[PIR] Epic: Perpetual Intelligence Runtime — ruvector work packages

Program epic — ruvector work packages. Part of the RuV Perpetual Intelligence Runtime (PIR), spanning five ruvnet repos: ruvector, LatentMesh, rvm, autogenous, RuView.

Canonical plan document

docs/research/perpetual-intelligence-runtime/03-program-plan.md in this repo is the authoritative plan. Sibling repos link to it rather than duplicating it. Supporting documents in the same directory:

• 01-evidence-review.md — verification pass over the eight founding claims. Six are grade A; two are UNVERIFIED and must never be cited as prior art.
• 02-asset-map.md — mechanism → component map, grounded in verified file paths, LOC counts, and ADR numbers after a second asset-inventory pass.

Prior art this program extends — read before writing any ADR

PIR does not originate this concept. It closes gaps that two sibling repos have already declared.

• ruvnet/autogenous ADR-401, "The Perpetual Intelligence Machine" (Status: Accepted, 2026-08-16) already defines this program's concept, including a 10-capability Built/Partial/Gap map, the operating loop, and an acceptance test. ADR-400 is Accepted — Implemented, first flywheel turn measured, and ADR-403 (verifiable execution loop) is Accepted. These are materially further along than LatentMesh's ADR-009.
• ruvnet/LatentMesh ADR-009, "Online Causal Control Loop" (Status: Proposed, 2026-08-18) specifies the execute → transfer latent state → counterfactual audit → measure causal value → update authority ceiling → persist → evolve topology loop and assigns a role to every component this program names. It is unwired — its own text states: "The statistical primitive (ADR-003) and admission gate (ADR-008) are implemented; the closed loop across live components is not wired."

PIR extends autogenous ADR-401 + LatentMesh ADR-009 and closes their declared gaps. It does not re-derive them. Any PIR ADR that restates their content as novel is wrong and should be rejected in review.

Two grounding corrections that matter when reading those repos:

• The real autogenous implementation is TypeScript in packages/radio-moe/ (~6.2K LOC source + 3.2K LOC test), not the Rust crates. Scope work against the TypeScript, not the crate listing.
• LatentMesh is ~1,407 LOC across 23 tests, with no network transport crate at all. Adopting ADR-009 means adopting a design contract, not inheriting a codebase.

Resolved: the "metaharness ADR-322" that autogenous ADR-400/401 cite has been traced — it is ruvnet/ruflo ADR-322, misattributed. See the correction block below.

What kind of program this is

PIR is consolidation and reconciliation across sibling repos, not greenfield design. Nearly every mechanism it needs already exists somewhere in the org — often in more than one place, built independently, with no integration between them. The program's real work is picking a canonical implementation per mechanism, wiring them together, and retiring the duplicates. Two concrete instances:

• Evaluation-before-promotion exists four times: ruvnet/dream-machine's engine (v0.1.0 on npm, ADR-0001 Accepted, ~2.5K LOC, 85 tests), ruvnet/ruflo's dream cycle, metaharness's Nightly Dream Cycle (docs at docs/dream-cycle/, not an ADR — the "ADR-251" citation for it is dead), and ruvector's own ADR-282 research-gate. WP2 consolidates them rather than adding a fifth.
• The cross-repo witness/promotion contract already exists: ruvnet/ruflo ADR-322 (Accepted, phases 0-2 implemented) specifies proposer/promoter separation of powers and the 322C verification stack (JCS, SHA-256, Ed25519 with domain separation, independent statistical recomputation). PIR adopts it; WP8 anchors to it instead of authoring a replacement.

Every work package should be read as "which of these do we keep, and what wires them" before it is read as "what do we build."

⚠️ ADR-322 misattribution — correct this wherever it appears

ruvnet/autogenous ADR-400 and ADR-401 cite a "metaharness ADR-322." That is wrong on two counts: ADR-322 is ruvnet/ruflo's ADR, and MetaHarness's own ADR series tops out at 250, so no ADR-322 exists there to cite. Redirect any downstream reference to ruvnet/ruflo. Correcting it at source is an acceptance criterion on ruvnet/autogenous#10.

Governing invariants

1. Every observation may change memory.
2. Every memory change is transactional.
3. Every agent communication is attributable.
4. Every behavioral mutation is tested.
5. Every promoted mutation must outperform its parent.
6. Every physical action produces new evidence.

Risk control shape: Darwin proposes, RVM gates what can change, the promotion gate requires statistically credible improvement.

Scope decision

Build 1 + 2 + 3 (self-evolving physical intelligence, persistent transactional memory, verified latent communication) as the combined PIR branch. Fast-follow with 4 (ruvLLM KV-cache migration — grade-A evidence, cheapest to ship, no dependency on the main branch). Treat 5 (universal RF/CSI vocabulary) as a separately-funded stretch bet, sequenced last.

Evidence base (correct citations — use these, not the brief's shorthand)

#	Mechanism	Source	Grade
1	SHAPER — frozen-weight skill/harness evolution	arXiv:2608.11350	A
2	WorldCycle — reversible-action verification (44% drift reduction, ~4x composite-action accuracy)	arXiv:2608.04964	A
3	LiveMem — fixed-capacity persistent memory state	arXiv:2608.02515	A
4	TARL (Transaction-Aware Reliable Ledgers) — 5-op executable memory ledger	arXiv:2608.03699	A
5	Causal audit of latent multi-agent comms (dated 2026-07-29, not Aug 5)	arXiv:2607.26773	A
6	"LATTE" anomaly quarantine	—	UNVERIFIED — no such paper
7	Cross-model KV-cache mapping (2.7–25x; 2 of 6 pairs degrade)	arXiv:2608.03893	A
8	Universal WiFi CSI "RF latent vocabulary"	—	UNVERIFIED — no such paper

Honesty constraints (binding on every ADR and issue in this program)

• Never cite "LATTE." No paper by that name on this topic exists; eight-plus unrelated papers use the acronym. WP7 is original work, informed by arXiv:2606.28958 (HMAC-SHA256 integrity manifest) for the integrity half only.
• Never present the universal CSI vocabulary as reproducing prior art. WP14 is a first-party bet; cite arXiv:2608.14694, arXiv:2604.18255, arXiv:2512.22143, arXiv:2605.14171 as motivation, not validation.
• Spell out "TARL (Transaction-Aware Reliable Ledgers)" on first use — the acronym is heavily overloaded in prior RL/robotics work.
• Date the causal-audit paper 2026-07-29.
• Name the repo when citing any ADR number. ADR-103 already means different things in ruvector vs. claude-flow/Ruflo, and this repo has ~15+ genuine duplicate ADR numbers (one number reused five times) — see WP0a. Resolved for ADR-103 specifically: ruflo ADR-322C cites "ADR-103's provider mechanism", confirming the witness-manifest ADR-103 is ruvnet/ruflo's (ruvector's ADR-103 is DeepAgents review amendments, unrelated). Keep naming the repo anyway — the habit is what prevents the next instance.
• Carry RVM ADR-285's hosted-vs-bare-metal discipline into this program's claims. autogenous self-labels "research prototype"; treat its APIs as unstable and budget for churn.

Cross-cutting security gates

• Structural frozen-weights enforcement (WP9): CI fails if any promotion-reachable mutation surface imports a training/fine-tuning API. Policy alone is insufficient.
• Witness-chain requirement: no RVF write without a corresponding RVM witness record (ADR-134 schema).
• Proof-gated promotion: the research-gate verdict is an input to ruvector-proof-gate/rvm-proof, never a bypass.
• Causal-audit CI gate (WP6): any PR touching latentmesh-align-consuming code or the quarantine module must pass a controlled-replacement audit before merge.
• Constitutional admission gate (WP11): capability-expanding mutations need a higher-bar approval, logged distinctly.

Top risks (revised after second-pass grounding)

1. Cross-repo coordination overhead is now the top risk — not Dream Machine. ruvector has zero existing wiring to either LatentMesh or Autogenous today, LatentMesh revised ADR-009's scope twice in one day, and autogenous self-labels research-prototype.
2. Two of eight founding claims are unverified. WP7 and WP14 are "build it for the first time," not "implement the paper."
3. ADR numbering hygiene is worse than expected — 283 files, true max ADR-304, ~15+ genuine duplicates including one number reused five times. WP0a must run early or this program's own ADRs will collide.
4. A dependency bug and a version split block the pipeline. GGUF weight downloads still fail in ruvllm (unexpanded glob + GGUF alias pointing at the safetensors repo), which is the real Darwin live-serve blocker — the HTTP-307 redirect bug often cited alongside it was already fixed in PR #590 (commit 946275a61, 2026-06-19); ADR-259's "blocked" wording is stale. Separately, ruvnet/dream-machine composes @metaharness/darwin 0.9.1 / flywheel 0.1.10 while ruvector pins 0.8.0 / 0.1.7. Both sit in WP0b and both block WP9.

Note: Dream Machine was the top risk in the plan's first pass. The second inventory pass found real CI-wired promotion machinery already exists under scripts/research-gate/ + crates/ruvector-sota-bench/harness (paired-bootstrap statistics, hard vetoes, Ed25519-signed replay bundles, ADR-282), which downgrades that risk substantially — WP2 is now adopt-and-wire, not build-from-nothing.

Sibling epics

• ruvnet/LatentMesh#1 — PIR: LatentMesh coordination
• ruvnet/rvm#36 — PIR: witness-chain convergence
• ruvnet/autogenous#12 — PIR: governance integration
• ruvnet/RuView#1658 — PIR: RF sensing substrate

Two further repos carry a single PIR issue each rather than an epic:

• ruvnet/dream-machine#22 — adopt the Dream Machine engine as PIR's evaluation core
• ruvnet/ruflo#3066 — ADR-322C as the canonical cross-repo witness/promotion contract

Child issues

Phase 0 — hygiene & unblocking (do these first)

• #845 — WP0a ADR numbering: freeze historical duplicates, generated INDEX.md, CI collision check
• #846 — WP0b GGUF download fix + @metaharness version bump + optionalDependencies compliance (blocks WP9)
• #848 — WP1 Adopt LatentMesh ADR-009's control loop as the runtime spine cross-repo

Phase 1

• #838 — WP2 Adopt ruvnet/dream-machine engine as the promotion gate; consolidate 4 implementations cross-repo
• #839 — WP3 Map LiveMem/TARL three-tier memory onto existing modules
• #840 — WP4 TARL five-operation ledger on ADR-194/047 proof-gated writes

Phase 2

• #849 — WP5 New crates/rvAgent/latentmesh-* crates (greenfield) cross-repo
• #850 — WP6 Causal-audit CI gate (arXiv:2607.26773)
• #853 — WP7 Latent-channel anomaly quarantine (net-new)

Phase 3

• #841 — WP9 Frozen-weight skill/harness evolution loop (SHAPER)
• #842 — WP10 Reversible-action verification stage (WorldCycle)
• #851 — WP11 Integrate autogenous's constitutional gate cross-repo
• #843 — WP12 30-day continuous-run acceptance harness

Parallel tracks

• #844 — WP13 KV-cache cross-model migration in ruvllm fast-follow
• #852 — WP14 Validate ruvector-perception against real CSI captures stretch

Sibling-repo issues

• ruvnet/LatentMesh#1 — epic: LatentMesh coordination
  • ruvnet/LatentMesh#2 — WP1 coordination request
  • ruvnet/LatentMesh#3 — WP5 wire-format compatibility review
• ruvnet/rvm#36 — epic: witness-chain convergence
  • ruvnet/rvm#35 — WP8 canonical anchor + shared record schema
• ruvnet/autogenous#12 — epic: governance integration
  • ruvnet/autogenous#10 — WP8 anchor autogenous records into rvm chains
  • ruvnet/autogenous#11 — WP11 expose constitution/admission-gate API
• ruvnet/RuView#1658 — epic: RF sensing substrate
  • ruvnet/RuView#1657 — WP14 real-CSI validation + pose-semantic pipeline
• ruvnet/dream-machine#22 — adopt Dream Machine engine as evaluation core cross-repo
• ruvnet/ruflo#3066 — ADR-322C as canonical cross-repo witness/promotion contract cross-repo

Start with docs/research/perpetual-intelligence-runtime/03-program-plan.md, then read 01-evidence-review.md and 02-asset-map.md before choosing one of the numbered child work packages. Review the referenced sibling-repository ADRs and implementation paths for the selected package. Done means a specifically scoped child issue is completed with its stated cross-repo integration, verification, and acceptance criteria; the epic itself is not a newcomer-sized change.