| 1 | Parser/language excellence (policy language, demos, perf, Lean-first AIR generator, build) | parser deputy gen 2 (Claude) | breadstuffs | L2-L3 | policy language elab+compile proved; two public demos; generator kills hand-written AIR; make check-fast < 2 min | demo (a) table now policyc (4b0146d52); union compile 1,515 s -> 28.9 s (7f0c7efae) -> 0.83 s (f45ef0ba5, proved-equal fast paths, bytes identical); make check-fast 13.9 s measured (bar met) and pinned equal to the crate table after minimize (db96c58ca); umbrella floor-ratchet green again (86db29a82, DfaRouting repair, no baseline growth); law1 classifier repaired (7b1f3657c). direct-logic front end cut over to the Lean compiler with program_sound/complete and a two-hot refusal control, law1 row 3 -> 0 (520727fb2, 3501b00f9). demo (b) real-session projection removed (939d676cc; synthetic is the demo). counted repeats \w{64} 4.38 s -> 0.62 s (20ab02054, hash-carrying normalizer, proved equal, bytes identical). In flight: in-process dregg-lean-ffi compiler (ember decision), zkOracle chain generator (generator lane declined it; custom-leaf stays with us, next); counted-repeat determinizer (simDecide) next perf target | 2026-11-07 (quarried → O4/O5) |
| 2 | Tower256 certificate leaf prover | Tower256 gen 5 (done) | minidregg main dcf53d2d1 | L3 | native 9104 kernel byte-identical to Lean on 5 vectors; ContentCert gaps closed (automaton root; nonempty-trace DfaAccepts iff); HonestRun 934 s → 45 s, FinalPolynomialCheck 12.3 GiB → 3.4 GB; umbrella green | next: ContentCert over a served automata AIR leaf; GPU prover via init 17 | 2026-11-06 (quarried → O2 UDR) |
| 3 | Dark learning (MPC custody, HE helpers, collaborative ZK proof, handover) | dark-learning gen 6 (deploy, prep, verifier, train lanes) | zkml-research (native/) | L2 | three custodians on forge/bb2/hbox over pinned TLS with distributed keys; real sealed MASCOT/LowGear tapes made per box; d = 4096 GEN6 step 94 s wired / 791 s via hbox; proof 2^-101/step (F_p^2), 43-47 MB, pipelined; handover rotation + self-handover with erasure all green; release M = 52 fits 12 GB/seat; budget 2000 steps 2^-15; MNIST plaintext 86.6% @2000 | run of record (N = 8, handover after 6, release) scheduled to end ~Nov 10 20:00 UTC; Python replay at d = 4096; 2,000-step run needs a wired third box (ember) | 2026-11-07 (bounded record kept; superseded → O2/O6 Tulva+Sable) |
| 4 | Governance device + attestation (seL4 chokepoint, NVIDIA/Blackwell, Chutes) | governance successor | degg-research, breadstuffs | L3 | IRQ exposure fixed; virtio TX; runtime halt falsifier; Blackwell verified live; narrator gate on | CAKEML-ROUTE.md, Blackwell | 2026-11-06 (quarried → O3/O6) |
| 5 | HOL4-Lean bridge (panSem export, row contracts, install) | bridge follow-on (done) | DreggNet, breadstuffs, degg-research | L3 | 64 contracts; Mac install 60 s; ladder on exported evaluate | ember: ledger policy A/B; vacuous demos | 2026-11-06 (vacuous wrappers retired; quarried → O1) |
| 6 | Mina/Pickles emitter rescue | rescue lane | breadstuffs | L0 | every module < 4 GB / 60 s or retired with census | census + diagnosis | opened 2026-11-06 (superseded → O2 DualPCS) |
| 7 | Iteration speed | iteration lane (today's scope done) | minidregg, breadstuffs, redregg | L2 | 8 modules fixed (ObjectLaw 16.5→2.7 GB, PiDeclarationDeployed 14.6→6.1, RotatedKernel 14.2→3.7, HostTeethProofs 8.3→1.45…); olean snapshots portable across machines/OS; make check-fast/check-full/boxes in both repos; policy + pin rules documented | ember: switch shared .lake/packages; bb2 flock shim; breadstuffs-codex Mathlib 7 GB; next: import Mathlib trimming (31 modules at 6.4 GB floor), TokenCensus, compiled guard exes | 2026-11-06 (converged → M0 obb cone + obb broker) |
| 8 | Audit theorem over a bit-exact recompute of gpt-oss-20b (FlashProve kept as a comparison point) | init-8 lane | zkml-research, minidregg, breadstuffs | L3 (G1-G5 reached; 2026-11-07) | gptoss-int v0 bit-identical on Mac, x86, T4, TPU v5e (layer 0); 24 layers agree with float64 at 301/302; auditor recomputes fail-closed with compiled-Lean verdicts (9 live receipts); steered transcripts accepted by a logit-tolerance checker and refused by exact recompute, on real weights | 24 layers in JAX on accelerator; a bf16 GPU server as the honest float backend | 2026-11-07 docs/EXACT-RECOMPUTE.md; Lean c9717142f (quarried → O6 Ozaki+Sable) |
| 9 | Attested transport + receipts (MPC-TLS call-tree certificates) | lane 9 (done) | breadstuffs | L3 | real tree: glm-5.3-flash → self-signed tool server → example.com, 4 MPC-TLS hops, 3 dataflow edges over authenticated bytes, key hidden; stranger re-verify from fresh GitHub clone ACCEPT in 57 s; 10 tamper classes refused | ember: keep bb2 notary/tool server running; publish call-tree page; independent notary; next: typed edges by call id, freshness challenge, renter-anchored root | 2026-11-06 99204bcb9 (quarried → O5 ZkStripe) |
| 10 | Third-party verification of a real Claude Code session (R0-R3) | NEW | breadstuffs, cv | L2 | a stranger verifies a cv-exported session through R3 with no host trust; microsite | GROUNDING.md | opened 2026-11-06 (model-head retired; executor-faithful → M0/O3) |
| 11 | Streaming certificates (prefix receipts over a live log, Büchi/ω) | streaming lane | breadstuffs | L3 (cross-model audit owed) | ω-automaton with three-valued verdict, window_sound/windows_compose, Büchi visit count (liveness never certified true from a prefix); stream-check follower; live runs to 27,775 lines; a real git add -A caught | ember: Codex re-auth; publish so-far page; publish window files | 2026-11-06 1f2569bd7 (quarried → O5/O6) |
| 12 | Tree automata over JSON/CBOR | tree lane (done) | breadstuffs | L2→L3 local | deterministic CBOR (not JCS: JCS aliases 2^53+1), canon_injective, schema_sound, path/corpus certificates via extracted tree-check; over 3,382 real tool calls: 13 mentions of pkill -f, 0 at the byte pattern; 18 tampers refused | ember: publish certificates+corpus; push microsite 59a5fdb; salted leaf digests; move onto Tower256 leaf | 2026-11-06 (quarried → O5 Det-CBOR TreeAut) |
| 13 | Eval-provenance / model-identity receipt | init-13 lane | zkml-research, dregg-microsites | L3 (strict mode waits on H100 recompute of E3) | receipt over real Llama 4 Maverick: own §11 conversion reproduces FlashProve wsums digest over 232 GB; per-expert cSHAKE roots; Lean identity_sound/hollow_refused; which-model page live | ember: who signs a real eval (H100 recompute of FlashProve dropped by ember) | 2026-11-06 45f11ea9 (quarried → O6 Passport CWT) |
| 14 | Hardware: tensor-core exactness model (F3) + holography accumulation (F4) | init-14 lane (done to L1) | zkml-research, minidregg lane/mma-exact, degg-research, redregg | L1 | F3 counts model + checked Lean exactness per accumulator + GPU protocol ready; F4 verdict "adapt" (no cheaper hash-based accumulation; keep commitment-free structure accumulation for later) | ember: one hour on a B300/B200 (+H100) | 2026-11-06 b21d99d6 (converged → O2 PikkuFold + O6/O7 Ozaki/RDNA2) |
| 15 | ArkLib: Selvage keystone derived from ArkLib's merged state-restoration theorem (finite alphabets); review draft finds no adaptive-adversary test in ArkLib's SR suite | init-15 lane (done) | minidregg lane/arklib-sr-bridge, zkml-research | L1 | bridge green (4,082 jobs, no sorry); coupling lemma proved; two codex checks; retirement list (recommend keep as cross-check) | ember: infinite-alphabet closure; post review; adaptive-test PR to ArkLib; blueprint (c) | 2026-11-06 f00e39c75 (quarried → O2/O6 AuditSampling) |
| 16 | Unlearning / forgetting certificates (C18-C20) | init-16 lane | zkml-research, minidregg lane/unlearning-certs | L2 | exact deletion for C18/C19 on UCI optdigits, 22/22 attacks refused, SGD as a bound, custody on Pedersen commitments, PQ commitment | umbrella + merge; ember: transcript leak, custody release, PQ commitment | 2026-11-06 2c7013d5 (quarried → O6 ClusterGov) |
| 17 | Accelerator readiness: dark-learning BFV kernels, Tower256 GPU prover, tensor-core probe; emulate locally, tune on Kaggle T4/TPU, one-hour runbooks for paid boxes | accel-readiness lane | zkml-research, degg-research | L0 | each workload passes a Kaggle dry run; RUNBOOK with box spec and cost; no paid hardware until ready | GROUNDING.md | opened 2026-11-06 (converged → O2/O7 hbox RX 6750 XT BOLT-FHE/RDNA2) |
| 18 | TPU first-principles performance: bound per kernel, static HLO/Pallas checks on the Mac, Kaggle profiles as %-of-bound; MXFP4 recompute, BFV NTT, sumcheck-as-matmul at ≥95% | tpu-perf lane | zkml-research | L0→L1 (deviceless TPU compile works on hbox: real libtpu compiler, v5e/v6e/v5p/v4) | rerunnable bound + static check + measured %-of-bound per kernel; ≥95% or named gap; project TPU style guide | bound calculator + HLO inspector on hbox; explain int8 dot copy fusions | 2026-11-06 b14717f6 (preserved as count-model note per Astra §4) |
| 19 | Veil 2.0 vs our temporal/program logics; Ariadne state continuity for the chokepoint device | veil lane | zkml-research notes/veil, veil-pilot, degg-research chokepoint-step | L2 | census; Veil read; two-protocol pilot; Ariadne modelled, model-checked, kernel-proved, refinement into the Veil model, BLAKE3 byte export | done: device kernel-proved (248f1cf); Veil model 1f83fe8; refinement refines (device step is a forward simulation of the Veil model, one-increment control does not simulate) veil-pilot 10fb360 (local); byte step with BLAKE3, stepBytes_complete/stepBytes_sound, @[export dregg_chokepoint_ariadne_step] ed3ee50 (green on 4.33.1 and 4.35.0). Next: HOL4 bridge consumes the export; hex layer still tested, not proved | 2026-11-06 (quarried → O1 Loom + O6 AriadneStep) |