre-registered V4 outcome — surfaced, not masked); driver≠producer; any new-node world/colony mutation; rows=[]/colony_count=0 while the legacy node still serves (transport broken); Director join/leave flapping (kick-fight = orphan camera = two variables).
  • INCONCLUSIVE (not FAIL): zero star/glide directives across the window (a calm colony under hold/narration beats has no bounded-time cut guarantee) → extend to 30 min → verdict WITHHELD if still zero. Perseveration tripwire: K=40 consecutive beats electing the same fenced action with zero camera directives ⇒ abort + INCONCLUSIVE, never a masked FAIL.
  • Mid-run ABORT (operator holds the verb): UNI count/name drift attributable in-window; Director flapping; any evidence of an unfenced spawn/cull/kill.
  • Rollback: stop uni-producer, restart uni-cam + original forwarder — restores Arm A exactly (both non-colony plumbing).

2. Ship-gate checklist

  • MERGED VERDICT SIGN-WITH-CHANGES recorded (§3); all 19 required changes integrated.
  • mix test test/sp/brain/ green (byte-identity, action-clone-invariance, novelty).
  • New unit tests green: fence_directives/2 identity on empty opts; fence replaces exactly {:spawn}/{:cull,_}/{:health,:tps}; RemoteRows.fetch(nil) local incl. failure branch; unreachable node → []; row normalisation.
  • Live pre-deploy receipts: one rpc probe pinning SP.Runtime.Board.all/0 present + side-effect-free in the v2 image's loaded code, capturing one real row's keys.
  • Gate row producer-camera-attached appended PENDING to evidence/gates.ndjson with pre_registration_path = this file, committed, pushed, BEFORE deploy.
  • Every chip mutation individually owner-approved; colony containers untouched (capture-before-destroy NOT triggered — if that ever changes, STOP and run docs/handoffs/GAIA_CAPTURE_BEFORE_DESTROY_2026-07-14.md first).
  • Verdict + receipts to docs/receipts/, gate row superseded with the verdict, every counted event re-derivable from the collector NDJSON + commit hash.

3. Review verdicts (lab-team adversarial review, 2026-07-15, protocol order)

  • Math-Breaker: SIGN-WITH-CHANGES — "Math object: None" survives all 8 checks (no new logit term, no reward channel, colony-agent math untouched, defaults byte-identical). Falsified the draft's pure-read claim (remote Colony.snapshot/0 = remote write via ensure_started); required MB1-MB5 (rpc retarget to Board.all/0 + 500 ms timeout + shape receipt/normalise + failure-branch anchor + load-bearing env receipted).
  • Systems Architect: SIGN-WITH-CHANGES — placement honest (lib/sp/show is the declared plumbing home; call sites stay in scanned dirs); supervision sound (rest_for_one: Producer crash leaves the camera up; driver re-asserts :producer); one-uni-law holds. Required A1-A6 (adopt MB verbatim + pin MFA live; lib/sp/show.ex joins the seam — local count reads 0 and gate 1 was UNREACHABLE as drafted; spec into docs/specs/; single-camera invariant across restarts; fence the local facade; deployment completeness incl. VIEWER_URL and the allowed-vs-forbidden RCON enumeration).
  • RED Experimentalist: SIGN-WITH-CHANGES — paired skeleton genuine, measurement layer rebuilt: E1-E6 (harness collector w/ frame-stamped ring; subject-REATTACHMENT metric on discriminating pairs — orbital motion is V4-blind; static-camera split into V4-FAIL vs INCONCLUSIVE→WITHHELD; seam-joined gate 1; restart repetition + preconditions + abort tripwires; evidence-bundle completeness + pre_registration_path).
  • Embodiment Designer: SIGN-WITH-CHANGES — no need proposed, no preference smuggled; fence is post-decision severance. Required D1-D4 (Learning row corrected — learn flags all false, so the standing prediction error is UNLEARNABLE; perseveration regime named + instrumented + tripwired, latent at the 6-UNI C-peak; forbidden-fix fence; two-minds honesty line).
  • AIF Core Theorist (merger): MERGED VERDICT = SIGN-WITH-CHANGES — 19 consolidated changes (all integrated above); reconciliation: the Math-Breaker/Embodiment contradiction on the Learning row resolved with code (genome.ex learn flags all false → Embodiment correct); A2≡E4 merged; A4/E5 complementary. "The rare clean case: a genuinely model-free change — transport + post-decision actuator severance — additive, opt-in, default byte-identical, claim fence intact, with its one honest cost named and instrumented rather than papered over."

Required follow-on artifacts: typed spec docs/specs/producer_remote_sense_observe_only.md (landed) · paired RED = §1 above (this file is the pre_registration_path) · ship-gate checklist = §2 above.

sha256 7dcbb4a6bc94a878 — of the original file, so what was ingested stays checkable.

Plain — written for this website, not the source document

Written for this website — not the document. This is a plain-language retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. (or choose Precise in the reading-level control above)

A proposal packet, reviewed and signed with changes, for attaching a real show-runner to a live camera without touching the colony it films. Two claims in the draft were falsified during the review and are corrected in place, with the correction noted rather than hidden. What had been called a pure read would in fact have written to the other side. And a claim about the mind adapting to dead actuators was false, because that configuration does not learn. Nothing in the model changes; the block applies after the decision.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 7dcbb4a6bc94a878

Clear — written for this website, not the source document

Written for this website — not the document. This is a clearer retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. (or choose Precise in the reading-level control above)

A proposal packet rather than a receipt of something already run. It describes work not yet done, carries the measured context it stands on, and records the review that signed it with changes.

The measured context comes first, with the time it was probed, including the state of the two components involved and one constraint that shapes everything: the image hosting the colony cannot host the camera at all.

The packet is then written in five declared elements. The maths object is none, which is the central claim: the model is untouched, and the change is confined to where the observations come from and to which chosen actions may reach the world. The intended effect and the no-go failure mode each get one sentence, which forces both to be sharp.

Two draft claims are corrected inline, with a note of who falsified them. What had been described as a pure read turns out to call something that writes when it finds nothing running, so it is retargeted at a genuinely read-only call with a timeout and a normalisation step. And a claim that the mind would adapt to dead actuators is false, because that configuration does not learn. The honest description is a standing prediction error that cannot be learned away, instrumented with counters and a tripwire, and named as latent under current conditions rather than absent.

A binding paragraph forbids the tempting fix. No per-action cost, and no editing preferences to quiet the choices the block refuses, because that would smuggle something into the model to hide a consequence; a separate variant would be lawful only through its own review. An honesty line records that two minds now sense the same colony while only one of them has hands.

The code touch-points are enumerated exactly, including one that exists only because the review found a gate would otherwise be unreachable, and one defence-in-depth refusal that leaves the default unchanged.

The test design is paired, with the control being the current measured state and the treatment one composite variable, because ownership physically cannot be shared. Collection is harness-managed rather than watched by a person, sampling on a fixed cadence into a timestamped file. The pass conditions are several and specific, including a definition of a discriminating event that rules out a camera merely orbiting, and a wide-shot fallback for when everything clusters too closely to tell apart. The results that would show it wrong are listed, and separately an inconclusive category, so that a calm colony issuing no cuts does not become a masked failure.

A ship-gate checklist follows, mostly unticked, with the conditions required to be written down, committed and pushed before any deployment. Then the reviewers' verdicts one by one, each with what it required, and the merged verdict with a count of consolidated changes. The merger's closing description is the one to read: a genuinely model-free change, additive, opt-in, byte-identical by default, with its one honest cost named and instrumented rather than papered over.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 7dcbb4a6bc94a878