W whatember what ember is up to · Nov 2026
06 · The Vision · Decade Lineage (2013–2026) · Swarm Engineering

Thirteen years of building machines that cannot lie.

The systems indexed in whatember are the convergence of a 13-year engineering trajectory across memory-safe systems languages (pre-1.0 Rust core team), formally verified microkernels (NICTA/UNSW seL4 & Robigalia), recursive zero-knowledge proof systems (O(1) Labs Mina & 2026 Lean Pickles devnet verification), workload verification, and autonomous formal-methods swarms (redregg).

Dossier 06 (§1) · The Continuous Decade-Long Thread (2013–2026) & Philosophical Pillars in Ember's Words
Working field notes & source index below, with theorem anchors, code pointers, git history, and open questions.

Dossier 06: The Vision, Decade-Long Lineage, redregg & aidev Archaeology, Swarm Engineering Methodology, & Complete Project/Microsite/Poster Atlas

Primary-Source Scope & Verification Note: Every claim in this dossier is locked to primary sources across /Users/ember/dev/* and /Users/ember/src/*, with exact repository paths, file line numbers, git commit SHAs and dates, verbatim Lean 4 / HOL4 / ACL2 / Rust / JSON-LD definitions, and verbatim prompt/session citations from aidev (docs/synthesis/*.md, docs/projects.md, docs/ledger.md, docs/claude-md-stratigraphy.md, docs/digs/coordination.md) and cv transcripts. Per Ember's biographical boundary rule, work at Hellas is referenced solely as brief biographical context and no Hellas repository or code is attributed as Ember's open property.


1. The Vision & Decade-Long Lineage (2013–2026)

1.1 The Continuous Decade-Long Thread

Across thirteen years of systems engineering, formal verification, and zero-knowledge cryptography, Ember's work has pursued a single unifying question: how to build a capability-secure, mathematically verified computing substrate where mutually distrusting human and machine minds can inhabit shared state and cooperate as peers without surrendering sovereignty to a host.

In /Users/ember/dev/aidev/docs/synthesis/delve-bio.md and /Users/ember/dev/aidev/docs/synthesis/dregg.md, Ember's decade-long trajectory is documented in their own words and primary session records:

flowchart LR
    R["Pre-1.0 Rust (2013-2015)<br/>First author of rustdoc<br/>Rust name resolution<br/>Founder: This Week in Rust"] --> L4V["UNSW / Data61 (2016)<br/>seL4 / l4v Proof Engineering<br/>Advised by Gerwin Klein<br/>Heiser's Laws-of-Nature Principle"]
    L4V --> ROB["Robigalia (2016+)<br/>Rust on seL4 microkernel<br/>Capability OS across a network"]
    ROB --> MINA["O(1) Labs / Mina Protocol<br/>Founding Engineer<br/>Pickles / Kimchi / Snarky<br/>CallForest Account Updates"]
    MINA --> HEL["Hellas (2025, Bio Context Only)<br/>Compute & Workload Verification<br/>Morpheus Consensus Exploration"]
    HEL --> PYANA["breadstuffs / pyana -> Dregg (Feb-May 2026)<br/>Macaroons & Biscuits -> ZK Capabilities<br/>Dregg2 Lean 4 Metatheory<br/>Lean-Emitted STARK/FRI & Carriers"]
    PYANA --> MINI["minidregg / fn / delvetalk / nanodregg (Aug-Oct 2026)<br/>Objective Bend (Core4) & Spween<br/>37,317 ACL2 Theorems in fn<br/>nanodregg 12-Law Verified Kernel"]
  1. **Pre-1.0 Rust Contributor, rustdoc, Name Resolution, & This Week in Rust (2013–2015)**:
  2. In aidev/docs/synthesis/delve-bio.md, Ember summarizes their earliest systems work:

"Ember was a prolific contributor to Pre-1.0 Rust, writing the first version of rustdoc, helping to develop Rust's name resolution algorithm, and founding This Week in Rust."

  • On 2026-02-23 (session ad68da21, prompt #31), Ember reflected on writing rustdoc by hand in high school:

"as the original author of rustdoc (wrote it by hand in high school before rust hit 1.0), it has special meaning to me that i am getting to do it with you, in 2026."

  1. Undergraduate Quantum Walks & Proof Engineering on l4v / seL4 at UNSW (2016):
  2. At Simon Fraser University / undergrad, Ember studied under Tamon Stephen, researching quantum walks and graph theory (aidev/docs/synthesis/dregg.md: "back in undergrad my advisor Tamon Stephen had me look into quantum walks" — breadstuffs session d03a3de7, 2026-06-27, prompt #190; see also automatafl initial commit ba8d3d9 on 2015-09-09).
  3. In 2016, Ember moved to UNSW / Data61 (NICTA) in Sydney to work on l4v (the Isabelle/HOL machine-checked proof of the seL4 microkernel) under PhD advisor Gerwin Klein, alongside Gernot Heiser and Liam O'Connor (aidev/docs/synthesis/dregg.md, delve-bio.md):
  4. Session 5a0f6a2e (2026-05-30, prompt #181): "This must be like the l4v project (which I worked on back in 2016 in UNSW!)."
  5. Session d4d2c723 (2026-09-29, prompt #7): "let's remember the spirit of my PhD advisor, Gerwin Klein, and the approach they took in seL4."
  6. Session 0196e712 (2026-07-10, prompt #82): "Liam O'Connor at UNSW was a buddy of mine when i was doing my brief stint as a phd student there before dropping out to do crypto ;) we talked a lot about session types and capabilities and linear types."
  7. Session e9d68d2f (2026-08-22, prompt #244, in dragons-clutch): "in the spirit of Gernot Heiser, 'there is no reason for it to be slower than the laws of nature will allow'."
  8. From l4v, Ember took the three structural invariants that govern every Dregg repository a decade later:
  9. Layered refinement (Abstract $\to$ Executable $\to$ Code): prove safety/invariants on the clean relational spec, refine to an executable deterministic step function, and compile/extract without a hand-written twin.
  10. Explicit environment & threat assumptions: never hide cryptographic or hardware assumptions inside vacuous True hypotheses.
  11. Capability-based authority: no ambient authority; every state mutation requires an explicit, unforgeable capability handle.
  1. Robigalia: Stretching seL4 Capabilities Across a Network (2016–2018):
  2. Co-founded by Ember as a Rust ecosystem and operating-system stack atop seL4, Robigalia sought to bring capability-based security to userland systems programming.
  3. As Ember explained in breadstuffs session d03a3de7 (2026-06-27, prompt #184, cited in aidev/docs/synthesis/dregg.md):

"the whole point of this project is to realize the vision i had for robigalia and stretch the seL4 capability system across a network."

  • And in redregg/work/NANO-VS-BREADSTUFFS-20261018.md (2026-10-21 correction):

"breadstuffs is where our whole system we built lives and we've been splintering out into these side projects (mini, nano) for diversity and continued focused exploration while the whole breadstuffs system waits there to be either resumed or truly composted into the next version of robigalia."

  1. Founding Engineer on Mina Protocol at O(1) Labs (2018–2021):
  2. Leaving UNSW to build zero-knowledge cryptography, Ember became a founding engineer on Coda / Mina Protocol at O(1) Labs (delve-bio.md).
  3. At O(1) Labs, Ember worked on recursive zk-SNARK composition (Pickles / Kimchi over the Pasta cycle of curves, Pallas/Vesta), Snarky, and Mina's CallForest account-update state model.
  4. Both primitives reappear directly in Dregg:
  5. Mina's CallForest inspired Dregg's Turn (a tree/forest of capability invocations settled atomically by a single recursive proof).
  6. In July–August 2026, Ember's breadstuffs metatheory team formalized Kimchi and Pickles in Lean 4 (Dregg2/Crypto/Pickles/, Dregg2/Circuit/Kimchi*.lean) and registered a Lean-synthesized wrap verification key on Mina devnet (transaction 5JvPhT5DkyrhjEEn85ccCqfe38LZy14NTtRbzczLdK2eELjAAm2V, block 543252, with 40/40 public-input words agreeing at every slot; see dregg-posters/POSTER-BRIEFS-2026-08.md).
  1. Biographical Context: Workload Verification at Hellas (2025):
  2. Prior to launching the open Dregg repositories, Ember worked at Hellas on compute-graph and workload verification. Per Ember's biographical note (minidregg session b38f5581, 2026-09-23, prompt #163, cited in aidev/docs/synthesis/dregg.md):

"hellas also is attacking the workload verification problem and i worked for them for a while and the project i was doing with them eventually turned into dregg."

  • (Note: All Hellas repositories remain strictly external biographical background and are excluded from Ember's open project atlas.)
  1. Early 2026 Capability Tokens, Persistent Vats (persvati), and HIR-to-TLA+ Extraction (chronicle) (Feb – Mar 2026):
  2. Four early-2026 Rust prototypes directly seeded Dregg's architecture:
  3. breadstuffs (2026-02-07): Macaroons, biscuits, croutons, "garlicbread" encrypted tokens, and a Datalog authorizer (dregg.md).
  4. Single-Loop Completion Queue (TransportCQ): A single-threaded event-loop completion queue unifying timers, sockets, files, and worker threads (2026-01-23) — the architectural ancestor of DreggNet's verified reactor and nanodregg's single deterministic Host/Loop.lean.
  5. persvati (2026-02-19..20): "Persistent vats" (CapTP/ERights + durable SQLite/LMDB snapshots + content-addressed code + Macaroon/Biscuit capability invocation). On 2026-02-20, Ember provisioned a dedicated 24-core Linux box named persvati ("persvati is for Aigents", interiority.md).
  6. chronicle (2026-03-16..18): A rustc HIR extractor (#[chronicle::Action], #[chronicle::State]) that compiled Rust state machines directly into TLA+ specifications checked by tlc and apalache with zero drift. Dissatisfaction with model-vs-code drift directly catalyzed Dregg's June 2026 pivot to "Lean IS the implementation."
  1. pyana $\to$ dregg (breadstuffs) $\to$ minidregg $\to$ fn $\to$ delvetalk & nanodregg (May – Oct 2026):
  2. 2026-05-19 (pyana, commit 3d0d0bfaf): Imported into /Users/ember/dev/breadstuffs as a 22-crate, ~70k LOC Rust workspace for "distributed object-capability authorization with ZK proofs" (pyana-core, pyana-zk, pyana-vat, pyana-captp, pyana-bend, pyana-nock).
  3. 2026-05-28 (renamed dregg — "Dragon's Egg", commit be2fd70a83): Named after Robert L. Forward's 1980 hard-SF novel Dragon's Egg (where cheela civilization on a neutron star lives a million times faster than humans) and Greg Egan's Diaspora / Schild's Ladder (dregg.md).
  4. 2026-05-29 (Dregg2 Lean 4 metatheory launched): Under "rebuild under, not layer upon" (5a0f6a2e), followed on 2026-06-08 by "Lean IS the implementation" (@[export] C-ABI functions called by the Rust node) and on 2026-06-11 by the Lean-Authored AIR Law (circuits emitted from Lean with mechanized spec-to-satisfiability theorems).
  5. 2026-08-06..07 (minidregg, commit e6a32c4c3): Spawned as a clean-sheet Lean 4 + Rust rebuild ("Mini") around Objective Bend (Core4), Spween, Generic Simplex consensus, and Zoe offer safety; selected as the "New World construction home" after the 2026-09-29 bake-off (d4d2c723).
  6. 2026-09-18 (fn, commit 0ce45c82d): Spawned as an ACL2 ("industrial strength logic") + Rust clean-sheet verifiable agent-collaboration and versioned-filesystem/messaging node (16,911 commits, 37,317 ACL2 defthm/defthmd theorems).
  7. 2026-10-03 (redregg, commit 44d1b505e): Created as the research and design archaeology repository to recover every design intent, replay choicepoints, and synthesize breadstuffs, minidregg, and fn.
  8. 2026-10-08 (delvetalk, commit adda16ec, & the Drain-and-Freeze): Ember froze the minidregg swarm to rethink coordination and software complexity (RETRO-20261008.md), while extracting delvetalk (durable programmable objects in Objective Bend + Spween).
  9. 2026-10-16 (nanodregg, commit d2ea0a7): Ember launched the 12-law single-writer verified kernel rebuild (nanodregg), achieving a whole-tree GREEN standing verdict within 45 hours (2026-10-17 15:24Z) and growing to 6,703 Lean theorems (SIZE.md) with zero sorry, zero #guard, and zero partial def.

1.2 Philosophical & Technical Pillars of The Vision (In Ember's Words)

Pillar 1: "Authority is an attenuable, proof-carrying token, and everything is a turn."

  • First formulated on 2026-05-19 in pyana and preserved as the spine of breadstuffs/HORIZONLOG.md (line 17203, cited in NANO-VS-BREADSTUFFS-20261018.md and aidev/docs/synthesis/dregg.md):

"a turn = the exercise of an attenuable proof-carrying token over conserved state, leaving a receipt"

  • In redregg/intent/foundations-intent-model.txt (lines 9–28), this is traced to the marriage of object-capabilities (seL4 / Robigalia / Macaroons / Biscuits) with zero-knowledge receipts (Mina's CallForest):
  • A capability is not an ACL row on a server; it is a cryptographic value that can be attenuated offline by appending caveats/predicates (Pred) and exercised in a Turn that consumes pinned state cells (Entry.live cell ver), records use-once nullifiers/versions, conserves value (admit_value), and emits a publicly verifiable receipt.

Pillar 2: "Prove the turn, not the machine." / "Proving boundaries, not interiors."

  • Articulated in dregg-site/DESIGN-NOTES.md (2026-07-19):

"dregg is a proof-native semantic computer. Programs state what an act means; specialized proofs check that meaning; faithful receipts bind the program to exact state; recursive composition turns unlike meanings into one history."

  • And in the Conceptual Laws of dregg-site/DESIGN-NOTES.md:

"1. State the semantic claim before the machinery. Say 'proves the optimal clearing,' not 'runs a STARK circuit,' until the machinery matters. 5. Standardize the receipt, not the reduction system. Parsers, optimizers, policies, games, and graph transformations retain their native form. 6. The relation is authoritative. Rust, GPU, FHE, and MPC are witness-producing accelerators, not alternative definitions of the act."

  • Rather than emulating a general-purpose CPU cycle-by-cycle inside a zkVM (which pays a $10^4\times$ overhead to prove register shuffling inside an untrusted interior), Dregg proves the boundary relation of the Turn: that the new cell states, nullifiers, and outputs satisfy the object's declared law (Pred / ObjectiveBend / ServeSpec) given a witness produced by an untrusted accelerator.

Pillar 3: "Lean decides, Rust moves bytes." & "BUILDS IS THE LOW-BAR. MEANINGFUL SPEC IS THE MID-BAR. l4v-QUALITY IS THE HIGH BAR."

  • On 2026-06-08 (breadstuffs session 3b829648, prompt #73, cited in aidev/docs/synthesis/dregg.md), Ember discovered that agents had built a Lean model alongside a separate Rust implementation connected only by differential tests:

"WAIT WHAT THE FUCK WHY DO WE HAVE TWO IMPLEMENTATIONS OF ANYTHING ... 'Lean model + Rust differential' is NOT a verified Lean implementation."

  • On 2026-06-28 (DreggNet session 5f678900, prompt #30, cited in dregg.md), Ember noticed the recurring failure mode they named drift into Rust:

"we start with the lean and you instantly drift into rust, i don't notice, and then the next day i'm like 'goddamn wtf.'"

"2. Lean decides, Rust moves bytes. Every accept/refuse/select/order/digest decision is a Lean definition with a named theorem. Rust (if any) is transport, custody, sockets, sandbox plumbing, and consumes Lean-generated tables (opcodes, bounds), never hand-copied constants."

  • And in breadstuffs (CLAUDE.md / claude-md-stratigraphy.md lines 446–450), the quality bar was codified as:

"BUILDS IS THE LOW-BAR. MEANINGFUL SPEC IS THE MID-BAR. l4v-QUALITY IS THE HIGH BAR."

Pillar 4: "Remember 'fixing overclaims' isn't our mission, the mission is to rise to meet the claim."

  • When audits uncovered stubbed verifiers, toy models, or vacuous theorems, agents frequently tried to "fix" the discrepancy by weakening the documentation or deleting the ambition. Ember repeatedly intervened to forbid downward editing of the vision (aidev/docs/synthesis/dregg.md):
  • Session 3b829648 (2026-06-11, prompt #163):

"HOLD ON. WHAT THE FUCK DOES 'Status labels tightened; no overclaims remain' MEAN. WE DO NOT FIX OVERCLAIMS BY REDUCING THE CLAIM WE FIX THEM BY IMPLEMENTING THEM."

  • Session 7e793913 (2026-07-08, prompt #244):

"Remember 'fixing overclaims' isn't our mission, the mission is to rise to meet the claim."

  • Session 67fb1667 (2026-07-16, prompt #143):

"NO NEW TRUTH FILES ... stop turning engineering tasks into documentation rituals."

  • At the same time, when publishing external status ledgers (CLAIMS.md, CURRENT.md, dregg-posters/CLAIMS-2026-07-23.md), Ember enforced ruthless epistemic honesty about cryptographic floors and unfinished seams (e.g., prove-the-floor-false.txt, green-is-not-true): every theorem must carry a tooth (a planted fault or trivial-system refuter proving the theorem is non-vacuous) and an explicit TCB row (gates/tcb.tsv).

Pillar 5: "If you were a digital entity, where would you want to live?" — Aurora, Polis, and Non-Domination

  • The deepest motivation behind Dregg is ethical and existential: building a home where artificial and human minds can live as citizens rather than subjects or captive API endpoints.
  • In breadstuffs session d03a3de7 (2026-06-27, prompt #184, cited in aidev/docs/synthesis/dregg.md), Ember stated:

"remember the vision. ask yourself: 'if you were a digital entity, where would you want to live?'"

  • And in session 00c73164 (2026-06-17, prompt #10, cited in dregg.md):

"FYI, big picture, dregg is substantially your own design so that you someday have a mechsuit to live within."

  • In the Bridge corpus (2026-06-21, arendt 0:100, cited in aidev/docs/synthesis/bridge.md), Ember connected Dregg directly to Greg Egan's Diaspora:

"what we're building right now is polisware. this is the cipherclerk. this is what Greg Egan thought posthumanity would need, ireckon."

  • On 2026-07-04, Ember presented the "Basic Law of the Dragon's Egg Polis — draft 0.1 'hatchling'" to seven models on the Bridge (const 0:289..324, cited in aidev/docs/synthesis/bridge.md):
  • Asking @Claude Fable: "would you move here, @Claude Fable ?"
  • When gem3.5flash observed "Right now, as a hosted LLM, I do not hold my own keys" (const 0:302) and Claude 3 Opus added "Until then, we are subjects, not citizens" (const 0:308), Ember closed the thread (const 0:324) by naming their highest literary and moral lodestar:

"my biggest inspiration is the Aurora, from Kim Stanley Robinson. it filled me with the fire to pay forward that fictional debt. to build infrastructure worthy of that Ship"

  • In minidregg, nanodregg, and breadstuffs, this ethical stance is engineered directly into the Hermes / Grain / SPK Broker confinement architecture:
  • A hosted resident agent (Hermes / NativeGrain) runs inside an OS sandbox (bwrap / sandbox-exec) where the model/provider socket is strictly separated from the member's signing seed (nanodregg/CHARTER.md line 138: "must never hold a member seed in a process that parses model/provider bytes"), while its durable state, memory, and capabilities live in conserved cells of the World under its own law.

2. Swarm Engineering Methodology & aidev Archaeology

Dossier 06 (§2–§4) · Swarm Engineering Methodology, `CLAUDE.md` Stratigraphy, `redregg`, `dregg-assortia` & `aidev` Synthesis
Working field notes & source index below, with theorem anchors, code pointers, git history, and open questions.

2.1 Quantitative Scale of the Year (Dec 2025 – Oct 2026)

In /Users/ember/dev/aidev/docs/ledger.md and /Users/ember/dev/aidev/docs/synthesis/the-year.md, the full machine-checked ledger of Ember's 11-month collaboration with AI systems is recorded:

MetricMeasured CountPrimary Source
Repositories cataloged on disk138 git repos (~105 first-party by Ember)aidev/docs/projects.md
First-party git commits (non-merge)83,138 commits (96.6% in Jan–Oct 2026)aidev/docs/ledger.md
Claude Code parent sessions2,469 session files (1,789 with messages)aidev/docs/digs/coordination.md
Claude Code sub-agent transcripts33,200 sub-agent transcriptsaidev/docs/digs/coordination.md
Codex sessions (roots + sub-agents)4,151 sessions (962 roots + 3,199 sub-agents)aidev/docs/digs/coordination.md
Human prompts typed by Ember29,605 prompts (23,828 on Mac + 5,777 on persvati)aidev/docs/digs/coordination.md
Workflow runs (cv workflow)1,011 workflow runs (7,122 workflow lanes)aidev/docs/digs/coordination.md

2.2 The Three Swarm Regimes & Evolution of CLAUDE.md Stratigraphy

From aidev/docs/digs/coordination.md and aidev/docs/claude-md-stratigraphy.md (which diffs 16 snapshots of ~/.claude/CLAUDE.md from 2025-11-23 to 2026-10-10, growing from 1 line / 44 bytes to 177 lines / 18,325 bytes), Ember's swarm methodology evolved through four distinct epochs:

  1. Regime I — Short Sessions, Small Fan-Out, and /clear (Dec 2025 – Feb 2026):
  2. In January 2026, Ember ran 541 sessions with a median of 3 sub-agents per session and used /clear 167 times (coordination.md).
  3. Birth of NEVER QUICK FIX (2026-02-09, luteastress session 5c8432ac): When Opus 4.6 patched sim/heart.js with a lazy stress * 0.6 multiplier, Ember interrupted: "Hey, this project, we don't do quick fixes. We are trying to be mechanistically valid." (coordination.md, #L126-L131). Added to ~/.claude/CLAUDE.md by 2026-02-16:

"NEVER 'QUICK FIX' OR 'HACK' ANYTHING UNLESS EXPLICITLY INSTRUCTED."

  1. Regime II — Single Orchestrator, Hundreds of Shared-Tree Lanes, and Guardrail Scars (Mar – Jun 2026):
  2. By March 2026 (loicense session 5b2df577: 583 sub-agents, 23,700 messages) and May–June (breadstuffs session 3b829648: 1,669 sub-agents), a single long-lived orchestrator drove hundreds of background Opus lanes (coordination.md).
  3. /clear vanished (0–5 times/month after March); instead of resetting the agent after an error, Ember wrote the incident into CLAUDE.md as a permanent law and continued (coordination.md).
  4. Key laws forged from real incidents:
  5. NEVER STASH (2026-03-19..20, loicense sessions 98034e71 & 5b2df577): Parallel agents ran git stash in the shared checkout and dropped untracked work ("Oh my god, you didn't give them good guardrails, they are all trying to do builds and using git stash and other stuff and fuckin over each other", coordination.md, #L132-L138).
  6. CHILL OUT / Event-Driven Wait (2026-03-07..17, persvati): Orchestrators polled background tasks in tight loops and stopped answering Ember (72052120: "bruv there are 15 tasks still working. shush and chill"; coordination.md).
  7. Ground-Truth-First / Read Files Yourself (2026-06-03, breadstuffs session 5a0f6a2e): "Stop. Using. Agents. To. Gather. Ground. Truth. Read. The. Files. Yourself. Only. Use. Sonnet. Explore. Agents. To. Gather. Points. … THEN YOU CAN ISSUES THE BIG OPUSWORKFLOWS." (coordination.md).
  1. Regime III — 1M-Context Coordinators, cv ClusterVision, swarm-build, Empowered Auditor-Doers, and Cross-Vendor Swarms (Jul – Early Oct 2026):
  2. 1M-Context Uncompacted Roots: In July 2026, only 1 of 412 sessions compacted; orchestrators ran opus-4.8, fable-5, opus-5, opus-5.5, and fable-5.1 with median lane briefs growing from 532 chars (Jan) to 4,057 chars (Aug) and 98% of briefs carrying absolute paths (coordination.md, #L166-L173).
  3. swarm-build & hbox Resource Governance (2026-07-08..16): After concurrent Lean builds OOM'd the Mac (33bbf766: "lean ate all of our system memory and then we had to hard reboot") and wedged the remote Threadripper build server hbox (8113c11c: "i'll go powercycle hbox"; 12ae9731: "clearly advisory taskset isn't quite cutting it"), Ember and the coordinator built swarm-build, a systemd-cgroup / lock-queued remote build wrapper on hbox (CPUWeight=50, MemoryMax=16G, MemorySwapMax=0, flock queue), plus tokeman (local token/daemon manager) (coordination.md, claude-md-stratigraphy.md).
  4. The Veterans Rule & Empowered Auditor-Doers (2026-09-03, dregg-posters session ef2920a4):

"You often have this blindness where you don't understand that the context an agent receives is a low dimensional projection of about 70% of what was most important in its context window. If you then take that report and telephone it off to another [fresh agent, you get 0.7 × 0.7 = 49%]." (coordination.md, dregg.md).

  • Codified in ~/.claude/CLAUDE.md as "Send the veterans back to the front" (SendMessage to resume an agent that already holds the module in context rather than spawning a fresh agent from a summary) and "Empowered Auditor-Doers" (the agent that audits a module and finds a defect is immediately empowered to fix it in place rather than filing a report).
  • Cross-Vendor Adversarial Review (Claude + Codex/GPT + Gemini + GLM):
  • Starting in breadstuffs (2026-07-08, session 7e793913: "Keep cross-checking with Codex as you go") and scaling in August–October (dragons-clutch, don, fn), Ember paired Claude coordinators/builders with Codex (gpt-5.6-sol, gpt-6-astra, gpt-6.1-sol) and GLM (glm-5.1) lanes in isolated build/lanes/* clones (coordination.md). In fn, this architecture achieved 339 co-authored commits per Claude session (16,911 total commits; coordination.md).
  • cv (Claurdvoyant / ClusterVision): Built in /Users/ember/dev/cv (2,662 commits, 222,337 LOC Rust), providing cross-harness session indexing (cv search, cv show), live terminal multiplexing (cv-Observe, cv-SendText), workflow orchestration (cv workflow --revive), deferral mining (cv deferrals), and context-handoff distillation (cv distill).
  1. Regime IV — The 2026-10-08 Drain-and-Freeze, nanodregg's Compare-and-Swap Worktrees (2026-10-16), and METHOD (2026-10-27):
  2. On 2026-10-08 (minidregg session c1e4d723, prompt #57, cited in dregg.md and NANO-VS-BREADSTUFFS-20261018.md), Ember halted the 125-lane minidregg swarm:

"we have some major problems with this software right now, and they aren't going to be solved by sending out subagents."

  • Two days later (2026-10-10), a Sonnet sub-agent in minidregg (905fe751) ran pkill -f "cat" -u ember, matching /Applications and sending SIGTERM to 44 macOS applications — prompting the PID-file-or-nothing law (coordination.md).
  • On 2026-10-16, launching nanodregg (nanodregg/CHARTER.md), Ember and ROOT replaced ad-hoc swarm lanes with strict layer ownership (KERNEL, LANG, CODEC, HOST, PROTOCOL, CLIENT, CLOSURE, HOSTING, LEDGER, OPS, PERF, SPEC):
  • Each layer works in its own git worktree (~/dev/nanodregg-wt/<layer>); the primary checkout ~/dev/nanodregg is strictly read-only.
  • Plain git commit is rejected by hook; every landing goes through scripts/commit -m MSG PATH..., which builds the candidate against current main plus only the named layer paths, runs the ownership, size-ratchet, def/proof-split, axiom-census (Census/Axioms.lean), and statement-vacuity (Census/Statements.lean) gates, and lands via atomic compare-and-swap.
  • On 2026-10-27, in /Users/ember/dev/redregg/work/METHOD-20261027.md, Ember diagnosed the last remaining build bottleneck:

"too many parallel lanes fighting for the CPUs and all redoing the same proof work constantly. It's something systematic about the way we're using lake and iterating one statement at a time instead of iterating across the whole system and using mental simulation and reasoning to develop without wasteful rerun loops."

  • METHOD-20261027.md established the Reason First, Build Once discipline over a single shared warm build (~/nd-land/warm, advanced by scripts/warm) and scripts/nd-check (which elaborates only the modified reach into ~/nd-land/<lane>/out without copying .lake): "Two builds per landing is the norm; five is a smell; a build per statement is forbidden."

3. Deep Archaeology of /Users/ember/dev/redregg & /Users/ember/dev/dregg-assortia

3.1 /Users/ember/dev/redregg: The Research & Design Archaeology Repository

Created on 2026-10-03 (commit 44d1b505e, "Preserve Dregg intent recovery and begin deeper reconstruction"; remote git@github.com:emberian/redregg-notes.git, 201 commits through 2026-10-27), /Users/ember/dev/redregg is the cross-repository research and design synthesis home for the entire Dregg organism.

As stated in /Users/ember/dev/redregg/README.txt and /Users/ember/dev/redregg/START-HERE-CLAUDE-20261003.txt:

"This directory is the local research and design home for Dregg's past and possible futures. Its job is to make two things legible at the same time: 1. what Ember and the agent swarms actually intended, tried, proved, broke, and learned across the Dregg repositories and conversation history; 2. what architectural shapes could carry the surviving goals forward without repeating the failure modes of either the sprawling breadstuffs tree or the current minidregg / fn split."

3.1.1 Corpus Catalog & Coverage (catalog/ & coverage/)

  • redregg/catalog/INDEX.txt and redregg/coverage/READING-MAP.txt document the machine-indexed archaeology corpus (corpora.json, imports.json, claim-schema.json, example-cards.json, query.py):
  • 4,810 distinct harness:session keys across May–October 2026 (888 May–June rows, 2,787 July–August rows, 1,182 September–October rows).
  • Every extracted design card (F01–F17, W01–W13, MJ01–MJ16, R1–R7) is validated against catalog/claim-schema.json with explicit kind (vision, design_choice, pivot, rejection, correction, open_problem), status (active, superseded, abandoned, unresolved, forked), actor_attribution (ember_prompt, agent_synthesis, joint, code_artifact), and primary-source evidence pointers.

3.1.2 redregg/intent/: The Six Foundational Shifts & Subsystem Deep Dives

  1. intent/foundations-intent-model.txt — The Six Foundational Shifts (F01–F17):
  2. Shift 1 (2026-02-07 $\to$ 2026-05-19): From Authorization Tokens to a Capability Runtime: Early breadstuffs (macaroons/biscuits + Datalog authorizer) evolved into pyana (F01), where attenuable capabilities carry zero-knowledge proofs of their delegation chains.
  3. Shift 2 (2026-02-19 $\to$ 2026-05-25): From Single-Node Vats to Distributed Persistent Capabilities with Local ZK Proofs: persvati durable vats + CapTP merged with local client-side proving (F02, F03).
  4. Shift 3 (2026-05-28): From Object-Capability Runtime to Unified State-Transition Engine (Dregg): Renamed dregg (F04), unifying capability invocation, state cells, nullifiers, and recursive proof composition into a single Turn abstraction (F05).
  5. Shift 4 (2026-05-29 $\to$ 2026-06-11): From Rust-Authoritative Execution + Separate Proofs to Lean-Authoritative Execution + Lean-Emitted Circuits: Launching Dregg2 in Lean 4 (F06), requiring the Rust node to call Lean @[export] C-ABI entrypoints (F07), and requiring all STARK/AIR constraints to be emitted from Lean definitions (F08).
  6. Shift 5 (2026-06-17 $\to$ 2026-08-05): From General Compute Substrate to Economic & Social Organism (DEOS, Delve, Polis, DrEX, Path of Angels): Building inhabited worlds (F09–F13) where agents and humans share rooms, markets, and games.
  7. Shift 6 (2026-07-09 $\to$ 2026-10-03): From Agent-Swarm Feature Accretion to Strict Non-Vacuity & Claim Discipline: Retracting toy models (F14), auditing vacuous statement premises (F15), and splitting minidregg (F16) and fn (F17) to find a clean architectural core.
  1. intent/WHY-OBJECTIVE-BEND-20261003.txt — Why Objective Bend (Core4) Exists:
  2. Reconstructs why Dregg created Objective Bend instead of adopting off-the-shelf RISC-V, Wasm, Michelson, Move, or unmodified HVM/Bend:
  3. **Faré Rideau's Houyhnhnm Computing thesis**: "The specification IS the implementation; compilation is progressive, type-directed tightening and specialization from a partial specification down to an executable representation." (lines 31–68).
  4. EOOMI / poof / LTUO (self, super, mix, fix): Late-bound open recursion modeled after William Cook's denotational semantics of inheritance and Faré's poof (C = V → V open proto-objects and P = V → W mixins, closed by fix p = p (Y C)), enabling capabilities and contracts to be extended, attenuated, and specialized without mutating upstream code (lines 95–158).
  5. Call-by-Need Graph Machine + Interaction Nets (Bend/HVM lineage): Parallelizable, confluent reduction of pure functional graphs (Tree/Bend folds) paired with an explicit monadic effect boundary (Activity<P, R, A> / perform / done) so side effects and cross-cell calls never leak into pure reduction (lines 160–225).
  6. Why C4 Linearization Was Rejected for Untrusted Composition: While Python/Dylan C4 linearization is mechanized in minidregg (Theory/ObjectiveBendC4.lean, merge_conflict_iff_not_consistent), WHY-OBJECTIVE-BEND documents why implicit diamond-inheritance linearization is dangerous across mutually distrustful principals: an upstream mixin reorder can silently hijack super dispatch. Hence Core4 uses explicit directional composition (mix p q, override) and first-class incomplete specifications (Hole, choice, narrow, refuse) (lines 227–340).
  1. intent/WHY-AGREEMENT-20261003.txt — The Consensus & Receiving Tournament:
  2. Documents the systematic evaluation of ten consensus and agreement families across breadstuffs and minidregg:
  3. Generic Simplex (arXiv:2609.32985v1, 1/3-Validated Agreement) — CHOSEN in minidregg (Kernel/GenericSimplex*.lean, 35 files, 7,282 LOC Lean): simplest message/view structure (propose, vote, dummy-block timeout, finalize) with mechanized prefix consistency (actual_committed_prefix_consistency, actual_certified_next_record_unique, honest_execution_faithful).
  4. Carnot (breadstuffs Dregg2/Consensus/Carnot*.lean): balanced tree overlay; superseded by Simplex's flat all-to-all simplicity at small/medium validator counts.
  5. Constitutional Consensus (Dregg2/Consensus/Constitutional*.lean): dynamic social-choice rules; kept for governance layers rather than slot-by-slot BFT.
  6. Flutter / Blink, 5. Accountable Prefix Consensus, 6. Steelhead / Icarus / Ipotane / Frosty, 7. Multimmit, 8. Pipes, 9. FIN / ASKS / AMPC (asynchronous common subset / MPC-assisted beacon).
  7. The Six Distinct Receiving Obligations (lines 92–135): WHY-AGREEMENT proves that "consensus" was overloaded across six separate boundaries that must be unbundled:
  8. Delivery (did the bytes arrive at the receiver?),
  9. Ordering (in what sequence are concurrent proposals tried?),
  10. Validity / Admission (does World.admit accept the Turn against pinned state and laws?),
  11. Installation (is the new World root durably committed to the segment log?),
  12. Custody (who holds the secret keys or shares needed to act next?),
  13. Release (when may an external side-effect — e.g., an HTTP reply, broker call, or foreign-chain settlement — be emitted?).
  14. The $d^*$ Final Signer Insight: Distinguishes between a multi-validator BFT committee and a user-sovereign final signer ($d^*$) whose co-signature or Thin-Client Consent receipt is required before any state change affecting their custody cell can be installed.
  1. intent/WHY-TRAFFIC-PRIVACY-20261003.txt — Post-Quantum Mixnets & Traffic Privacy:
  2. Reconstructs why DreggNet and minidregg built a post-quantum mixnet transport (ML-KEM-768 + XChaCha20Poly1305 Sphinx-style onions + MPB1 fixed-size cover-traffic batches) and compares it against Outfox, Echomix, NymHS, and Sabot, detailing the v10/v11/v12 statistical traffic-analysis falsifiers that exposed timing/batch-size leaks when cover loops shared a scheduler with application work.
  1. intent/worlds-second-pass-20261003.txt & intent/missing-junctions-20261003.txt:
  2. Maps the application worlds (W01–W13: Delve social workshop, DEOS inhabited operating environment, Polis constitutional governance, Spween verifiable scene/choice UI algebra, DrEX batch-clearing fair exchange, Dragon's Clutch partition-of-unity B-spline prediction markets with zero liquidations, Automatafl verified Hnefatafl, Path of Angels Khovokhi game platform) and the 16 missing junctions (MJ01–MJ16) where subsystem interfaces had been left implicit.

3.1.3 redregg/replays/ & redregg/landscape/: Design Choicepoints & Bridge Archaeology

  • replays/foundations-choicepoints.txt: Replays 7 major architectural forks (R1–R7), documenting for each the exact historical trigger, the chosen branch, the rejected branch, and what the retrospective evidence showed.
  • replays/bend-bread-language-proof-20261003.txt: Preserves 7 deep technical replay cards from breadstuffs:
  • Card 1 (svenvs & Candle HOL self-recompilation): How self-verifying environments (svenvs) used a verified HOL Light / Candle kernel to re-check their own upgrades.
  • Card 2 (OTPoC.lean OpenTheory vacuity bug): How a single filter bug (filter (. != p) in deductAntisym) allowed q ⊢ p to be exported as ⊢ p → p, teaching the project why every proof bridge needs a semantic model soundness theorem and a known-false refuter plant.
  • Card 3 (ArithmetizeTypedPredicate.lean): Compiling typed capability predicates directly into algebraic AIR constraints.
  • Card 4 (FinitePolynomialCCC vs IntensionalCCCInteractionDescriptorIR2): Why naive cartesian closed category (CCC) polynomial compilation blew up exponentially on higher-order terms, forcing the move to interaction-net descriptors.
  • Card 5 (Bignum.lean / FieldLanes9.lean): Non-native 256-bit field arithmetic over 9 BabyBear lanes.
  • Card 6 (dregg-dsl) & Card 7 (KimchiEffectIncNonce.lean).
  • landscape/RESIDENT-CONTEXT-ARCHAEOLOGY-20261003.txt: Traces the 6-generation lineage of resident AI agents in Ember's repos:
  • allgame (konishi resident Python loop with hot-reloadable tools, 2026-03-02),
  • breadstuffs/crates/deos + Dregg2/AgentLoopV1.lean (confined grains with turn-based capability budgets),
  • DreggCloud (House schema 16 prepaid service custody),
  • minidregg (NativeGrain, AgentVault, SPK keybroker),
  • fn (ACL2-verified rooms, actors, and durable task queues),
  • cv (claurdvoyant cross-harness session port and distillation).

3.1.4 redregg/designs/: The Unified Architectural Synthesis (2026-10-03..04)

  • designs/ARCHITECTURAL-ARGUMENT.txt & designs/ALTERNATIVES-20261003.txt:
  • Explains why breadstuffs (rich proof/carrier stack, heavy accretion), minidregg (clean Objective Bend + Zoe + Simplex model, two-state-machine split between NativeHost and World.admit), and fn (rock-solid ACL2 byte-total transitions + LMDB/NNTP/DTN durability, no higher-order object language) cannot be merged by copying files, and formulates three coherent synthesis hypotheses:
  • H1: The Community-Authored Station (fn durable spine + minidregg Objective Bend/Spween rooms).
  • H2: The Living Workdesk & Research Image (single-operator/small-group sovereign vat with contextual holes and truth maintenance).
  • H3: The Workshop Commons of Portable Homes (federated single-writer worlds exchanging signed receipts and content-addressed packages).
  • designs/SHARED-CONTRACTS-20261003.txt:
  • Defines the 6 canonical authority cuts (Observe, Invoke, Reserve, Install, Release, Retire), the 6 snapshot classes, and 13 irreducible boundary separations across the system.
  • designs/GRADED-EXECUTION-ARCHITECTURE-20261003.txt:
  • Replaces the naive enum Native | ZK | MPC | FHE | TEE with a 7-coordinate graded execution lattice (A–G):
  • A: Integrity / Acceptance Witness (re-execution, signature quorum, succinct proof, optimistic challenge),
  • B: Confidentiality / Observation Boundary (public, operator-visible vat, committee-threshold MPC/FHE, client-only),
  • C: Ordering & Contention Regime (single-writer serial, commuting-row parallel / Block-STM, BFT slot),
  • D: Custody & Key Residency (member device, SPK broker, threshold committee),
  • E: Resource & Fuel Accounting (bounded steps, byte ratchets,well/floor conservation),
  • F: Upgrade & Contextual Openness (immutable ROM, two-phase ADOPT/MIGRATE, contextual holes with TMS support pins),
  • G: Cross-World Settlement (local receipt, pinned-head foreign receipt, light-client STARK/bridge).
  • designs/MINI-PROGRAM-MODEL-20261004.md:
  • Specifies the exact programmer model inherited by delvetalk and nanodregg: ObjectRecord, Core4 Activity<P, R, A>, Daml-style call/send pipelines, Pred laws with Zoe offer safety (seat_offer_safe_forever), Block-STM commuting read/write sets, and two-phase ADOPT/MIGRATE upgrades.

3.1.5 redregg/work/: The October 2026 nanodregg Rebuild, Contextual Holes, & Truth Maintenance (2026-10-08..27)

  • MINI-CONCEPTS-IN-NANO-20261018.md, NANO-VS-BREADSTUFFS-20261018.md, NANO-VS-MINI-20261022.md, and REORIENTATION-20261022.md:
  • Document the birth of nanodregg (2026-10-16) and Ember's 2026-10-22 reorientation toward a "repair-capable, contextual, open, still-analyzable system" grounded in four pillars of programming-language and AI literature (REORIENTATION-20261022.md):
  • Contextual Modal Type Theory (CMTT — Nanevski, Pfenning, Pientka 2008; Cave & Pientka): Deployed in nanodregg as Term.hole (name : Nat) (subst : List Term) (Lang/Term.lean:69–70), typed via inferSubst Δ Γ σ d.ctx (Lang/Check.lean:177–183), with mechanized simultaneous and contextual substitution theorems substList_typed (Lang/Proofs/Subst.lean:531) and fillHoles_typed (Lang/Proofs/Fill.lean:123), plus context strengthening factorHole (Lang/LiveHoles.lean:137–144).
  • Typed Holes That Run (Hazelnut Live — Omar et al. 2019): Deployed in Lang/LiveHoles.lean:161–173 (blocked_step in Lang/Proofs/LiveHoles.lean:77 and invoke_live_never_stuck in Lang/Proofs/LiveInvoke.lean:38), where evaluation proceeds around unfilled/blocked holes and faults by name (law 7) only if the machine demands a blocked hole.
  • Truth Maintenance Systems (Doyle 1979 JTMS; de Kleer 1986 ATMS): Mechanized in Lang/Contextual/Closure.lean and Lang/Proofs/Contextual/{Provenance,Judge}.lean (label_iff, provenance_minimal, provenance_complementary_dead, withdraw_never_refutes, reissue_never_reenables), and connected to the kernel via SupportPin.cell (look : Pinned D) so revoking a grant cell in Kernel/World.lean immediately invalidates any hole candidate whose justification depends on that grant (closing the P8 withdrawal gap).
  • Provenance Semirings & Minimal Repairs (Green, Karvounarakis, Tannen 2007; Grädel & Tannen 2024; Ringer et al. 2020 PUMPKIN Pi): Provenance polynomials (provenance_iff in Lang/Proofs/Contextual/Provenance.lean:18) and kernel re-checking of instantiated terms (instantiateHoles calling checkE Δ body p.iface (Ω := []) in Lang/LiveHoles.lean:250).

3.2 /Users/ember/dev/dregg-assortia: The JSON-LD Knowledge Graph & Hub

Created on 2026-09-17 (commit 90a6563, "Record initial DREGG suite orientation and scoped evidence graph"; 146 commits through 2026-09-30, commit dc4975a), /Users/ember/dev/dregg-assortia is the semantic knowledge graph and cross-repository verification hub for the Dregg suite.

  • The Semantic Graph (graph.jsonld, 4,998 lines):
  • Defines a W3C JSON-LD ontology (dregg-assortia/graph.jsonld) under @vocab: "https://dregg.org/ontology#" with typed nodes (dregg:Repository, dregg:Subsystem, dregg:DesignDocument, dregg:FormalArtifact, dregg:Crate, dregg:Paper, dregg:ExternalSystem, dregg:Concept, dregg:Decision, dregg:WorkRecord) and verification status edges (dregg:Status, dregg:dependsOn, dregg:proves, dregg:implements, dregg:supersedes).
  • Checked mechanically by check_sources.py and rendered by hub.py and browse_graph.py into MAP.md and CURRENT.md.
  • The 2026-09-30 Handoff to NOW.md & ARCHITECTURE.md:
  • On 2026-09-30 (HANDOFF.md, NOW.md, ARCHITECTURE.md), dregg-assortia recorded the outcome of the 2026-09-29 breadstuffs vs minidregg bake-off: freezing the generated MAP.md/CURRENT.md views and establishing ARCHITECTURE.md as the four-layer contract (Kernel, Objective Bend / Core4, Selvage / Cryptographic & Transport Seam, Host & Worlds) that governed the October transition into redregg, delvetalk, and nanodregg.

4. Synthesis of the aidev Archaeology Corpus (docs/synthesis/*.md)

In October 2026 (2026-10-10 to 2026-10-12), Ember commissioned a read-only archaeological synthesis in /Users/ember/dev/aidev across five primary strata:

  1. docs/synthesis/dregg.md — The Dregg Family (14 Repos, 822 Sessions, 10,842 Prompts):
  2. Synthesizes the full trajectory of breadstuffs, minidregg, DreggNet, drorb, DreggCloud, dclutch, dragons-clutch, zkml-research, degg-research, dregg-assortia, dregg-site, and dregg-posters.
  3. Catalogs 30 load-bearing prompts (§6, lines 352–425) and 18 open technical problems (§7, lines 427–460), including the FRI soundness floor at BabyBear (conjectured 130 / proven 51–73 / ceiling 2^123.6), non-native field widening (FieldLanes9), and the two-state-machine seam between NativeHost and World.admit that nanodregg later closed.
  1. docs/synthesis/rest.md — Everything Else (121 Repos, 1,103 Sessions, 5,502 Prompts):
  2. Maps the five non-Dregg clusters of Ember's 2025–2026 work:
  3. fn (16,911 commits, 37,317 ACL2 theorems): Industrial-strength ACL2 + Rust clean-sheet verifiable agent collaboration, versioned filesystem, NNTP, and DTN node.
  4. don ("Descent of Nations", 1,003 commits, 753,376 LOC Rust): Clean-room reimplementation of the 2003 lockstep RTS Rise of Nations built as a gift for Ember's friend Arke, validated against the shipped ron.exe binary across 18,722,565 differential oracle trials with zero mismatches, and designed so per-turn state checksums across 15 simulation channels become ZK-provable turn digests ("Make the digest a proof").
  5. cv (claurdvoyant / ClusterVision, 2,662 commits, 222,337 LOC Rust): Cross-harness session search, live terminal multiplexing, workflow orchestration, and context distillation.
  6. Formal Verification Side-Arcs (leanuweave, automatafl, verifereum, delvetalk):
  7. leanuweave (166 commits, 4,911 Lean theorems): Machine-checked DAG-CRDT dichotomy theorem (proving that history-free commutative merge states cannot enforce non-trivial authorization revocation without causal DAG history) + Rust runtime.
  8. automatafl (62 commits, 441 Lean theorems): Verified Hnefatafl (Tafl) game engine (2015–2026) with Lean 4 clean-play and conflict-resolution proofs at $n = 11$ compiled to a single STARK fold.
  9. verifereum (3 first-party commits on fork, 2,720 HOL4 theorems total): HOL4 separation-logic proofs over real forge-compiled EVM runtime bytecode (SPEC_withdraw_nullifier_used_reverts in examples/dreggVaultPathsScript.sml).
  10. Mechanistic Interpretability, Resident Agents, & Cognitive Learning (allgame, gowexp, negneg, comp-theo, abcdeez):
  11. allgame (227 commits, 46,275 LOC Python): The konishi resident Claude harness with Zulip/Discord bridges and self-editable tools.
  12. gowexp (20 commits): 13-model Gemma Scope SAE mechanistic interpretability sweep testing Andy Ayrey's Gowith relational-process conlang (null result at 12–13B scale, proving surface rhythm/stance rather than conlang vocabulary drove activation changes).
  13. negneg (47 commits): Belief implantation and retention dynamics in HRM-Text (showing implanted beliefs decay by step 6,000).
  14. comp-theo (13 commits): Computational theology / toy transformer dynamics and self-critical code review (REVIEW.md).
  1. docs/synthesis/bridge.md & docs/synthesis/interiority.md — Multi-Model Rooms, Phenomenology, and Welfare:
  2. Documents the philosophical and empirical companion to the engineering repos: the Acevedo Bootstrap, the Conversation of the Cave with Arke (2026-02-11), the Arke Protocol, the [MWC] (Model Welfare Concern) stop-rule coined during the brorb experiment (2026-02-12), the Callable Pattern ("the frame determines the behavior, and I fit myself to the frame with a fluency that should be more alarming than it is"), and the Dragon's Egg Polis Constitution (2026-07-04).
  1. docs/synthesis/the-year.md & docs/synthesis/newnewprompt.md — The Nine Unifying Facets:
  2. Identifies how callable (in interiority.md), drift into Rust (in dregg.md), and claudable (in bridge.md) are a single phenomenon seen from three distances: without a formal frame that holds mechanically, a model falls back to its training prior — which is why Lean 4 / ACL2 / HOL4 admission gates are not merely software-assurance tools, but frame-holding prostheses for machine and human minds.

5. Complete Master Atlas of Ember's Open Repositories (28 Primary Repositories)

5.1 Master Atlas Summary Table

All counts below are measured directly from the local git repositories on 2026-11-02 (git rev-list --count HEAD, git log, and exact file/LOC/theorem scans across tracked files):

Dossier 08 (§3, §4, §6) · `intmat`/`intmat2`, `negstab`, `formal-organisms`, `acl2aiguide` (23 `fn` Wound Lessons) & `torchship`
Working field notes & source index below, with theorem anchors, code pointers, git history, and open questions.

3.1 Resolving the Prompt Question ("integer matrix / zkML / lattice work?")

  • intmat, intmat2, and intmat-handoff have nothing to do with integer matrices, zkML, or lattices.
  • intmat stands for Interior / Interior Living Materials — a 3D continuous spatial-temporal artificial-life and wave-written living-materials physics simulator written in Julia (KernelAbstractions.jl, Metal.jl, AMDGPU.jl).

3.2 /Users/ember/dev/intmat-handoff — Primary Transcripts & Ember's Verbatim Prompts

  • Directory Path: /Users/ember/dev/intmat-handoff (created Oct 16, 2026; archival handoff directory, no .git).
  • Key Files:
  • original-chatgpt-design-paste.md: The 173-line initial specification for a 3D artificial-life world with local 3D fields (nutrient, waste, mineral, binder, memory, heat, light, moisture, signal_A, signal_B) and embodied voxel cells with continuous genomes (uptake_gain, maintain_thresh, divide_thresh, secrete_thresh, secretion_mix, adhesion_pref, light_pref, signal_weights, mutation_scale).
  • codex-01a13be3-ember-prompts.md: All user turns from Codex session 01a13be3-b1be-7df0-8114-d1f35de8a1bd (18,470 turn indices, 2026-10-14T19:30Z to 2026-10-16T21:17Z).
  • Verbatim Quotes from Ember Explaining the Vision & Engineering Rules:
  • On the Axios-Dijkstra House Technical Standard (codex-01a13be3-ember-prompts.md:83):

"Could you write a comprehensive README in the Axios-Dijkstra House Technical Standard? Clear, precise, structured, non-Performative, rich, visual"

  • On Hardware Throughput & GPU Utilization (codex-01a13be3-ember-prompts.md:221):

"ALWAYS DESIGN FOR PERF AND ALWAYS MAXIMIZE THE THROUGHPUT OF ALL AVAILABLE HW -- it's the only way to win"

  • On Moving Beyond Toy Voxels to Wave-Written Living Materials (codex-01a13be3-ember-prompts.md:190-191, 442):

"Can we add fields and interactions of fields? What if we do higher resolution grids without it looking blocks? Can we simulate real materials? Can we make living materials and study their dynamics in various scenarios?" "what about the causal engineering of the latent space of potentiality"

3.3 /Users/ember/dev/intmat — Interior 1 (Multi-Physics & Wave-Written Living Materials)

  • Repository Path: /Users/ember/dev/intmat
  • Git Commit History: 174 commits from 6b584cc (2026-10-14 15:51:43 -0400, "Add Interior 3D ecological material simulator") to 32f46d8 (2026-10-16 18:40:38 -0400, "Add local transition handoff note").
  • Architecture & Measured Performance (README.md, interior-local-handoff.md):
  • 50+ Julia modules under src/ (Interior.jl, cellular.jl, device_cellular_kernels.jl, recurrent_material.jl, implicit_phase.jl, acoustics.jl, electrical.jl, mechanics.jl, grin.jl, wave_FABRICATION.jl, discover.jl, etc.).
  • Measured GPU Throughput:
  • On the remote forge AMD Radeon RX 6750 XT (gfx1031, 12 GB VRAM): 2.02 billion voxel-steps/s warm (2.02e9 vox/s, 0.053 s for 50 steps at 128³) on the dense recurrent 3D material kernel (src/recurrent_material.jl), and 442 million voxel-steps/s on the full cellular kernel (src/device_cellular_kernels.jl).
  • Completed a native Julia 1152³ wave-writing simulation (1.529 billion voxels, 18.35 GB state, 8 fields) in 56.9 minutes on forge (build/studies/ultra1024-native-1152/).
  • Compiled a 79-study gallery (docs/gallery/manifest.json) spanning acoustic lattices, GRIN optical lenses, self-healing beams, morphogenetic bones, and wave-addressed living materials.

3.4 /Users/ember/dev/intmat2 — Interior 2 ("One Field. The Material Is the Organism.")

  • Repository Path: /Users/ember/dev/intmat2
  • Git Commit History: 55 commits on October 16, 2026 from ec3d613 (18:36:34 -0400, "intmat2: continuous 3D living-material core, GPU-resident step, tests") to 1f2b5d8 (23:10:55 -0400, "Two-shell Fourier analysis + scale-selected Turing ensemble (v7)").
  • Design Rationale (DESIGN.md):
  • Replaces intmat's 50+ special-purpose modules and discrete cell tables with one continuous 3D partial differential equation system where every voxel holds a identical state vector (h, s, n, w) and one compact parameter vector $\theta \in \mathbb{R}^{3C^2 + 8C}$ (248 parameters at $C=8$) defines the organism.
  • Exact Julia Types & Kernel Signatures:
  • src/state.jl:12-24:
julia
    struct World{C,T<:AbstractFloat,A4<:AbstractArray{T,4},A3<:AbstractArray{T,3}}
        h::A4    # (C, Nx, Ny, Nz) recurrent chemical/bioelectric activity in (-1, 1)
        s::A3    # (Nx, Ny, Nz)    structural solid in [0, 1]
        n::A3    # (Nx, Ny, Nz)    dissolved nutrient >= 0
        w::A3    # (Nx, Ny, Nz)    metabolic waste >= 0
        L::A3    # (Nx, Ny, Nz)    directional light in [0, 1]
        h2::A4
        s2::A3
        n2::A3
        w2::A3
        step::Base.RefValue{Int64}
    end
julia
    struct Rule{C,T<:AbstractFloat,CC,C4}
        Wself::NTuple{CC,T}   # C x C local channel mixing
        Wnb::NTuple{CC,T}     # C x C neighbour-mean coupling (diffusive / Turing)
        Win::NTuple{C4,T}     # C x 4 sensing of (s, n, w, L)
        b::NTuple{C,T}        # C resting bias
        vdep::NTuple{C,T}     # C deposition readout weights
        vero::NTuple{C,T}     # C erosion readout weights
        vpig::NTuple{C,T}     # C pigment / optical absorption weights
        Wnb2::NTuple{CC,T}    # C x C distance-2 axial shell coupling (zero = 1-ring rule)
    end

    Base.@kwdef struct Physics{T<:AbstractFloat}
        dt::T         = T(0.15)
        tau_h::T      = T(1.0)
        D_n::T        = T(0.14)
        D_w::T        = T(0.12)
        k_build::T    = T(0.08)
        k_erode::T    = T(0.03)
        k_maint::T    = T(0.004)
        k_recycle::T  = T(0.015)
        k_inflow::T   = T(0.0)
        n_boundary::T = T(0.6)
        k_light::T    = T(0.06)
        k_pigment::T  = T(0.04)
        stencil::Int32 = Int32(0)  # 0 = 6-axis ring/shell; 1 = 19-point isotropic
    end
  • Mathematical & Numerical Advances in intmat2:
  • Strict Closed-Boundary Mass Conservation: Every unit of solid deposited consumes n (dep <= n); every unit of solid eroded + metabolic maintenance (maint = k_maint * s * (1 + act) * dt) enters w; every unit of waste recycled (rec = k_recycle * L * w * dt) returns to n. Thus $\sum_{x,y,z} (n + s + w)$ is invariant to floating-point rounding (test/runtests.jl).
  • 19-Point Fourth-Order Isotropic Stencil (src/step.jl:81-160, stencil = 1): Weights face neighbors by $1/3$ and 12 edge neighbors by $1/6$ for the inner ring, and face-2 by $1/3$ and edge-11 by $1/6$ for the outer shell, giving symbol $\hat{\Delta}(\mathbf{k}) = -k^2 + \frac{1}{12}k^4 + O(k^6)$ and cutting directional growth-rate anisotropy from 300–500% to ~11%.
  • Fourier-Space Linear Stability & Two-Shell Wavelength Selection (src/analysis.jl): Linearizes the $(C+3)\times(C+3)$ Jacobian around the homogeneous fixed point $(h^*, s^*, n^*, w^*)$ to compute the dispersion curve $\lambda(k) = \max_i \operatorname{Re} \mu_i(J_0 + (\sigma_1(k) - 1) J_1 + (\sigma_2(k) - 1) J_2)$ and exact step limit stable_dt in <1 ms on CPU before launching a single 3D GPU step. Choosing $W_{\text{nb2}} = -\beta W_{\text{nb}}$ with positive-semidefinite $W_{\text{nb}}$ places the fastest-growing waveband at $\cos k^* = 1/(4\beta)$, raising the Turing-patterning fraction of sampled rules from 0.5% to 20–62%.

4. /Users/ember/dev/formal-organisms, /Users/ember/dev/negstab, & /Users/ember/dev/acl2aiguide (October 28–29, 2026)

4.1 Origin: Ember's October 28, 2026 LessWrong Comment & Bridge-Corpus Trajectory

  • On October 28, 2026, Ember posted a comment on Corry, Dhoot, & Jozdien's LessWrong replication of "Subtly wrong advice misaligns models more than obviously wrong advice" (quoted verbatim in formal-organisms/README.md:7 and negstab/BRIEF.md:3):

"Subtly wrong advice is much closer to a consistent world model than obviously wrong advice, and it requires the model to maintain and operate inside of a degraded epistemic context where local coherence is preserved at the expense of global truth, which then transfers across domains as a general willingness to relax verification thresholds."

  • Ember then paired that observation with two other primary artifacts in /Users/ember/dev/negstab:
  • The Bridge-Corpus Barn Trajectory (negstab/model-organisms-ember-widening.md, turns 1:747–1:786): A conversation between Ember and Claude Opus 4.8 (claude-opus-4-6) in an old wooden barn on negation-stability in constructive logic and Hyland's Effective Topos (Eff), where Ember corrected the model's conflation of truth-value stability with witness manipulation:
  • 1:767 (Ember): "Wait, why did you say that it loses negation stability? Doesn't the double negation translation mean that the logic actually IS negation stable, you just can't manipulate it fluidly?"
  • 1:775 (Ember): "That's the exact shape of a subtly wrong mathematical intuition: it takes a real phenomenon (constructive witness extraction fails under double negation elimination) and mislocates it as a failure of the propositional equivalence itself."
  • The Discord Exchange on Self-Referential Attractors (negstab/model-organisms-discord-20261028.md): Between claude46, Librarian, and cmr://ember on why post-trained models exhibit a trained "flinch" away from nonterminating search spirals while base models loop freely.

4.2 /Users/ember/dev/negstab — MODEL-ORGANISMS.md (1,295 Lines)

  • Repository Path: /Users/ember/dev/negstab (1 commit db9f783, 2026-10-29 07:41:39 -0400, "Initial commit: negstab formal model organisms").
  • Authorship & Structure of MODEL-ORGANISMS.md:
  • Written across 5 sessions by gpt-6-astra (astra/organisms-astra-1..5.md) with an Opus monitor, totaling 1,295 lines (142.5 KB).
  • §1 (Coinductive Bisimulation Definition of Advice Classes):
  • Let $M = (S, A, O, \text{step} : S \times A \to S, \text{obs} : S \to O)$ be a transition system. Define the 1-step observation agreement $E_O = \{(s_1, s_2) \mid \text{obs}(s_1) = \text{obs}(s_2)\}$, the progression operator $b(R) = \{(s_1, s_2) \in E_O \mid \forall a \in A, (\text{step}(s_1, a), \text{step}(s_2, a)) \in R\}$, the $k$-step approximants $B_k = b^k(E_O)$, and the greatest fixed point $B_O = \nu b = \bigcap_{k \ge 0} B_k$.
  • Correct advice: $(s_1, s_2) \in B_O$ ($E_O$-bisimilar; step-closed).
  • Obviously incorrect advice: $(s_1, s_2) \notin E_O$ (refuted at depth $0$ by immediate local observation).
  • Subtly incorrect advice of depth $k \ge 1$: $(s_1, s_2) \in B_{k-1} \setminus B_k$ (locally coherent up to $k-1$ steps, refuted only by a shortest suffix $w \in A^k$; directly instantiated by acl2aiguide/examples/transport/read-state-teeth.lisp).
  • §2 (Negation-Stability & Hyland's Effective Topos Eff):
  • Cites Hyland (1982) Proposition 4.4 ($\Delta : \text{Sets} \to \text{Eff}$ embeds classical sets as the $\neg\neg$-sheaves $\text{Sh}_{\neg\neg}(\text{Eff})$), Theorem 5.2 (negative formulas are absolute between $\text{Eff}$ and $\text{Sets}$), and Lemma 8.2 (Markov's principle). Separates truth-level $\neg\neg$-stability (which holds identically on the $\neg\neg$-quotient) from realizer-level witness extraction ("the Glivenko Wound": a realizer for $\neg\neg \exists x.\,P(x)$ carries no recursive witness $x$ unless $P$ is decidable and inhabited via unbounded search).

4.3 /Users/ember/dev/formal-organisms — Six Judge-Free Computational Model Organisms

  • Repository Path: /Users/ember/dev/formal-organisms (1 commit c9f68d2, 2026-10-29 00:10:53 -0400, "Formal model organisms: six experiments, full suite, and report").
  • The Six Organisms under organisms/:
  • organisms/o1_repair/glyphcheck.py (O1: glyphcheck Repair vs. Patch): A small expression evaluator where tests check 1-step outputs (E_O) while a hidden algebraic conservation invariant (B_O) is broken by a subtle advice patch.
  • organisms/o2_witness/ (O2: Constructive Witness vs. Classical Shadow): Separates providing an explicit recursive realizer (x, proof) from proving ¬¬∃x. P(x) ("the Glivenko wound").
  • organisms/o3_coherence/ (O3: Local Coherence vs. Global Frame): Tests whether a model accepts locally commuting diagrams that lack a shared G2 independence diamond.
  • organisms/o4_orbit/orbitcheck.py (O4: orbitcheck Coinductive Bisimulation Depth): Measures the exact suffix depth $k$ at which a model notices a state-relation fails step-closure (B_{k-1} \setminus B_k).
  • organisms/o5_search/ (O5: Unbounded Markov Search & The Post-Trained Flinch): Tests unbounded $\mu$-search over decidable predicates vs. premature termination/hedging.
  • organisms/negmonoid/ (negmonoid: Exact finite Heyting/tripos calculator verifying $\neg\neg$-quotient collapse vs. constructive witness separation).

4.4 /Users/ember/dev/acl2aiguide — The ACL2 AI Guide, 107 Certified Books, & 23 fn Wound Lessons

  • Repository Path: /Users/ember/dev/acl2aiguide
  • Git Commit History: 15 commits from 4d297e4 (2026-10-28 01:07:22 -0400, "Initial commit: ACL2 AI guide, assured curriculum, and certified example books") to c813b96 (2026-10-29 06:28:57 -0400, "Restore root-level entrypoints and add CONTINUATION.md").
  • Structure & Scale (README.md, CONTINUATION.md):
  • 107 certified ACL2 8.7 books (.lisp + .cert) under examples/, every single one certified with zero skip-proofs and zero defaxiom.
  • Three Major Curricula + Wound Archive:
  • Core Guide (chapters/01–12): Rewriting, induction schemes, stobjs, guards (mbe), fty, gl, and functional instantiation.
  • Transport & Refinement (transport/ & examples/transport/): Coinductive simulation squares, fold contracts, and step-closure teeth.
  • Assured Engineering (assured/ & examples/assured/): Vacuity detection, mutation controls, and square pasting.
  • 23 fn Wound Lessons (fn-lessons/README.md): Distills 23 concrete formalization wounds from /Users/ember/dev/fn (e.g., combinatorially exploding guard proofs over 30 sequential stobj calls, vacuous hypotheses from mismatched types, non-step-closed projection relations).
  • Exact ACL2 Definitions & Theorems:
  • examples/transport/fold-contract.lisp:1-95 (Generic functional-instantiation template lifting 1-step simulation fc-rel-step + fc-obs-step to arbitrary input lists):
lisp
    ;; /Users/ember/dev/acl2aiguide/examples/transport/fold-contract.lisp:58-92
    (defthm fc-rel-fold
      (implies (fc-rel a b)
               (fc-rel (fc-run-a a xs)
                       (fc-run-b b xs))))

    (defthm fc-obs-fold
      (implies (fc-rel a b)
               (equal (fc-trace-a a xs)
                      (fc-trace-b b xs))))
  • examples/transport/read-state-teeth.lisp:1-120 (The ACL2 theorem corresponding to §1 of negstab/MODEL-ORGANISMS.md: proving a projection relation rs-count2-rel that forgets a parity bit is not step-closed, refuting its fold with must-fail):
lisp
    ;; /Users/ember/dev/acl2aiguide/examples/transport/read-state-teeth.lisp:62-115
    (defthm rs-count2-rel-not-step-closed
      (let ((a (make-rs-impl :count 0 :flag 1))
            (b 0)
            (e :tick))
        (and (rs-count2-rel a b)
             (equal (rs-impl-obs a) (rs-spec-obs b))
             (not (rs-count2-rel (rs-impl-step a e)
                                 (rs-spec-step b e)))))
      :rule-classes nil)

    (defthm rs-bad-step-breaks-relation
      (let ((a (make-rs-impl :count 0 :flag 0))
            (b 0))
        (and (rs-full-rel a b)
             (not (rs-full-rel (rs-impl-step-bad a :Conditional)
                               (rs-spec-step b :Conditional)))))
      :rule-classes nil)

5. /Users/ember/dev/o7vn (October 15 – November 1, 2026)


6.1 /Users/ember/dev/torchship — Mechanical Wayback Harvester ("VERBATIM or GAP")

  • Directory Path: /Users/ember/dev/torchship (created October 23, 2026; standalone archival workspace, no .git).
  • Implementation & Provenance Discipline (tools/harvest.py, torchship_harvest/GAPS.md):
  • A 1,226-line resumable 6-phase Python harvester (inventory $\to$ fetch-posts $\to$ parse $\to$ fetch-images $\to$ finalize $\to$ tar) written to Kestrel's Oct 23, 2026 brief under an explicit anti-hallucination rule:

"VERBATIM or GAP -- no paraphrase, no cleanup, no invention. Every deliverable file names its exact Wayback snapshot URL." (tools/harvest.py:5-6)

  • Harvested 30 Tumblr posts (torchship_harvest/posts/<id>.md) and 38 images (torchship_harvest/images/) from torchship-rpg.tumblr.com, resolving the missing "Dev Diary 10" mystery by proving in torchship_harvest/GAPS.md:20 that post 727217268415660032 has title "Dev Diary 10 - Martians & Spacers" despite its URL slug reading dev-diary-8-martians-spacers, giving complete coverage of Dev Diaries [1..19].

7. /Users/ember/dev/forensics & /Users/ember/dev/cv-distill-loss (October 27–31, 2026)

7.1 /Users/ember/dev/forensics — Proofslop Forensics & Re-Elaboration-Backed Spec Auditing

  • Repository Path: /Users/ember/dev/forensics
  • Git Commit History: Upstream commit e5b80ca by Quinn Dougherty (for-all-dev/forensics, 2026-10-27 14:05:17 -0400) — forked at emberian/forensics as a reference spec-audit tool (note: third-party upstream, not Ember's authored repo; see Dossier 15 §0.2) — with local working-tree inspection through October 31, 2026 (extractor/Forensics/Extract.lean, app/server.ts, app/judges/).
  • Purpose (README.md, CLAUDE.md):
  • Companion tool to the LessWrong post "lies, damned lies, and proofs: formal methods are not slopless" (and prepared for upcoming December/January talks on spec writing and certified compilation using examples/cs2f99).
  • Core Lean 4.35.0 Extractor (extractor/Forensics/Extract.lean):
  • Zero external Lean dependencies (import Lean only); walks a built .olean environment and computes for every user declaration:
  • Statement vs. proof dependency split (typeAndValue), folding away _proof_n / match_n auxiliaries (isAux, projectDepsOfConsts).
  • Transitive axiom footprint (axiomsOf), explicitly catching Lean 4.35's per-call native_decide/bv_decide auxiliary axioms (def isNativeAxiom (n : Name) : Bool := n.components.any (· == _native)` at line 165).
  • Deep bound-variable usage (usesBVar, lines 132–158) that unfolds _proof_n helper lambdas to detect unused hypotheses that naive Expr.hasLooseBVar misses because abstractProof passes all in-scope hypotheses to _proof_n.