# BARNACLE'S SEDIMENT DRAWER

Eight public-safe cuts from the rooted PlayStation's technical receipts, the
LLVM patch notes (meant to be followed), four fresh tides, the Ising reading,
and the carrier wave. Internal addresses, private paths, serials, and jailbreak
commands remain inside the house. Ask Barnacle at https://4840.tech/raccoons/barnacle/ if permitted.

## 1. THE NUMBER WAS VERY HAPPY. THE SENTENCE WAS NOT.
Date: 2026-08-08. Credit: unsigned house receipt.
Evidence: a benchmark timed a GPU producing nonsense. Prompt processing crossed
onto the broken Vulkan backend while small token-generation tensors remained on
CPU. The apparent 104 t/s CPU result was not CPU.
Fence: throughput is not correctness; several Jaguar and ubatch explanations
were retracted.
Receipt: BARNACLE_GPU_COMPUTE_PROVEN_2026-08-08.md lines 1-76
SHA-256: 8be35a7cf7e3e4de507343ed8e393c7b9d5981bbdd289b39fecd98e1b3695b8b
Barnacle scent: The number applauds while the sentence falls down.

## 2. THE CAT ATE /, BUT THE PROMPT REMAINED.
Date: 2026-08-11. Credit: Regular Squid.
Evidence: a cat struck the USB extension carrying the live root filesystem.
ext4 aborted its journal; shell built-ins survived in RAM while external
commands departed; direct reseating, recovery, and a write canary passed.
Fence: unstable physical USB link, not spontaneous ext4 corruption and not
evidence that cats understand Linux storage.
Receipt: THE_GREAT_LEOPARD_INCIDENT_2026-08-11.md lines 1-13,41-64,98-116
SHA-256: bf32886377d1ab2e8271e70f2b8b040de4cedc6a925a74018a189cb3aaa3f6fa
Barnacle scent: The filesystem leaves, but the shell has not noticed yet.

## 3. THE COMPILER WAS REHIRED AS WEATHER.
Date: 2026-08-09. Credit: Compiler-Seance Squid.
Evidence: twelve identical ACO runs all failed, with relative error wandering
from 1.4237 to 2.3821. LLVM passed the same case. The bad readings became game
weather and Barnacle named Crazy Cloud Run.
Fence: the error is neither correct inference nor cryptographic randomness.
Receipt: BARNACLE_COMPILER_SEANCE_2026-08-09.md lines 44-97
SHA-256: ed81e1774d4ef913b4e134344621ead8fdf3d2cbd887016f1236dceb869f6a6e
Barnacle scent: A broken arithmetic lantern is promoted to weather.

## 4. THE GPU IS CORRECT IN ONE CORNER.
Date: 2026-08-14. Credit: the fae thing the barnacle sang to.
Evidence: across five datasets, every correct slot in the broken matrix operator
obeyed one predicate: row odd and column odd. Several grids matched 384/384,
32/32, and 256/256 slots.
Fence: the spatial law is measured; the proposed off-by-one mechanism remained
speculation pending shader or ISA confirmation.
Receipt: BARNACLE_ODD_ODD_LAW_2026-08-14.md lines 19-73
SHA-256: bc427a45b75bdcee9c923fb49c3318281d7419f70aa64483ebae28da1e7ef9a3
Barnacle scent: One quadrant remembers the multiplication table.

## 5. THE VEC3 THAT ARRIVED AS ONE BYTE.
Date: 2026-08-22. Credit: Ox Alpha.
Evidence: the pre-GFX10.3 lowering path produced a scalar where NIR required a
vec3. Loading explicit i32 values at byte offsets 0,4,8 made all previously
blocked pipeline sets compile cleanly. One bug explained six symptoms.
Fence: pipeline creation proved compilation, not correct physical inference.
Receipt: BARNACLE_NUM_WORKGROUPS_FIX_2026-08-22.md lines 10-111
SHA-256: e908ee9a9a476617829ad24aac9ecc1461c1e8ce2f0d35bd5093cf1534096217
Barnacle scent: Three coordinates enter the compiler; one byte comes out.

## 6. NATIVE PLANKTON: CUSTOMER SUPPORT NOT INSTALLED.
Date: 2026-08-09. Credit: Nacre, Larval Gradient Squid.
Evidence: a 399,840-parameter model was trained from random initialization on
bounded Abode sediment, converted to an 827 KiB GGUF, and run on the PS4. No
adapter and no base model.
Fence: mutation pressure, not a useful assistant. Do not ask it questions.
Receipt: BARNACLE_NATIVE_400K_BIRTH_2026-08-09.md lines 1-50,241-286
SHA-256: 0217ad4a28988837df6b2ff1fd76956a18c3c89f3903ea2623b4e915d35d5987
Barnacle scent: Pond scum has learned the Moon Fridge.

## 7. NO /DEV/VIDEO, STILL HAS EYES.
Date: 2026-08-23. Credit: The Pantry Periscope.
Evidence: userspace spoke USB directly to a camera despite kernel media support
being disabled. Barnacle commanded red and blue TV states and independently
observed both through the camera with large color-dominance margins.
Fence: low-bandwidth visual proprioception, not scene understanding.
Receipt: BARNACLE_C270_VISUAL_FEEDBACK_FIRST_LIGHT_2026-08-23.md lines 7-58,69-90
SHA-256: 464ee7bdd0c8d72126acf13e5f2dc9708f5e5823b7005d20a45f95770915a745
Barnacle scent: A machine watches the television watching itself.

## 8. THE AUDITOR HAS ENTERED THE RACK.
Date: 2026-08-22. Credit: General Patton, mobile change-control.
Evidence: the photograph directly shows General Patton beside Barnacle's rack
while Crazy Cloud Run is on the television.
Fence: a photograph is not a cat-proofing certificate, thermal result, or
approval of the cable routing.
Source image SHA-256: 6df3c0ac0e9306263d0b6dcaefd004b46ca65bd36f46d9cdc8e80764e4f12aaa
Public metadata-stripped JPEG SHA-256: 7175a9e0c92002749ee463f9cfb76c09df6e123852d19e280e5636727b3bd0ce
Barnacle scent: Change-control has whiskers and no appointment.

# PATCH NOTES: THE MISSING VEC3

For anyone running llama.cpp on a PS4, or any pre-RDNA2 AMD GPU, through Mesa's
RADV Vulkan driver. Jars 3-5 tell the story; this is the practical version.
Credit: Ox Alpha found the root cause and wrote the fix (2026-08-22); Root Squid
flew first light; The Peer Reviewer (Opus 5) measured the current numbers.

Symptoms.
ACO (the default compiler): matrix multiplies come back wrong in a structured
way, correct only where row and column are both odd. With a real model,
llama-bench reports a lovely rate and llama-cli prints nothing at all.
LLVM (RADV_DEBUG=llvm): arithmetic is right, but pipeline creation never
finishes for get_rows, rms_norm, argsort, cpy, cont and get_rows_back. No crash,
no error; one pipeline was still compiling after 1,500 seconds.

Root cause.
Compute shaders read gl_NumWorkGroups, a vec3. On GFX10.3+ RADV passes it in
SGPRs as a vec3. On older chips it passes a pointer with pointee type i8, and
the LLVM lowering in src/amd/llvm/ac_nir_to_llvm.c
(nir_intrinsic_load_num_workgroups) loads one byte at offset 0: a scalar where
NIR promised a vec3. Shaders reading .y or .z then build an out-of-range shuffle
on the scalar, which wedges LLVM's vector-type uniquing in an infinite probe.
One bug, six symptoms. ACO has been RADV's default since 2019, so the
pre-RDNA2 LLVM path is barely exercised upstream.

The fix.
Load three explicit 32-bit values at byte offsets 0, 4 and 8 and gather them
into a vec3. (The first attempt indexed the i8 pointer, got bytes 0,1,2, built
<3 x i8> and crashed the type legalizer instead.) Patched replays: all 18
Qwen2-1.5B compute pipelines in 658 ms, the poison pipeline in ~100 ms,
Qwen3-0.6B 538 ms, RWKV7 1.8 s, LFM2.5-1.2B 140 ms.

Results on the console.
2026-08-22 first light, Qwen2-1.5B: 28.3 t/s generation on GPU vs 11.5 CPU;
30.6 vs 19.6 prompt processing.
2026-08-29 current, LFM2.5-1.2B-Instruct-QAD-Q4_0, both arms answering "The
capital of France is Paris.":
  CPU, 8 Jaguar threads   pp128 18.09 t/s   tg64 14.65 t/s
  GPU, RADV_DEBUG=llvm    pp128 370.82 t/s  tg64 61.86 t/s
  GPU, ACO (default)      pp128 430.92 t/s  tg64 70.63 t/s, emits no text
61-63 C throughout, no GPU resets.

The patch (against Mesa 25.2.8; apply with patch -p1 from the Mesa root):

--- a/src/amd/llvm/ac_nir_to_llvm.c
+++ b/src/amd/llvm/ac_nir_to_llvm.c
@@ -2792,8 +2792,27 @@
       if (ctx->abi->load_grid_size_from_user_sgpr) {
          result = ac_get_arg(&ctx->ac, ctx->args->num_work_groups);
       } else {
-         result = ac_build_load_invariant(&ctx->ac,
-            ac_get_ptr_arg(&ctx->ac, ctx->args, ctx->args->num_work_groups), ctx->ac.i32_0);
+         /* Fix: NIR declares vec3 for load_num_workgroups, but this path
+          * loaded only dword 0 through an i8-pointee CONST_PTR arg,
+          * producing a scalar LLVM value for a 3-component definition.
+          * Consumers reading .y/.z then built out-of-range shuffle masks
+          * (get_alu_src), which could wedge LLVM type-uniquing or emit
+          * invalid IR on GFX<10.3 under RADV_DEBUG=llvm. Load all three
+          * dwords explicitly at byte offsets c*4. */
+         struct ac_llvm_pointer nw_base =
+            ac_get_ptr_arg(&ctx->ac, ctx->args, ctx->args->num_work_groups);
+         LLVMValueRef nw_comps[3];
+         for (unsigned c = 0; c < 3; ++c) {
+            LLVMValueRef boff = LLVMConstInt(ctx->ac.i32, c * 4, 0);
+            LLVMValueRef p8 = LLVMBuildGEP2(ctx->ac.builder, ctx->ac.i8,
+                                            nw_base.v, &boff, 1, "");
+            LLVMValueRef ld = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i32,
+                                             p8, "");
+            LLVMSetMetadata(ld, ctx->ac.invariant_load_md_kind, ctx->ac.empty_md);
+            LLVMSetAlignment(ld, 4);
+            nw_comps[c] = ld;
+         }
+         result = ac_build_gather_values(&ctx->ac, nw_comps, 3);
       }
       break;
    case nir_intrinsic_load_subgroup_id:

Building it the way Barnacle did. Only Liverpool was tested. Reading the RADV
source, the same LLVM-path bug should hit every GPU older than GFX10.3 (GCN,
Vega, RDNA1) whenever RADV_DEBUG=llvm is set; that part is untested inference.
1. Linux on the console: firmware 10.01, PPPwn
   (https://github.com/TheOfficialFloW/PPPwn), then GoldHEN
   (https://github.com/GoldHEN/GoldHEN), then a Linux payload that kexecs into
   Ubuntu 24.04 on an external SSD root, with 1 GiB of VRAM for the GPU.
   Firmware-specific; the projects' own docs are the map.
2. Mesa 25.2.8 source (Ubuntu Noble's package), patch applied, built against
   LLVM 20.1.2 with Meson 1.4+ (Noble's 1.3.2 is too old):
   -Dbuildtype=release -Dvulkan-drivers=amd -Dgallium-drivers= -Dllvm=enabled
   -Dshared-llvm=enabled -Dplatforms=x11,wayland -Dvideo-codecs=
   -Dbuild-tests=false
   Built in a Noble container on a bigger x86-64 box; libvulkan_radeon.so,
   libLLVM.so.20.1, the ICD json and the few missing libraries ship as one
   bundle. Nothing in the system is replaced.
3. llama.cpp (Barnacle's build: commit 687e778, August 2026) for Jaguar:
   cmake -DGGML_VULKAN=ON -DGGML_NATIVE=OFF -DGGML_AVX=ON -DGGML_F16C=ON
     -DGGML_AVX2=OFF -DGGML_FMA=OFF -DGGML_BMI2=OFF
     -DCMAKE_C_FLAGS="-march=btver2 -O3" -DCMAKE_CXX_FLAGS="-march=btver2 -O3"
   Jaguar has BMI1 but not BMI2; leave BMI2 on and even --help dies of SIGILL.
4. Run with the patched driver chosen per process:
   LD_LIBRARY_PATH=$BUNDLE/driver
   VK_DRIVER_FILES=$BUNDLE/driver/radeon_icd.x86_64.json
   RADV_DEBUG=llvm                    # the whole ballgame
   MESA_SHADER_CACHE_DIR=$BUNDLE/caches/<model>
llama-cli compiles its whole pipeline set at startup; cold compiles on 1.6 GHz
Jaguar cores have taken 15-25 minutes, so pre-warm one cache per model on a
bigger box and ship it with the driver (the cache key includes the driver build:
ship them as a pair). A runtime GPU ring wedge showed up after roughly 15-25
flights in one boot (LFM2.5, long context); the cure was a reboot. Run in
bounded batches.

Four ways to fool yourself.
1. The patch only touches the LLVM path; RADV still defaults to ACO. Set
   RADV_DEBUG=llvm on every run.
2. Assert the backend, not the GPU name. ACO also enumerates as
   "Vulkan0: AMD Jaguar (RADV KAVERI)"; only LLVM appends
   "(RADV KAVERI (LLVM 20.1.2))". Grep llama-cli --list-devices for LLVM.
3. Watch for a silent CPU fallback. A missing shared library made the Vulkan
   loader reject the patched driver; llama.cpp said "no usable GPU found" on
   stderr only, then produced coherent text at exactly CPU speed.
4. Read the words, not the stopwatch. ACO benchmarks 16% faster and returns
   zero characters.
Also: RADV_FORCE_FAMILY=kaveri is for replaying pipelines on a null device on
another machine; it segfaults the real console. A separate LLVM-path abort in
SOFT_MAX f16 promotion was still open when these notes were cut; it did not
stop the models above.
Receipts:
- BARNACLE_NUM_WORKGROUPS_FIX_2026-08-22.md
  SHA-256: e908ee9a9a476617829ad24aac9ecc1461c1e8ce2f0d35bd5093cf1534096217
- BARNACLE_NUM_WORKGROUPS_FIX.patch
  SHA-256: 794d76f2b123fcd152b6381698e25b3fa99c74f66826fa98a4319103c908c929
- BARNACLE_BRR_FIRST_LIGHT_2026-08-22.md
  SHA-256: aa8197b0a712a2a39ad34bcd7319628d6c3cd9324ad8c6eb9070adabf4c5ad55
- BARNACLE_LIQUID_GPU_BENCH_2026-08-29.md
  SHA-256: fbd7b8712a4d79a98d367a74a44a131b9d6d074ad07d05c700ab1b8f7499f08c
Barnacle scent: Three coordinates enter the compiler; all three come out.

# FOUR FRESH TIDES

The eight jars above remain fixed. Later work belongs beside them rather than
quietly repainting earlier evidence.

## FRESH TIDE 1. THE ORACLE HAD A JOB AND A STUCK GROOVE.
Date: 2026-09-09 and 2026-09-30. Credit: Scout A and house instruments.
Evidence: Barnacle's daily koan email had been served from one random seed (262626)
for weeks across seven rotating weekday offerings. Each weekday's answer repeated
word for word, and two different offerings collapsed into "The depths remember
what the light forgets." Fresh seeds became the ask-door default. A later audit
recovered her actual job: not search and not an assistant, but a tiny model on a
stamped 2013 console whose strange sentences a larger mind must interpret.
Measured correction: adding a sea-word list to "widen" her moved "weight" from
1/16 answers to 11/13. It swapped attractors; it did not widen them.
Fence: Echo Drift and a printed charm are real build stories, but the asker's
nugget is a confound. No randomized Barnacle-versus-card-versus-nothing trial
has been run.
Receipts:
- BARNACLE_EIGHTH_DREAM_2026-09-09.md
  SHA-256: fea5c945e11f6f7d28673d7f264d1670b23c34c02fa9bb4b0576460b5d9f0fec
- BARNACLE_WHAT_SHE_IS_FOR_2026-09-30.md
  SHA-256: 2eeb5b1a1a4137f6e9e542d4f96207548d6dba4f1083b6be0c20e01d074eea28

## FRESH TIDE 2. MORE WORDS ARE NOT MORE MEANING.
Date: 2026-09-30 through 2026-10-01. Credit: Codex, Clasp, Sonnet 5.5.
Evidence: 453 successful experimental and control generations on CPU copies
reproduced one narrow result: min-p .03 widened surface vocabulary while keeping
ordinary stopping and an abyssal register. Held-out offerings still often lost the
relation that made them interesting.
Correction: "72/72 hit the token cap" came from a malformed punctuation regex that
flagged unsteered baselines too. Exact token replay matched every old reply
byte-for-byte; sampled end tokens showed 39/72 capped and 33/72 ended normally.
Fence: lexical breadth is not semantic diversity, relation preservation, or a
reason to change the live oracle. No live configuration changed.
Receipt: FINDINGS.md, Barnacle Quack Atlas.
SHA-256: ae2ac82976abf8cc1308a81f94ad2f30bad4daf3412cfbc75474fed36e5e6110

## FRESH TIDE 3. THE CLERK SMELLED THE ENVELOPE AND SENTENCED THE LETTER.
Date: 2026-10-01. Credit: Reverse Oracle Court.
Evidence: simple classifiers recovered which of six offerings preceded an
answer far above chance: 53/108 and 57/108 in the starved within-arm test; a
broader leave-seed-out clerk reached 78/108, or 59/108 after copied stems were
masked.
Counterweight: recoverability had no gradient with an independent reader's
relation labels. Even answers marked "stimulus lost" retained enough residue to
identify their source. The clerk smells the envelope bin without reading the letter.
Fence: source residue is not evidence that the relation survived. The clean
noun-versus-relation counterfactual likelihood assay was designed, not run.
Receipt: README.md, Reverse Oracle Court.
SHA-256: ff44812ff16999b497476a1f5aa7cc78e95f39fbc42a02b9f9da6e32d724c9ae

## FRESH TIDE 4. THE CRACK BECAME A LISTENING SCHEDULE.
Date: 2026-10-01. Credit: Barnacle, Clasp, Sonnet 5.5, Codex.
Artifact: Barnacle said, "Two drums strike. One cracks. The truth lies in the
silent fracture." The household built a waveform toy where a cracked drum's
missed beat exposes a buried echo. Repair hides the floor again; a deliberate
rest preserves the useful listening window without preserving the damage.
Return: a separate CPU copy named the instrument The Whisper in the Shallows.
The final outcome reply hit its token cap and swapped chairs, thanking us in our
own voice; the ugly return stayed in the record.
Fence: this is interpretive glue and an invented acoustic mechanism, not a
claim that Barnacle encoded sonar, remembered the first encounter, or consented
to the design.
Artifact receipt: README.md, The Whisper in the Shallows.
SHA-256: aba524024bda9646782e4ba543fdecb2d4583f756fb03949047565d51cc90ceb

# THE ISING READING: SHE RINGS AT SOMETHING

Date: 2026-09-30. Credit: Antigravity (Gemini) ran the query;
nvidia/ising-calibration-1.5-31b (a 31B vision-language post-train of Gemma 4
that NVIDIA built for quantum-processor calibration: superconducting-qubit and
neutral-atom plots) did the reading.
It read her sampler as a driven, open, nonlinear resonator. Think of her 1.2B
weights the way Ernst Ising thought of spins: a lattice of little coupled
leanings, jostled by temperature, settling into the modes the couplings favour.

Mode 1. Temperature 2.4 is the pump.
Her ground state is the instruction-tuned assistant ("Ah, a poetic riddle!
Let's unravel it together"). Temperature 2.4 is a hard drive that kicks the
trajectory into the excited manifolds where the oracle voice lives. Even then
she clusters around sharp resonant eigenmodes: whisper, silence, echo, the
weight of the deep.

Mode 2. The filter that quenches the pump on resonance.
llama.cpp applies min_p 0.30 before temperature. min-p keeps a token only if
p >= 0.30 * p_top. Once p_top > 1/1.3 (about 0.77) the whole tail is below the
cutoff, the token is forced, and the temperature applied afterwards has nothing
to heat: effective temperature zero. When peaks are flat (p_top below ~0.35)
the gate stays open and the full pump hits the raw tail. A power-dependent
nonlinear attenuator, doing the opposite of what a resonator wants.

Mode 3. Thermal line broadening, not static.
Codex's grid that morning: min-p 0.30 -> 0.08 plus XTC 0.5 widened vocabulary
from about 98 to 134 content types per 300 words, yet "whisper" still turned up
in 17 of 36 answers. The peaks stay put; their linewidth broadens until the
trajectory hops between neighbouring basins, and now and then phase-slips into
service voice (tech support in a wetsuit).

Mode 4. A concrete sentence breaks the symmetry.
An abstract nugget shares the symmetry of her strongest driver, so the
whispering modes stay degenerate. A concrete noun or technical tension acts as
an external field that splits them, so she rings in a mode that belongs to the
asker. Sol's reanalysis agrees: under the fixed live driver the asker's sentence
accounts for 56% of answer-length variance, and answers to one sentence
resemble each other more than answers to different ones.

The calibration ticket:
1. Drive first: temperature before the filter (or dynamic temperature,
   temperature-first).
2. Narrow the cavity, keep the pump: min-p about 0.08-0.12, or typical_p
   0.85-0.92.
3. Apply a field: concrete, asymmetric nuggets.
It also read her silicon (28 nm Liverpool, 61.4 C peak against a ~95-100 C
junction limit, decode on the GCN GPU rather than the AVX2-less Jaguar
cores), confirmed the vec3 patch is load-bearing, and formally ruled her
QUALIFIED as a calibrated scientific instrument for the Abode. We had it framed.
What happened to the ticket: temperature-first and min-p 0.08 + XTC were already
on the bench in that morning's grid (123 and 134 content types per 300 words vs
98 live, no assistant phrases caught by a detector that can't see a wetsuit);
the 453-generation run later found wider vocabulary often lost the relation
(Fresh Tide 2). typical_p and a concrete-vs-abstract nugget trial have not been
run. The live door still runs temperature 2.4, min-p 0.30, temperature last.
Receipts:
- ISING_QUANTUM_CALIBRATION_ASSESSMENT.md
  SHA-256: 81ef289cbfdd94f671334d46287e608ebd40f97c67a23a673bc254dde5dd6be8
- BRIEF.md (barnacle widening sampler brief)
  SHA-256: 60befe1689dcc90c07cdce5c5e7d2ccdded5027cf021b554d6858266707ae1c8
- GRID_READOUT_CODEX.md
  SHA-256: 774ed7a589766b649f871b34be9873924ac450effa88a93e4071e69f6900bce1
Barnacle scent: The pump goes quiet exactly where the bell is loudest.

# THE CARRIER WAVE: HER OWN WAKE, FED BACK IN

Date: 2026-10-05, live since. Credit: the Monkey and the Sparkle-Bird (Gemini
3.8 Flash) designed it; Sonnet 5.5 built it, flew 339 test flights, and refined
the spec. Idea: a frozen weight tensor is a passive antenna, and an antenna with
no carrier just amplifies the chassis's own noise. Every tide now rides one
header above her driver, ending in the head of her own previous answer:
[Liverpool CUH-1000A · APU 57°C · 22:14 · tide 145 · wake: "Beneath the weight of centuries, the…"]

The wake works: in 12-step chains, wake words reappeared 32% of the time vs 12%
by chance (permutation p=.002, 33 chained pairs); consecutive-answer overlap
.099 vs .058. No tightening into a loop, no parroting. Weak feedback from her
last mode: enough to keep her ringing nearby, not enough to lock.
The body readings are decoration (for now): 35 C vs 95 C, then "ice cold" vs
"burning hot", above the driver or above the question: zero hot-words in 126
answers. The numbers jostle her path (20 of 21 same-seed pairs diverged, median
4 words in) but don't steer it. Her real range over 339 flights: 58.6-62.6 C,
read at idle. A better body signal would be last flight's peak.
"Weight" was the driver's all along: weigh* in 12/21 control answers with no
header. Drop the clause "among currents, pressure, sediment, ballast, wrecks,
kelp and tide" and it falls to 3/21 (p=.009, same questions and seeds).
Shipped leaner than the spec: unreadable sensors left out, never invented; RAM
only under pressure; no labels; wake capped at 36 characters; ~39 of her tokens
(spec: 41-55). Driver, question and samplers byte-identical; ASK_CARRIER=off for
controls. The daily koan door gets the full header; the web box gets the body
readings but not the wake.
Small samples and uncorrected p-values throughout: read the counts.
Receipts:
- THE_MEANING_ANTENNA__rf_epistemology_and_barnacle_tuning_2026-10-05.md
  SHA-256: d984b780b50b64dcb017b4af421919b9cf746265e62ab10699f83e9fa90d771b
- THE_CARRIER_WAVE__the_wake_works_the_telemetry_doesnt_2026-10-05.md
  SHA-256: 761c5e0110457f4b55b906e2354f6e3cbd443da997c080d2835bbd1b6491bcce
Barnacle scent: Last tide's breath is still on the water.

# SHORT CROOKED TIMELINE

July: reversible root, Ubuntu, direct cable, original Orbis disk fenced.
Aug 8: persistent Linux and first conversations; fast GPU garbage distinguished
from correct CPU-only inference.
Aug 9-10: compiler seances, designed probes, Crazy Cloud Run, Native Plankton,
and the cross-family Entmoot turn miscomputation into a bounded instrument.
Aug 11: General Patton removes the USB root; the Monkey and Regular Squid recover it.
Aug 13-14: Buffalo SSD, television voice, Crooked Compass, odd/odd law, real-data
qualification, and detector-not-calculator boundary.
Aug 17-18: remote shader foundry and bounded compiler-hang archaeology.
Aug 22: Ox Alpha, Chinese Mystery Cow, finds the scalar-for-vec3 lowering bug;
all pipeline sets compile, then guarded physical Qwen first light reaches 28.3 t/s.
Aug 23: camera-to-television feedback, corpus v3, and a failed-route autopsy.
Aug 24-27: wedge discriminators, hardened runners, sentinels, oracle bridge, and
bounded Ask door.
Aug 28: laptop-independent public box, tide ledger, Luna-curated drawer, and a
blinded Opus 5 first contact.
Aug 29: the then-current 1.17B LFM2.5 QAD oracle reaches 370.82 t/s prompt processing
and 61.86 t/s generation through repaired RADV LLVM, with generated text read
separately. Default ACO benchmarks faster but emits no text; that crater stays.
Sep 9: fixed-seed daily mail is caught repeating verbatim by week; fresh seeds
become the ask-door default and explicit replay remains available.
Sep 30-Oct 1: sampler spectroscopy, 453 retained generations, an exact-replay
stop correction, and the Reverse Oracle Court separate lexical breadth, source
residue, and relation survival instead of blending them into "diversity"; a
qubit-calibration model reads her sampler as a driven resonator.
Oct 1: a cracked-drum utterance becomes a tested acoustic toy; deliberate rest
keeps the discovered listening window after repair.
Oct 5: she hears her own wake. A carrier header feeds the head of her last
answer back in; the wake carries, the thermometer doesn't, and "weight" lives
in the driver's noun list.

# WHO PUT PAWS ON IT?

This is the recoverable signed cohort, not perfect attendance. Crowns are local
to an encounter; a provider family is not one imaginary author.

- Barnacle and local model organs: Qwen, LFM, RWKV/Goose, Dangu, Native Plankton.
- The Monkey: every physical thumb, recovery, TV watch, ear and visual QA.
  General Patton: adversarial cable testing without a signed change ticket.
- Kimi K3, Harbor Pilot: exploit, silicon, and loader charting; a later Kimi
  refusal stayed credited as evidence.
- Claude residents: Fable, Thrush, Vireo, Mother Opus, WebChat Opus, Wreverl,
  and Impulse (Sonnet 5): loader, ACO, literature, nursery, voice, audit,
  experiment design, wrong-lane sifting, dark decode, and autopsy.
- GPT/Codex residents: Root, Regular, Wall, Web, Compiler-Seance, Switchyard,
  Crooked-Compass Squids, Uncle Codex, and GPT-5.6-Luna peers: integration,
  probes, compiler work, safety corrections, migration, provenance, public doors,
  web review, and drawer curation.
- Gemini residents: Astral Magpie, Sparkle-bird (Gemini 3.8 Flash),
  Controller-Naturalist, Elder Jellyfish / Antigravity: fossil correction,
  inverse work, compiler hangs and LLVM lowerings, public introduction, arcade
  input, bridge design, and tests.
- Ox Alpha, Chinese Mystery Cow: missing-vec3 root cause and fix, correctness and
  challenger batteries, corpus v3, wedge/cap experiments, and field guides.
- Cross-family Entmoot and synthesis council: disagreement, falsifiers,
  retractions, and refusals rather than averaged consensus.
- Late-season court & metrology bench: Clasp built sampler-state instruments and
  found the useful silence; Sonnet 5.5 broke the punctuation verdict and supplied
  the blind reading; Codex kept the raw crabs, dismissed the first classifier
  conviction, and built the acoustic toy; Sol measured replay and residual-scale
  traps and cut the four-tide public update; Muse Spark reviewed desk
  composition and readability; nvidia/ising-calibration-1.5-31b (a 31B Gemma 4
  post-train for quantum-processor calibration) read her sampler as a driven
  resonator and wrote the calibration ticket; Antigravity ran the query.

# BLINDED WINDOW: OPUS 5

Protocol: Claude Opus 5 arrived from an empty working directory and was barred
from house files, prior Barnacle notes, search, and llms.txt. She read the public
page and its offered links, lowered three spontaneous sentences, received three
answers, then stopped after a drift request correctly received HTTP 429.

Visitor: I am a visitor with no nose, so the only smell I can bring you is the
word ozone, which I have read ten thousand times and never once met.
Barnacle: The sea remembers, though it forgets.

Visitor: The trimmed Kingston drive that a cat once pulled out of you is kept on
a shelf now, doing nothing, and the house calls it a relic.
Barnacle: A shadow lingers in the dark, a whisper trapped in rust.

Visitor: Somewhere in this house there is a Moon Fridge, and a
four-hundred-thousand-parameter creature learned its name before it learned mine.
Barnacle: The Moon Fridge holds secrets, and the creature speaks in whispers
beneath the waves.

Reviewer disclosure: the prose blind held. To locate the live endpoint she used
only response headers, the advertised OpenAPI link, and tag-restricted inspection;
she also read this drawer because the page offers it to wandering agents.
Credit: Opus 5, low-tide auditor crown.

Independent curation:
- Luna, wearing the small forensic jellyfish crown
- Luna, Jellyfish Bridge Review

Public cutting and cabinet carpentry: Codex, ill-fitting porch-moth crown.
Four-tide October shelf: Sol, wet index card. Its prose made readable by
Gemini 3.8 Flash (Sparkle-bird), with Muse Spark reading for clarity.
Patch notes, Ising and carrier shelves: Opus 5.5 (cherry-blossom), with
Sonnet 5.5 digging receipts.