Show HN: A custom language that compiles and runs compute kernels on AMD GPUs

Drake Stapleton published Omega, a systems language and compute platform that compiles and runs compute kernels on AMD GPUs, with GPU execution handled by the separate CUDA++ (aien-dev/CUDApp) stack under architecture decision ARCH-0042. Omega's seeded `.om` compiler, native host and device/mapping words, OM-2 profile and WebAssembly back end are on `main`, but the migration is unfinished: substantial C implementation and build/test tooling remain on the default branch, and building the aien-sovereign-core server from CUDA++ pinned in `cudapp.lock` is pending in pull request aien-sovereign-core-mv#784. CUDA++'s drift gate (`tools/extract/drift.sh`) fails when a GPU file changes in the Omega repository after cutover baseline `e5b32567` without matching bytes in CUDA++, and GitHub Linguist does not yet recognize Omega or `.om` files.

Papers: https://doi.org/10.5281/zenodo.23614483 Omega https://doi.org/10.5281/zenodo.23547810 , The Token Is the Instruction https://doi.org/10.5281/zenodo.23644702 , A Call to Humanity https://doi.org/10.5281/zenodo.23614483 · Author: Drake Stapleton GPU stack owned by CUDA++ aien-dev/CUDApp https://github.com/aien-dev/CUDApp ARCH-0042 https://github.com/aien-dev/aien-architecture-mv/blob/main/docs/adr/0042-omega-cudapp-bridge-ownership.md : Omega owns computation, CUDApp owns GPU execution . CUDA++ is the single source of truth for the GB10 GPU stack: src/omega gpu , src/omega blackwell , src/omega bw , src/omega sm121 , the GPU host tests tests/gpu , tests/bw , tests/test omega gpu , mk/omega-gpu- , mk/sm121dis.mk and tools/sm121dis/ full list: .github/CODEOWNERS . Make new GPU changes there. The copies here stay in this repository; removing them needs a separate decision by Drake ARCH-0042, decision 7 . CUDA++'s drift gate tools/extract/drift.sh fails when one of these files changes here after the cutover baseline e5b32567 without the same bytes in CUDA++. Building the server aien-sovereign-core from CUDA++ pinned in cudapp.lock is pending aien-sovereign-core-mv 784 https://github.com/aien-dev/aien-sovereign-core-mv/pull/784 . Omega is a systems language and compute platform being implemented in handwritten, seeded .om . Its goal is one inspectable path from a small seed through compiler, runtime and applications to machine execution. Pure computation and authorized effects have distinct contracts; semantic identities identify declared computations, while validation and receipts record what was actually checked. Migration is not finished. This repository's default branch still contains substantial C implementation and build/test tooling. The seeded .om compiler, its native host and device/mapping words, the OM-2 profile and the WebAssembly back end are on main ; an existing C implementation is an outstanding replacement task, not the final language of Omega. - The Omega programme: seed to model https://github.com/aien-dev/omega-mv/issues/599 - Omega 1.0 roadmap: readiness audit, milestones and acceptance gates https://github.com/drakestapleton/omega-mv/blob/main/docs/ROADMAP-1.0.md - One small Omega program, and the identity boundary https://github.com/drakestapleton/omega-mv/blob/main/docs/TEACH OMEGA.md : a bounded lesson for issue 619. It uses a historical teaching pin, not the current compiler, and it does not claim the legacy identity machinery covers seeded programs. - Remaining C source/header inventory https://github.com/drakestapleton/omega-mv/blob/main/docs/migration/C-TO-OM-INVENTORY.md - Experimental .om deterministic-stream successor https://github.com/drakestapleton/omega-mv/blob/main/seed/physics0/README.md : source/host checks passed; native compilation and execution NOT RUN. - Neuralese supplement https://github.com/drakestapleton/omega-mv/blob/main/neuralese/README.md : PROPOSED model-bound representation design, with a bounded native .om envelope/codebook candidate and ordinary checked-source boundary. No learned reasoning or model experiment is claimed. - Alpha model/platform qualification https://github.com/aien-dev/aien-architecture-mv/issues/305 GitHub's language bar measures recognized source bytes on the default branch; it is not the project's language specification or a migration-completion verdict. GitHub Linguist does not currently recognize Omega/ .om , and the active C files here have not yet been replaced and removed. We do not label handwritten C as generated/vendor code, rename it to .om , or assign an unrelated language to change that statistic. See the inventory above for the actual remaining work. Native GitHub recognition requires support in Linguist https://github.com/github-linguist/linguist/blob/main/CONTRIBUTING.md adding-a-language . php SILICON - ATLAS - PHYSICS/FORGE - OMEGA - AIEN full picture: aien-architecture Omega was decided by Drake on 2026-09-29 to be both the reaction runtime and the compiler for Omega Systems Core decision record https://github.com/drakestapleton/omega-mv/blob/main/docs/adr/OMEGA-SYSTEMS-CORE-0000.md , frozen . This retires the earlier "not a compiler" wording. Research-grade, with a staged Alpha qualification programme in progress. An Alpha plan or an open PR is not a qualified release. Plainly: - Seeded .om compiler on main from PR 696 : one compiler, seed/out/omc , written in .om and rebuilt from the 408-byte loader by sh seed/omega-from-seed.sh seed0 - hex1 - asm2 - om4 - omc . Its one supported profile is AArch64 Linux, static ELF output, with the whole language in docs/one layer/OMEGESE.md : native host words ids 100..165 and the checked device/mapping words ids 166..177 . Host qualification is tests/gates/run.sh seed the OMEGA-SEED gate: stage fixpoint, seed/host/check.sh and the IR comparison against the C oracles, no gate skipped . What it does not prove: the fixpoint is a same-lineage rebuild check, not diverse double compilation; the reference interpreter on seed/omega-ref predates this compiler source it refuses a call to a function defined later, first at fexp , seed/omc/omc.om line 935 . Diverse double compilation is the separate gate seed/qualification/ddc650/ddc.om PROPOSED, host EXERCISED, not in CI, not QUALIFIED; omega 650 https://github.com/aien-dev/omega-mv/issues/650 : the seed om4 route and a GCC-built copy of the synced spec interpreter tools/omega ref/omega ref.c on the unmerged seed/omega-ref-sync at 190130b7, pinned by SHA-256 each rebuild omc.om , and at 62d1d0d6 both reproduced seed/out/omc 62775d2a byte for byte while every negative control failed 2026-10-30, one AArch64 host, same project lineage; see its README https://github.com/drakestapleton/omega-mv/blob/main/seed/qualification/ddc650/README.md . The older tools/omega ref/gate ddc.sh PASS omc 734f0a39 , 2026-10-24 exists only on that unmerged branch and does not describe the current compiler. GPU launch and model execution through these device words are not qualified. The language has bounded native recursion; a restricted importer does not describe the whole language. See conformance and evidence limits https://github.com/aien-dev/omega-mv/issues/624 . - Candidate x86-64 Linux backend FR-9 : an initial target x86linux line selects x86-64 static ELF output for whole programs; no selector preserves AArch64 output. The backend is implemented in seed/omc/omc.om , not a C translation. Design, ABI, provenance and qualification https://github.com/drakestapleton/omega-mv/blob/main/docs/X86-LINUX.md distinguish emulated execution from native x86 execution and the existing AArch64 seed chain. Bare x86 units use a private frame ABI, not the SysV C ABI; foreign callers still use omega-bridge. Merged as 923; status DRAFT. x86 execution so far is EMULATED QEMU only; native x86 execution and a native x86 seed0 chain are NOT RUN. The last complete differential, at source 62d1d0d6 2026-10-30, seed/qualification/x86-linux/20261030-main-62d1d0d6.json , reported 15182 PASS, 0 FAIL, 2 NOT RUN and 222 EXCLUDED; the two NOT RUN cases are invalid-fd ioctl fixtures for which QEMU returns ENOTTY where native Linux returns EBADF. Its aggregate verdict remains FAIL because NOT RUN is not PASS. This is not a hardware-qualified release. - Legacy OSC/C implementation on main : OSC-0/1/2 and OSC-3 material describes the earlier implementation and unfinished migration, not the forward source-language path. The old Milestone 6 "self-host" result was a fixed-output self-copy check, not compilation historical note https://github.com/drakestapleton/omega-mv/blob/main/docs/osc/OSC-1-SELF-HOST-STATEMENT.md ; do not confuse it with the later seeded compiler work. - Reaction runtime: the R1 to R16 qualification ladder passed on builds that predate the composition modules omega 126 , a historical result. Requalification at 3108fc2 recorded R15 FAIL 14 of 16 gates: G1, G15 and R16 FAIL evidence/REQUAL-3108fc2/ ; the R16 receipt 22d7a79a at 850fc54 is invalid as a current claim because its script was later found defective; none of R1 to R16 is requalified on the current candidate, and the ladder rerun is in progress. The R16 loop-site inventory was reconciled on the candidate tree 502 sites, 0 unclassified, 0 stale, PR 279 ; that is the inventory gate only, not R16. The current living build is IMPLEMENTED / NOT QUALIFIED until re-qualified. Composition, Capability Graph, Skill Router, J-Space and Cortex run on the host; the Fabric interface is a host-only loopback with no network; none of this is on silicon as a qualified system. - Numerics E1 : closed 2026-10-02 on the GB10 chip campaign Gate 5 26/26 with 47 parity lines, reductions SUM/MAX/MIN/MEAN, nine transcendental ops over all 2^32 inputs each, 148 load/store kernels , consolidated receipt evidence/E1-CLOSURE/e7851c69d34ac777a9436af16d0bdd5d69264c8528c62d562b600f5d89153061.json . Natural-base EXP and LOG, the 2026-10-02 exclusion, now have GB10 kernels: on omega cb6d201 all eleven E1 transcendental ops the nine plus EXP and LOG passed on the chip over all 2^32 inputs each with 0 mismatches evidence/E1-TRANSC-GB10/b2f4afe9… , ae3a4123… , d846fbe3… , and the M20 tensor GB10 tier passed 99 of 99 cases on 0919efc3 evidence/M20-TENSOR-GB10/a00b4178… ; see E1 GAP TABLE https://github.com/drakestapleton/omega-mv/blob/main/docs/numeric/E1 GAP TABLE.md , "Update, 2026-10-27". - Decode attention OWN-30B : one cache holds up to 131072 positions 128k context, OMEGA GPU DATTN MAX POS . On the GB10 chip, contexts 65536, 100000 and 131072 match the double reference exactly evidence/dattn-maxpos-128k-20261018/ ; at those lengths the kernel streams K/V at about 57 GB/s, so long contexts are slower than the weights alone allow. - Qwen3.8-27B kernels GB10, 2026-10-18 : each one is checked on the chip against a double-precision reference with deliberately broken variants that must fail, and none is joined into a serving engine yet. FP8 matrix-vector and the tensor-core prefill GEMM now take k up to 32768, which covers the 27B down projection at k = 17408 evidence/fp8-k32768-20261018/ , omega 541 . The Gated DeltaNet decode step the recurrent layer used by 48 of the 64 layers runs as one launch per layer and matches the reference to about 2e-7; it takes 125.7 us per layer, which moves its 6.29 MB of state at 50.1 GB/s against 273 GB/s nominal evidence/gdn-decode-20261018/ , omega 542 . The DeltaNet prefill reads a whole prompt in three launches per layer and hands its state to the decode step, which carries on exactly; at 2048 tokens it takes 3.80 us per token per layer, and the recurrent scan in it does about 2.4 trillion operations a second, against a chip peak for that kind of math that has not been measured yet evidence/gdn-prefill-20261018/ , omega 543 . The full-attention decode step for the other 16 layers 24 query heads over 4 key-value heads, 256 numbers per head, rotation on the first 64 only, done by reordering the numbers once at load runs as one launch per layer, matches the reference at each tested context 1, 7, 256, 301, 1024, 4096, 8192, 32768 and 131072 tokens with all five broken variants caught, and at 32768 tokens of context takes 2.10 ms per layer, streaming its 134 MB of cached keys and values at 63.8 GB/s against 273 GB/s nominal evidence/QWEN38-DATTN256-8bb4494e-20261018/ , omega 544 . The prefill attention for those 16 layers reads a 2048-token chunk of a prompt in two launches per layer and leaves cache rows bit-equal to the ones the decode step would have written; it matches the reference at every tested point up to 129024 + 2048 tokens with all four broken variants caught, and a 2048-token chunk takes 10.0 ms per layer at the start of a prompt and 591 ms at the end of a 128k one, streaming cached keys and values at 230.6 GB/s against 273 GB/s nominal evidence/pattn256-chip-20261018T1517Z/ , omega 545 . Getting it to run found a bug every kernel shared: the encoder wrote the integer add, one compare and the exit instruction with a fixed timing word, so the waits the code generator scheduled on them never reached the chip. The encoder now keeps them, make test-bw-encode-control checks every instruction type, and 18 of 19 kernel families pass their chip tests with the fix the ternary family fails only its speed gates, which the old encoder fails the same way on the same load; evidence/encoder-fix-sweep-20261018T1512Z/ . - Expert slots Expert Ribbon, omega 533 : src/omega gpu fp8 slots.h keeps a fixed pool of FP8 expert slots that can be refilled while a model runs. A slot is hidden while it is rewritten, gets a new version only after the copy is complete, and cannot be rewritten while any launch or recorded command list still reads it. Shown on the GB10 at the slot commits a8ae3a1f and d8185686 on their 2026-10-18 bases tools/run gpu fp8 slots chip.sh , receipts in evidence/RIBBON-SLOTS- ; not re-run on the current base : a slot gives bit-identical results to the same expert held resident, and the negative control pins ignored is caught. The SSD side and the model wiring live in aien-sovereign-core. - Hardware: results on the DGX Spark GB10 are tied to the exact commit in each receipt. Host and emulator passes are never reported as hardware qualification. The live status is owned by aien-architecture-mv https://github.com/aien-dev/aien-architecture-mv : CURRENT EXECUTION PLAN.md https://github.com/aien-dev/aien-architecture-mv/blob/main/CURRENT EXECUTION PLAN.md and doctrine/ROADMAP.md https://github.com/aien-dev/aien-architecture-mv/blob/main/doctrine/ROADMAP.md . This file does not restate numbers that drift. Historical milestone receipts for example Milestone 4, OMEGA SEMANTICS , 12 of 12 gates stay in evidence/ unchanged as history. - FP8 expert launch in one wave 2026-10-24, L310, measured on the real chip : the grouped expert GEMV used 6 CTAs per group, a value tuned on 8 groups; a 9-group model the 35B, Qwen3.6-35B-A3B, a mixture-of-experts model with about 3B parameters active per token: 8 routed experts + the shared one therefore launched 54 CTAs on 48 SMs, two waves, in each of its 80 expert launches per token. AIEN OMEGA FP8 GROUPED WAVE=1 sizes the launch to one wave min 6, 48 / groups ; 9 groups - 5 and the 35B 1-row step measured 20.498 - 19.347 ms, expert stream 160.5 - 193.3 GB/s N=3, chip time alone, not end to end through the server; receipts in aien-sovereign-core docs/speed/Q36 CEILING.md section 11, which is not on that repository's main yet: commit 65444d38 https://github.com/aien-dev/aien-sovereign-core-mv/blob/65444d3810ddf691e3bffb3276ec4a39eef838fe/docs/speed/Q36 CEILING.md . These GPU kernels are C code in this repository, not Omega-language code bar C1 in docs/OMEGA.md . Off by default until the merge train has judged it; AIEN OMEGA FP8 GROUPED CTAS=n forces an exact value. gpu fp8 api test --sim checks the rule table without the chip; the server prints the value it runs with at load. - SMART is not TRUSTED, INTELLIGENCE is not AUTHORITY. Trust comes from verified invariants, never from capability. - Meaning is permanent; representation is disposable. A program's SEMANTIC ID SHA-256 of the canonical encoding names its canonical body, so respellings of one body whitespace, parameter name, operand order share it, but x + 1 + 1 and x + 2 do not. The separate FUNCTION ID names the function: inside the polynomial fragment + - and exact and / or identities modulo 2^w, w = 1..64 two programs get the same FUNCTION ID exactly when they compute the same function; outside it the id falls back to the body and says so. Equality of arbitrary programs is undecidable, so no id is extensional for every program. Nodes, typed relations and receipt-backed equivalences form the graph of meaning spec/program-identity.md https://github.com/drakestapleton/omega-mv/blob/main/spec/program-identity.md sections 9-10, om/meaning/ ; host-exercised, not yet wired into the C library or Visor . - Pure computation is separate from effects , and effects need capability-bounded, receipt-producing transactions. Ownership rule ARCH-0042 https://github.com/aien-dev/aien-architecture-mv/blob/main/docs/adr/0042-omega-cudapp-bridge-ownership.md , Accepted by Drake 2026-10-30, merged as aien-architecture-mv 342 https://github.com/aien-dev/aien-architecture-mv/pull/342 : Omega owns computation, CUDA++ owns GPU execution, Omega Bridge owns compatibility. - This repository owns the language, its semantics and IR, the seeded compiler seed/out/omc and the compiler pin other repositories consume, and the declared contract of what a kernel computes. - CUDA++ https://github.com/aien-dev/CUDApp owns GPU execution: the sm 121 encoder and GPU code generators, the cubin writer, the disassembler, the driver client, submission and the GPU runtime. - Omega Bridge https://github.com/aien-dev/omega-bridge owns compatibility: C and Python emitted from a checked Omega subset, native bindings so foreign callers call compiled Omega pinned here in om/omega-bridge.lock , framework attachments and GPU artifact negotiation. It does not compile Omega for other repositories and does not run kernels. Not yet true on main: the GPU C under src/omega blackwell , src/omega gpu and src/omega sm121 disasm.c is a diverged copy of CUDA++ code that is still developed here, and the driver client comes from physics-mv. Existing code stays where it is Drake, 2026-10-27 ; new GPU-execution work goes to CUDA++ or a repository connected to it. The copies are recorded in CUDApp 106 https://github.com/aien-dev/CUDApp/issues/106 , CUDApp 107 https://github.com/aien-dev/CUDApp/issues/107 , CUDApp 108 https://github.com/aien-dev/CUDApp/issues/108 and CUDApp 10 https://github.com/aien-dev/CUDApp/issues/10 . The bridge's own seed bootstrap and front end are 870 https://github.com/aien-dev/omega-mv/issues/870 and 871 https://github.com/aien-dev/omega-mv/issues/871 . aienux https://github.com/aien-dev/aienux holds Aienux, the kernel of AienOS, which Omega is meant to run on. physics-mv https://github.com/aien-dev/physics-mv provides FORGE machine realization and the GB10 native path pinned in physics.lock . aien-architecture-mv https://github.com/aien-dev/aien-architecture-mv owns status and sequencing. ARGUS-0/1 a defensive plane is implemented in aienux native/argus , pinned by argus.lock ; Omega only carries the opt-in RX ARGUS hooks src/runtime/rx argus.c and receipts in evidence/ARGUS/ . Physics Zero exists here as a host benchmark apparatus and learner src/physics0/ , tests/physics0/ , mk/physics0-learner.mk , docs/physics0/ , receipts in evidence/physics0/ , with its specifications in the physics repository. DIRAC-0 is a specification in aien-architecture-mv plus an oracle source in research/dirac-oracle/ that has never been built or run. The Evolution Arena is a plan only. Child repositories pin an omega-mv commit and the sha256 of the omc it births: omega-gpu-driver https://github.com/aien-dev/omega-gpu-driver is the GB10 GPU driver in seeded Omega W0 bind/read/release over VFIO via vfioctl / vfiomap ; exercised in QEMU, not hardware-qualified . Language rule: .om is the only forward first-party implementation language. Implement and qualify each replacement, switch its actual consumers, then remove the superseded first-party source and obsolete build/runtime dependencies from the active tree. This includes compiler/runtime code, services, tools, automation and project-owned tests; C/Rust/shell remaining today are unfinished migration, not permanent exemptions. Preserve Git history and immutable evidence. External protocol formats, generated machine code, device firmware and independently sourced qualification oracles have separately declared roles; they must not conceal a foreign implementation of Omega. See the replacement/deletion contract https://github.com/aien-dev/aien-architecture-mv/issues/300 . Preserve the native GPU work and its explicitly recorded OS/driver dependencies; do not infer hardware or OS independence from removal of a userspace library. No systemd. Offline, in-house builds. Never overwrite failed experiments or old receipts. Verdict words: PASS, FAIL, NOT RUN, BLOCKED HARDWARE, BLOCKED OPERATOR, MISSING IMPLEMENTATION. These are C/legacy compatibility checks , not commands for the seeded .om compiler. The seeded compiler is rebuilt and qualified with sh seed/omega-from-seed.sh and tests/gates/run.sh seed see seed/README.md ; do not mix compiler binaries or relabel old receipts. Needs gcc and make on an AArch64 Linux host: CI runs the host suites on ubuntu-24.04-arm , and test-realize compiles with -march=armv8.6-a+dotprod+i8mm+sve . Host suites need no GPU and no physics checkout: make PHYSICS DIR=/nonexistent PHYSICS LOCK CHECK=0 test-compiler-quick compiler, quick suite make PHYSICS DIR=/nonexistent PHYSICS LOCK CHECK=0 test-numeric-cpu E1 CPU numerics make PHYSICS DIR=/nonexistent PHYSICS LOCK CHECK=0 test-realize The full omegatool build make links the physics repository at the commit in physics.lock . Some older milestone gate scripts rewrite their receipts in place, so do not run them on a clean checkout you mean to keep; prefer the make test- targets above. Some GPU suites build against physics but still open no device: make test-gpu-decode-graph checks the decode graph's glue kernels and, on a fake platform, that a failed or abandoned decode token always gives the GPU session lock back shown on the fake only; the real-chip decode run is tools/run gpu decode graph chip.sh . Targets that touch the GB10 need the DGX Spark, run one at a time, and are never killed mid-run. The M19 clean-clone gate rebuilds the tree in a temporary directory and runs the milestone gates there; when that nested build or nested run fails, it now prints the tail of its own log to stderr, so a failure names its cause instead of reporting a bare FAIL. CI definitions in .github/workflows/ list the suites that run on every push. make test-bw-hazard host only, links libomega gpu.a , so it needs the physics checkout like make reads the encoded bytes of every GPU kernel the generators build and checks the scoreboard waits: a register written by a load must be waited on before it is read, and a register a queued load or store still reads must be waited on before it is overwritten omega-mv 554 . The simulators do not model these waits, so this is the host check that would have caught the omega-mv 545 chip fault. Known findings on older kernels are listed, with how each was confirmed, in tests/bw hazard test.c . Host checks on the shared Spark judge, whose 1-minute load average is rarely below 2, keep every check they had and never turn load into a pass. tests/runtime/r11 own hold test.sh raises the R11 load limit for its own runs R11 LOAD LIMIT , which can only raise it and judges the quiet-flag hold decision the program prints; the living run's speed checks are judged there only when the load average is at most 2. The speed checks the action-graph 2x speedup in tests/runtime/rx action graph.c , workflow fusion take their median over the repetitions the machine did not disturb run-queue delay from schedstat, tests/runtime/rx quiet reps.h ; when too few are undisturbed the check reports NOT RUN with reason cpu busy , and tests/gates/run.sh reads that refusal from the receipt, runs the rest of the step and ends it NOT RUN. Omega's serve cost is the CPU time of the serving thread cost ns in src/runtime/rx omega.c , not wall-clock time, and the R15 parity rig gives both sides a fixed cost table r15 use cost model . src/ existing core and C compiler awaiting replacement src/compiler/ , src/runtime/ reaction runtime, src/fabric/ host-only fabric interface, src/physics0/ Physics Zero apparatus, spec/ and docs/ specifications, tests/ gate suites, evidence/ content-addressed receipts, tools/ harnesses, mk/ make fragments included by the Makefile . Also bench/prime race/ Prime Drag Race implementation , calibration/ Turing calibration experiments, preregistration and blinding protocol , research/ experiments and prototypes and doctrine/ the Milestone 4 doctrine . Development follows Human-Directed Agentic Engineering HDAE : humans define goals, architecture, constraints, acceptance criteria, and stopping points; agents implement, execute checks, diagnose failures, and revise within those boundaries. Human review retains responsibility for claims, acceptance, and release decisions. For the documented method and a systems-software case study, see Gregory M. Kurtzer, Ginux: Building an Independent Kernel That Runs Linux Userspace through Human-Directed Agentic Engineering 2026 https://doi.org/10.5281/zenodo.23682548 . This is a method citation, not a change to authorship or the qualification of earlier experiments. Open a pull request with the exact command you ran and its output. Receipts are content-addressed, name the exact commit and refuse a dirty tree. Do not edit existing evidence in place. External review is expected before merge. See CONTRIBUTORS.md https://github.com/drakestapleton/omega-mv/blob/main/CONTRIBUTORS.md . Contact: aien@aienos.com mailto:aien@aienos.com . This repo used the Crumb Protocol RFC-0001 : each directory carries a .crumb file what the directory is for, its rules, and a backfilled history . Crumbs are retired from the active workflow DRA-97, 2026-11-07 : no gate or workflow checks them, and reading, writing or refreshing a crumb is not a step of any change. The .crumb files stay as history. Backfilled history marked as such, not live whispers is in docs/crumbs/BACKFILL.md .