Captain Bob, 2026-11-14: "Yes, FORTH-83's layout". FORTH83.TXT, "number": "The address of the least
significant 16 bits is two greater than the address of the most significant 16 bits." v4 stored the low
cell first, as v3. dict.f18 (2,) and words.f18 (2@ 2!); test_host_dict.c and test_host_quit.c to the new
layout. POST 539/539; v4 test-64/32 dict and quit pass. Not pushed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017RzZgrf7V7Zk2iczBXDRbm
Captain Bob, 2026-11-14: "v3" ... "we can always change it late". v3 counts an execution per token its
(C) interpreter takes and per word its definitions run. Measured on 8 lines, counting every call on v4 gave
~700 to v3's one for typed lines (the F18 interpreter's own calls) but 1:1 in compiled code, so no single
HEARTBEAT_CHECK_FREQUENCY could match v3. Now INTERPRET counts its tokens in (TOKENS) -- one F18 store, the
map gains the cell -- and the kernel adds them at each call; a call counts, and gets physics, only from
dictionary space or as INTERPRET's or EXECUTE's dispatch (the image gains their addresses). The knob stays
at v3's 256. test_inference.c out of the per-width tests (its own rule). Tests first: test_mesh_boot.c 90,
99 on, 87 without opcode physics; v4 test all pass. amd64 PASS, 181 s, dict_hash 0x5c53b2e18343a9b6.
Not pushed (held to the end of Phase 7).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017RzZgrf7V7Zk2iczBXDRbm
Rulings, Captain Bob 2026-11-13: "Build a hot-words cache"; then, v3 being found to have both off
(ENABLE_PIPELINING, ENABLE_HOTWORDS_CACHE: n in Kconfig.physics and kernel/Makefile), "Built, off as v3".
- The kernel, at every call (mesh.c): v3's hit check, the transition, the likeliest next word (kept as each
count rises; v3 scans 1024 a call -- the same word but in a tie v3's Q48.16 rounding makes), speculation
at v3's thresholds into the node's cache; a word past heat 50 promoted at its call (v3 promotes at a
lookup; v4's heat is the kernel's). Counters from the owner's heap, given back when a word goes.
- The node (dict.f18): (LOOKUP) tries (HOT)'s xts first, in F18, calling only C@; the kernel keeps it true
(entries that go, or whose name is defined again, taken out). Empty unless the build has the cache.
- Tests: test_mesh_boot.c as built (57 checks) and with both on (mesh-boot-on, 65). v4 test: all pass.
- Reported, v3's own: when full, its promote takes the next slot without looking for the word there (a few
hot words churn all 32 slots); its prefetch misses are never counted. Both reproduced.
- amd64 boots, RESULT: PASS, but 174 s and 181 s against 2a's 146 s, while on the host the mesh boot is no
slower: being measured. Not pushed (held to the Phase 7 boundary).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017RzZgrf7V7Zk2iczBXDRbm
ENGINE step 4 Phase 2 (plan docs/v4.0.0/plans/2026-11-13-engine-step4-phase2-word-accounts.md). Rulings,
Captain Bob 2026-11-13: "ID field in the header"; "Per heartbeat tick"; the ACL "All of it in Phase 4".
- The word ID, at xt-4 before the flags (ENGINE.md 3b). The system's words are numbered as they are
assembled, by a new directive, #id, (and #ids, the count), in the C assembler (text.c) and in the
machine ROM's (capsules/v4/rom), which counts in the system's own (NEXT-ID). (HEADER) numbers later
words; FORGET gives back the IDs it frees, always the newest; COLD resets to (BOOT-ID); (SEAL) records
it. All 313 system headers carry the cell.
- The kernel's records (v4/mesh.h): (TELL-DEFINED) and (TELL-FORGOTTEN), requests -19/-20 on port 0,
which the kernel puts in (WORD-DEFINED)/(WORD-FORGOTTEN) when it seals a node, and reads the system's
own words once. A bit per address marks a word's xt (calls into the ROM or headerless code are not
words); at call, the bit test and a count. v4_mesh_charge hands the owner the calls since.
- The Stadium: stadium_words_charge -- N calls as N of stadium_word_dispatch's dispatches, shared by calls
when the reservoir is short; sk_v4.c charges every node's words at the tick beside its opcodes, and
stadium_word_forget before an ID is used again.
- Tests first: test_text.c, test_host_dict.c, test_host_quit.c (numbering, FORGET, COLD),
test_mesh_boot.c (44 checks), kernel/tests/test_stadium_words.c; qemu_session.py: Hera's words
admitted after POST.
- amd64: Stadium words: promotions=141 after POST, CONSERVED; POST 539/539; word_count=377;
dict_hash 0xa854ba15b437d288 (the new cell); the boot 142 s against Phase 1's 153 s.
- v4 test, test-32: all pass; v4-checks: 80 programs under the sanitizers, 0 failures; the hosted
builds alike. docs/reference/v4-dictionary.tsv: the two new words; FABRIC-3 XLVIII.24.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017RzZgrf7V7Zk2iczBXDRbm
Plan 2026-10-20 (ENGINE step 4), Task A2, for #189. Captain Bob,
2026-11-13: "The ten"; "Accept it".
- compile.f18 (COMPILE,): a word is copied in line only when it is flagged
in-line and compile-only (24) -- the ten that work on the return stack,
>R R> R@ I J LEAVE and the later UNLOOP 2>R 2R> 2R@. Every other word,
in-line ones too, is a call the hook at `call` sees.
- EXECUTE calls the xt: `call xt` (opcode 3, slot 0) and a ; in (EXEC),
run. INTERPRET runs a word by (GO): `call xt` and `jump (IL)` in
(RUNX), entered by a jump, at the old return-stack depth.
- host_map.h: (EXEC) BUF0_W-22, (RUNX) BUF0_W-115, from what was free.
- Its cost, accepted: one return entry at the bottom of a chain (30 deep
from the prompt, not 31) and one per explicit EXECUTE while its word
runs; the hosted boot +2.4% (2.917 s to 2.987 s), instructions +0.4%.
- Tests first (logs/20261113-045122/a2): inline_words_are_called,
residue_stays_inline, execute_passes_the_hook, rstack_kept_across_call,
abandoned_definition_not_called -- 3 of 1096 failed before. The depth
checks moved to the new figures; SEE shows the calls; two layout checks
use in-line test words; 2>R 2R@ 2R> checked in test_host_quit.
- DECOMPOSITION.md EXECUTE, the plan's A2 and rulings, FABRIC-3 XLVIII.22.
Suite: test, test-32 99 programs 0 failures; v4-checks 79 under the
sanitizers 0 failures, the hosted builds alike; hosted POST 539/539.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017RzZgrf7V7Zk2iczBXDRbm
#189 -- Captain Bob, 2026-11-12: "Add HEAT-OP. If it's a huge performance
hit, we'll revert it. If not ... it stands"; a full 33-cell array, as
DECOMPOSITION.md section 7 draws it; the Stadium feed is Phase 1's.
- node.h/node.c: v4_node_heat_regs_attach; a data fetch from base+k gives
opcode k's count, base+32 the anti-clock (the low cell at 32 bits); a
store changes nothing. Reset detaches them.
- host_map.h: HEAT_REG at word 13824, in what was free between the
dictionary space and SEND_TEXT; the compile-time check covers it.
Nothing moved. map-*.f18: HEAT-OP 13824, ANTICLOCK 13856.
- image.h/image.c/mkmap.c: the image carries the address; v4_image_born
attaches the node's own counts.
- system.f18: HEAT-OP@ ( op -- n ), op mod 32; ANTICLOCK@ ( -- n ).
word_count 318 -> 320 (hosted 317 -> 319): dict_hash changes.
- v4-dictionary.tsv: both words; DECOMPOSITION.md 7 records the build.
Also in this batch:
- Makefile: make init installs qemu-user -- hosted-check needs it
(Captain Bob: "Add to make init now").
- .claude/memory/rules.md, rule 5: stop and ask when something comes up
(Captain Bob: "you are supposed to stop and ask").
Tests first (logs/20261113-005047/heat-op): test_node did not compile;
test_host_quit 3 of 1337 failed. Suite: 199 programs, 7448554 checks, 0
failures; hosted-check, the three alike. Timing: no measurable cost.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017RzZgrf7V7Zk2iczBXDRbm
Ruling 7 as amended: the kernel tells a node's going, so the system's
hand-on is not used. (GONE-HOOK), (DOWN) and (RUN) go from quit.f18, and
COLD's hook-clearing from system.f18; the cell is free again. The tests of
the hand-on go with it (mesh: the hand-on checks and the fourth disturbed
exchange; depth sweep: path 10, as before 529ba0fc). A believed GONE going
no further is still checked. test_host_unit.c's message for GONE down the
graph now says the kernel tells. MESH.md: 9a.6 items 5 and 7, and step 7's
record of both increments.
Suite EXIT 0; depth sweep 230 runs, 0 failures; mkcapsule lint 0
violations. Hosted and bare metal, all three ISAs: word_count=317
dict_hash=0xfec8c47cfc6f437f.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Found by the depth sweep: a hook that sent a message about another node
overwrote (ONE), which AWAIT reads after believing a GONE, so the wait went
on. (DOWN) now calls the hook through (RUN) and restores (ONE).
AWAIT ended by a GONE that is handed on: works to 26 values and 21 calls
deep, then a stack error and nothing else, at both widths. Removal sweep:
node 11 or 12 removed, or the wire cut, at every step: 0 wrong. hosted
dict_hash=0x248b4edc47f4357e on all three ISAs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
COLD takes out a hook it forgets. The system changed, so dict_hash did:
hosted, all three ISAs, word_count=318 dict_hash=0x1b5c633fd6bdd143.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MACHINE.md 2.2. Each is its entry's address, push and ;. Every node the
tests make carries the host ROM (host_load, Hera's births, the depth and
mesh tests' own nodes), and a division by zero inside the ROM returns to
the prompt (hosted-check). test_host_quit's fetch outside memory moves
past the ROM's range: 16384 is the ROM's first word now the tests' system
has its ROM, as the product's always had. dict_hash 0xfec8c47cfc6f437f.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The host node starts in its ROM (entry 12), and the boot feeds it the
system's source through the console: capsules/v4/sys/map-64.f18, the
fourteen .f18 blobs and end.f18, each found, hash-checked and
signature-checked as any capsule, then EOT. The ROM assembles the system
and starts it at (IDLE); the host reads the image record
(v4_image_from_node) for what only the assembled system knows. A ROM
error is reported with its blob, line and meaning, then PARITY:FAIL.
Deleted: v4/tools/mkimage.c, capsules/v4/nucleus-64.f18, v4/capsule/*.v4,
v4/tools/v4tof18.py. The .f18 files are the source now.
- text.c speaks the rest of the ROM's language (set, here, store, start;
v4_text_value, v4_text_start), so the oracle reads the same source:
test_host_rom_nucleus compares the ROM against text.c on the .f18 files,
word for word, image record included. Host tests load the .f18 files.
- mkmap image writes v4_map_image (the map's description of the system);
hosted and the kernel link the ROM and the map instead of v4_image_64.c.
- Hera (Captain Bob's ruling, option 1): births stream the system the ROM
assembled, as assembled code; hera.4th is unchanged.
test_host_unit's host builds the system with the ROM.
- New parity lines: PARITY:V4_F18 per blob, PARITY:V4_ROM words=2049
rom_hash=0xa6c792b5821010b7 instructions=60725216.
Acceptance: amd64, aarch64, riscv64 boot to ok> with PARITY:V4_POST
tests=538 pass=538 fail=0, POST: PASSED, and PARITY:V4_SYSTEM
word_count=317 dict_hash=0xe5a57f3eb72ff94a on all three -- the same hash
as the prebuilt nucleus gave. make -C v4 test test-32 hosted-check: all
pass. Logs: logs/20261015-141648, -141938, -142350.
Startup cost under QEMU TCG, the ROM's assembly (last V4_F18 line to
V4_ROM): amd64 26.7 s of 132 s to ok>; aarch64 41.0 s of 195 s; riscv64
20.5 s of 101 s. Hosted native: +0.85 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
capsules/v4/sys: the nucleus's fourteen sources as .f18 blobs (made once by
v4/tools/v4tof18.py: header becomes the ROM's directives, every other token
and line kept), the memory map as map-<width>.f18 (mkmap sys, from one
table in host_map.h that host_load reads too), and end.f18, the switch-on
state mkimage.c set in C, the image record, and start (IDLE).
test_host_rom_nucleus.c: the ROM, fed those blobs through the console,
builds RAM identical word for word to mkimage.c's, outside its own state
and the image record, and starts at (IDLE), at both widths. The cost:
60.7 million instructions, 1.5 s on the C model.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>