Kev-0.8B Core AI: Metal-kernel graph, 128 tokens per call (K128) replaces the unrolled S=16 bundle; card v2

.gitattributes

@@ -35,3 +35,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 gpu-pipelined/kev_0_8b_decode_fp16_pf16/kev_0_8b_decode_fp16_pf16.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
 gpu-pipelined/kev_0_8b_decode_fp16_pf16/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/kev_0_8b_decode_fp16_metal_pf128.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -24,16 +24,16 @@ Apple Core AI (`.aimodel`) conversion of [jaredpalmer/kev-0.8b](https://huggingf
 
 | Variant | Path | Size | Requires | Tested on |
 |---|---|---:|---|---|
-| Decoder, fp16, with the host's pointer head | `gpu-pipelined/kev_0_8b_decode_fp16_pf16/` | 1,521 MB | macOS 27 / iOS 27 | M4 Max, macOS 27.0 (26A428), 2026-10-03; iPhone 18 Pro, iOS 27.0 (24A437), 2026-10-04 |
+| Decoder, fp16, with the host's pointer head | `gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/` | 1,520 MB | macOS 27 / iOS 27 | M4 Max, macOS 27.0 (26A428), 2026-10-04; iPhone 18 Pro, iOS 27.0 (24A437), 2026-10-04 |
 
-One question of 94 tokens takes 112 ms on the M4 Max and 157 ms on the iPhone 18 Pro, in Swift ([measured below](#time-per-decision-on-the-mac)).
+One question of 94 tokens takes 29.9 ms on the M4 Max and 37.8 ms on the iPhone 18 Pro, in Swift ([measured below](#time-per-decision-on-the-mac)).
 
 ```swift
 import Kev
 
-let bundle = URL(filePath: "Kev-0.8B-CoreAI/gpu-pipelined/kev_0_8b_decode_fp16_pf16")
+let bundle = URL(filePath: "Kev-0.8B-CoreAI/gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128")
 let kev = try await KevDecider(bundle: bundle)              // asset: nil = the .aimodel, specialized here (GPU, frequent reshapes)
-let response = try await kev.decide(requestJSON: requestData, shared: true)   // shared: the state's chunks run once
+let response = try await kev.decide(requestJSON: requestData, shared: true)   // shared: the state's whole calls run once
 print(PythonFormat.dumps(response, asciiOnly: false))
 ```
 
@@ -56,14 +56,14 @@ option's closing token against the question's last token. The author's card repo
 results; none of them are re-measured here.
 
 This port merges the adapter into the base with the author's own merge script and exports the text
-backbone as one Core AI graph. The graph takes token ids, 16 per call, and returns the final-norm
-hidden state at every position; it has no vocabulary head. Weights are fp16 (1.51 GB). The host runs
-the pointer head from the author's head weights. The gate is probability parity with the author's own
-fp32 code, on every option of every question, on a fixture of 434 questions and on a held-out set of
-130.
+backbone as one Core AI graph. The graph takes 128 token ids per call and returns the final-norm hidden
+state at every position; it has no vocabulary head. Each Gated DeltaNet layer runs its recurrence in an
+fp32 Metal kernel, one GPU dispatch per layer per call. Weights are fp16 (1.51 GB). The host runs the
+pointer head from the author's head weights. The gate is probability parity with the author's own fp32
+code, on every option of every question, on a fixture of 434 questions and on a held-out set of 130.
 
-It runs on the Mac GPU and on an iPhone 18 Pro. On the phone the shipped `.aimodel` is specialized on
-the device; an AOT asset compiled for the phone's GPU (not shipped) was measured too.
+It runs on the Mac GPU and on an iPhone 18 Pro. On both, the shipped `.aimodel` is specialized where it
+runs.
 
 ## Readout contract
 
@@ -96,19 +96,19 @@ row form (`rows_of`), with the base model's tokenizer:
   `(p_max − 1/K) / (1 − 1/K)` and every p; `score` → Σ level · p, the legend, every p and a confidence
   `max(0, 1 − E|level − mode| / D)`; values rounded to 4 decimals. `usage` = `{input_tokens,
   output_tokens}` with the author's counting (the packed request; the tokens of `json.dumps(answers)`).
-- Length: a row holds at most **4,080 tokens** (the last padded call must end below the graph's 4,095
-  position bound). The author's server accepts states of 65,536 tokens; here a longer row is refused
-  before any call.
+- Length: a row holds at most **3,968 tokens**. The last call is padded to 128 tokens and must end at or
+  below the graph's position bound of 4,095. The author's server accepts states of 65,536 tokens; here a
+  longer row is refused before any call.
 - Shared prefix (optional, exact): every row of a request starts with the same state tokens, so the
-  first ⌊Ls / 16⌋ chunks can run once, the four states be copied per question, and each row continue
-  from there. Every row's hidden state equals its direct run bit for bit (below).
+  state's whole calls (⌊Ls / 128⌋ of them) can run once, the four states be copied per question, and each row
+  continue from there. Every row's hidden state equals its direct run bit for bit (below).
 
 `conversion/kev/oracle_kev.py` runs the author's package unchanged (`Checkpoint.load("cpu", fp32)`,
 `kev.model.admit`, the row-form `forward`, one CPU thread) and records the reference:
 [`fixtures-kev-0.8b.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/fixtures-kev-0.8b.json). The fixture is 384 records and 434 questions:
-the first 60 lines of the author's transfer-v4 development file (`evals/v4/transfer-v4/development.jsonl`,
-tag `kev-1.0`; MMLU, 4 options), the first 20 records of each of its seven other sources (140, `choice`
-and `noul`), its first 20 `score` records, the 144 `choice` items of SemIf authored144, and 20 records
+lines 1–60 of the author's transfer-v4 development file (`evals/v4/transfer-v4/development.jsonl`,
+tag `kev-1.0`; MMLU, 4 options), records 1–20 of each of its seven other sources (140, `choice` and
+`noul`), its `score` records 1–20, the 144 `choice` items of SemIf authored144, and 20 records
 written for the port (70 questions: `noul`, `choice` with 2–10 options, `score` with 2–7 levels,
 JSON states, three states of 1,431–1,746 tokens, one record of 8 questions). The longest row is 1,802
 tokens. Fourteen questions have an oracle top-2 margin of 0.02 or less (near-ties) and are listed apart.
@@ -119,27 +119,31 @@ The held-out set is the next 130 transfer-v4 records (mmlu 61–100, records 21
 
 `conversion/kev/qwen3_5_kev_decoder.py`: the overlay's stateful Qwen3.5 text decoder
 (`Qwen3_5StatefulForCausalLM`) on the merged weights, the vocabulary head replaced by the identity, the
-output the final-norm hidden state at every position. One function, `main`, at a static 16 tokens; the
-Gated DeltaNet recurrence is unrolled in the graph.
+output the final-norm hidden state at every position. One function, `main`, at a static 128 tokens. Each
+Gated DeltaNet layer runs its recurrence in the overlay's fp32 Metal chunk kernel
+(`qwen3_5_gdn_metal`): one GPU dispatch per layer per call, the recurrent state kept in fp32 through the
+call.
 
 | | name | shape, type |
 |---|---|---|
-| inputs | `input_ids` | [1, 16] int32 |
+| inputs | `input_ids` | [1, 128] int32 |
 | | `position_ids` | [1, seq] int32, the ramp 0 .. seq − 1 |
 | states | `keyCache`, `valueCache` | [6, 1, 2, ctx, 256] fp16, ctx up to 4,096 |
 | | `convState`, `recState` | [18, 1, 6144, 3], [18, 1, 16, 128, 128] fp16 |
-| output | `hidden` | [1, 16, 1024] fp16, every position |
+| output | `hidden` | [1, 128, 1024] fp16, every position |
 
-Per row of T ids, from zeroed states: ⌈T / 16⌉ calls, call k with ids[16k : 16k + 16] and position_ids
-0..16k + 15; the last call is padded with `<|endoftext|>` (248044) and its padded rows are dropped. The
+Per row of T ids, from zeroed states: ⌈T / 128⌉ calls, call k with ids[128k : 128k + 128] and position_ids
+0..128k + 127. The last call is padded with `<|endoftext|>` (248044) and its padded rows are dropped. The
 rows of every call, cut to T, are the backbone's final-norm `last_hidden_state [T, 1024]`. The bundle's
-`metadata.json` says `kind: decision-backbone` and carries this order, the row layout, the delimiter ids,
-the render rules, the head formula and the response shape.
+`metadata.json` says `kind: decision-backbone` and `language.prefill_chunk: 128`, and carries this order,
+the row layout, the delimiter ids, the render rules, the head formula and the response shape.
 
 **Host.** Builds the rows, runs the calls, converts the hidden rows at the readout positions to float64,
 applies the head (`head/head.safetensors`, `head/kev_head.json`) and the per-question softmax in float64,
 rounds p to fp32 once, and writes the response. `conversion/kev/decide.py` is the Python reference;
-[`apps/Kev`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/Kev) is the Swift one.
+[`apps/Kev`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/Kev) is the Swift one. With `shared: true` the state's whole calls run once. A
+prepared state (`prepare(state:)`, then `decide(prepared:questionsJSON:)`) runs the same calls in two steps,
+so questions that arrive later skip the state.
 
 No engine is involved: the hosts drive the low-level runtime (`AIModel` + `loadFunction`, four zeroed
 states per row). No runtime patch.
@@ -148,10 +152,10 @@ states per row). No runtime patch.
 
 The bar, fixed before any graph ran: the argmax equal to the oracle's on every question whose oracle
 top-2 margin is above 0.02 (near-ties listed apart), max |Δp| ≤ 0.02 over every option of every
-question, the mean over rows of each row's mean |Δp| ≤ 0.002, and every process re-running its first row
+question, the mean over rows of each row's mean |Δp| ≤ 0.002, and every process re-running its opening row
 bit for bit. The Python gates load the AOT `.aimodelc` (`coreai-build compile … --platform macOS
 --preferred-compute gpu --architecture h16c --expect-frequent-reshapes`) with
-`SpecializationOptions.default()`; their times are on a GPU shared with other work.
+`SpecializationOptions.default()`.
 
 ### Before any graph: the merge and fp32 torch
 
@@ -161,11 +165,12 @@ bit for bit. The Python gates load the AOT `.aimodelc` (`coreai-build compile
   merged tensors bit for bit (the adapter moves them by 2.8–3.6 %). The base alone moves p by up to
   0.789 (argmax 25/53), the base's Gated DeltaNet projections put back by up to 0.282 (46/53): the
   check can fail ([`gate-kev-0.8b-merge.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-merge.json)).
-- The decoder module in fp32 on the CPU, driven like the graph, read through the author's head: 434/434
-  argmax (the 14 near-ties included), max |Δp| 3.5e-6, lowest per-position hidden cosine 0.99999999999
-  on the 21 rows the oracle keeps. It equals the overlay's plain text decoder bit for bit (3 rows), and
-  S = 32, 64 and 128 give the bits of S = 16
-  ([`gate-kev-0.8b-torch-parity.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-torch-parity.json)).
+- The decoder module in fp32 on the CPU, driven like the graph with the recurrence unrolled, read through
+  the author's head: 434/434 argmax (the 14 near-ties included), max |Δp| 3.5e-6, lowest per-position
+  hidden cosine 0.99999999999 on the 21 rows the oracle keeps. It equals the overlay's plain text decoder
+  bit for bit (3 rows) ([`gate-kev-0.8b-torch-parity.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-torch-parity.json)).
+- The Metal kernel has no torch values: its `torch_defn` returns zeros of the right shapes. The GPU gate
+  below is its parity check.
 
 ### The graph alone on the Mac GPU
 
@@ -173,39 +178,60 @@ The oracle's ids in, the fp16 hidden rows read through the author's fp32 head:
 
 | set | questions | argmax (margin > 0.02) | near-ties agreeing | max \|Δp\| | mean of row means | bar |
 |---|---:|---:|---:|---:|---:|---|
-| **fixture** | 434 | **420/420** | **12/14** | **0.0117** | **0.00097** | **PASS** |
-| held out | 130 | 125/125 | 5/5 | 0.0057 | 0.00081 | PASS |
+| **fixture** | 434 | **420/420** | **13/14** | **0.0124** | **0.00096** | **PASS** |
+| held out | 130 | 125/125 | 5/5 | 0.0058 | 0.00083 | PASS |
 
-The two near-ties that flip are `tv4_023` (oracle margin 0.0020) and a SemIf item (0.0032), each by at
-most 0.0027. The worst row is a SemIf item (0.0117). Every value is finite and every process's re-run is
-bit-equal. Two red arms move p on the same graph: a state swapped for another record's moves 4 of 5
-argmaxes (max |Δp| 0.905); a grammatical "not" in five yes / no questions moves 2 of 5 (0.759); fp32
-torch moves the same rows by the same amounts. Transcripts:
-[`gate-kev-0.8b-readout-fp16_pf16.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-readout-fp16_pf16.json),
-[`gate-kev-0.8b-heldout.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-heldout.json).
+The near-tie that flips is a SemIf item with an oracle margin of 0.0032; it moves by 0.0026. The worst
+row is a SemIf item (0.0124). Every value is finite and every process's re-run is bit-equal. Two red arms
+move p on the same graph: a state swapped for another record's moves 4 of 5 argmaxes (max |Δp| 0.905); a
+grammatical "not" in five yes / no questions moves 2 of 5 (0.759). Transcript:
+[`gate-kev-0.8b-readout.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-readout.json).
 
 ### From the request: Python and Swift
 
 - `conversion/kev/host.py`, without the author's package, rebuilds every oracle row from the raw
-  request (ids, `<|fim_suffix|>` / `<|box_end|>` indices, keys): 564/564 rows with two tokenizer
-  implementations; fed the oracle's p, its response bodies equal the author's on every record but
-  `latency_ms`. `decide.py` on the AOT graph: hidden rows bit-equal to the gate's on 564/564 rows; the
-  shared prefix equal to the direct run on every row, the 20 multi-question records taking 662 calls
-  instead of 1,830 ([`gate-kev-0.8b-host.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-host.json)).
-- The Swift host ([`apps/Kev`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/Kev), Release): ids from the raw request 564/564; on the same
-  AOT asset every row's hidden state, every p and every answer set's `json.dumps` equal the Python
-  reference's bit for bit and byte for byte, and the bar is the gate's (fixture 0.0117 / 0.00097, held
-  out 0.0057 / 0.00081). Its text half equals CPython on 2,787 rendered values, 599 requests (21 refused
-  with the same messages), 203,635 doubles (`str`, `repr`, `round`) and 1,093 `json.dumps` texts. One
-  word changed in one question turns its record red.
+  request (ids, `<|fim_suffix|>` / `<|box_end|>` indices, keys): 434/434 and 130/130 rows with two
+  tokenizer implementations. `decide.py` on the AOT graph: hidden rows bit-equal to the gate's on 434/434
+  and 130/130 rows ([`gate-kev-0.8b-host.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-host.json)).
+- The Swift host ([`apps/Kev`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/Kev), Release): ids from the raw request 434/434 and 130/130;
+  on the same AOT asset every row's hidden state and every p equal the Python reference's bit for bit
+  (434/434), every answer set's `json.dumps` and usage byte for byte (384/384), and the bar is the gate's
+  (fixture 0.0124 / 0.00096). With the shared prefix every row's hidden state and p equal the direct run's
+  (434/434). Its text half equals CPython on 2,787 rendered values, 599 requests (21 refused with the
+  same messages), 203,635 doubles (`str`, `repr`, `round`) and 1,093 `json.dumps` texts. One word changed
+  in one question turns its record red.
 - JIT: the `.aimodel` specialized by the Swift runtime (GPU preferred, `expectFrequentReshapes`) equals
-  the AOT asset bit for bit on 25 rows of 10 records; the specialization took 5.4 s and added
-  2,510,934,016 bytes to the runtime's cache.
-- Rows near the limit (a diagnosis, not a gate set): rows of 4,072, 4,030, 3,039 and 2,997 tokens built
-  from the fixture's long states keep every argmax, max |Δp| 0.0023, and the hidden state's cosine to
-  the oracle's is at least 0.99994 at every position
+  the AOT asset bit for bit on 25 rows of 10 records. The specialization took 3.34 s (4.05 s with the
+  tokenizer and the head); with the runtime's cache warm a load took 0.69 s
   ([`gate-kev-0.8b-swift.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-swift.json)).
 
+### Time per decision on the Mac
+
+Swift, Release CLI, the AOT asset, in one machine-wide GPU lock window (2026-10-04 12:13–12:27 JST). Two
+processes per graph counted only when no other GPU job ran and the one-minute load average was at most 12;
+each made 10 decisions per item after one warm-up, and the table gives the median (p10–p90) over the 20.
+The earlier upload's graph (below, under Bundle) ran in the same window:
+
+| request | tokens | calls | ms | p10–p90 | earlier upload's graph, ms |
+|---|---:|---:|---:|---|---:|
+| one question | 94 | 1 | 29.9 | 29.6–30.1 | 100.1 |
+| one question | 380 | 3 | 89.2 | 88.7–90.4 | 404.4 |
+| one question | 1,518 | 12 | 357.0 | 355.4–359.3 | 1,600.1 |
+| one question | 1,802 | 15 | 447.9 | 446.6–449.8 | 1,897.6 |
+| five questions on one 137-token state, each row from zero | 240 | 10 | 296.6 | 294.6–298.5 | 845.7 |
+| the same five questions, the state run once (shared) | 240 | 6 | 186.4 | 185.4–187.7 | 327.7 |
+| eight questions on that state, each row from zero | 320 | 16 | 474.2 | 472.2–475.9 | 1,377.0 |
+| the same eight questions, shared | 320 | 9 | 278.8 | 277.2–281.7 | 461.0 |
+| four questions on one 1,477-token state, each row from zero | 1,622 | 49 | 1,455.9 | 1,453.7–1,460.4 | 6,366.1 |
+| the same four questions, shared | 1,622 | 16 | 485.4 | 483.8–489.4 | 1,749.0 |
+| a follow-up question on a prepared 137-token state | — | — | 30.3 | — | 34.1 |
+| a follow-up question on a prepared 1,477-token state | — | — | 31.5 | — | 51.5 |
+
+Preparing the two states took 34.1–34.7 ms and 335.9–339.6 ms (the two processes' medians). In round 11's
+lock window (09:43–10:45 JST) the `.aimodel` specialized by Swift, the shipped path, took 29.7 ms for the
+94-token question and 186.5 ms for the five questions shared, against 30.0 and 189.1 ms for the AOT asset;
+the Python reference took 33.4 and 207.1 ms. Transcript: [`gate-kev-0.8b-timing-mac.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-timing-mac.json).
+
 ### iPhone 18 Pro (iOS 27.0 24A437, Core AI arch h19p, 2026-10-04)
 
 Through [`apps/KevGate`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/KevGate), a headless gate app on the Swift host, Release, without the
@@ -213,73 +239,95 @@ increased-memory-limit entitlement (3,529 MB available at launch). The phone was
 battery at 80 % and charging; every bench item started at thermal state nominal. Every p the app wrote
 was re-scored on the Mac from its bit patterns.
 
-| load | cache | wall s | `AIModel` s | `main` s | runtime cache added | peak footprint |
-|---|---|---:|---:|---:|---:|---:|
-| `.aimodel` specialized on the phone (GPU preferred, `expectFrequentReshapes`) | cold | 6.65 | 5.87 | 0.42 | 2,505,661,047 B | 316 MB |
-| the same | warm | 0.61 | 0.00 | 0.25 | 0 | 216 MB |
-| h19p AOT asset (2,505,674,625 B), `.default` | cold | 4.64 | 3.07 | 1.22 | 2,505,681,950 B | 223 MB |
-
-| set | asset | questions | argmax (margin > 0.02) | near-ties agreeing | max \|Δp\| | mean | bar |
-|---|---|---:|---:|---:|---:|---:|---|
-| fixture | specialized on the phone | 434 | 420/420 | 13/14 | 0.0102 | 0.00098 | PASS |
-| held out | specialized on the phone | 130 | 125/125 | 5/5 | 0.0049 | 0.00083 | PASS |
-| fixture 60 + held out 30 | h19p AOT | 90 | 80/80 | 10/10 | 0.0071 | 0.00078 | PASS |
-
-On the phone the AOT asset and the device specialization agree bit for bit (90/90 rows). No hidden row
-is bit-equal to the Mac's (M4 Max, AOT h16c): their p differ by at most 0.0027, and one near-tie
-(`tv4_023`, oracle margin 0.0020) lands on the oracle's side on the phone and not on the Mac. The shared prefix equals the direct run on
-all 20 multi-question records. The first decision after the cold specialization took 3.58 s; in the bench
-below a 16-token call's median was 25.9–26.6 ms on every item, whatever the row length.
-
-Bench (60 s of rest before each item, one warm-up, then the decisions back to back; `latency_ms` = state
-resets, graph calls and the head):
-
-| request | tokens | calls | median ms (specialized on the phone) | runs | h19p AOT median ms |
-|---|---:|---:|---:|---|---:|
-| one question (`tv4_000`) | 94 | 6 | 156.6 | 5: 156.3–157.3 | 157.2 |
-| one question (`own_j03` q0) | 380 | 24 | 629.2 | 5: 628.9–630.1 | 627.8 |
-| one question (`own_L02` q0) | 1,518 | 95 | 2,500.0 | 5: 2,499.4–2,502.5 | — |
-| one question (`own_L01` q2) | 1,802 | 113 | 2,975.9 | 5: 2,974.1–2,981.4 | — |
-| 5 questions, 137-token state (`own_m01`), direct | 240 | 51 | 1,330.5 | 5: 1,329.1–1,333.5 | — |
-| the same, shared prefix | 240 | 19 | 507.8 | 5: 507.1–508.8 | 508.2 |
-| 8 questions (`own_m01`), direct | 320 | 83 | 2,167.5 | 3: 2,163.8–2,182.6 | — |
-| the same, shared prefix | 320 | 27 | 722.4 | 3: 721.2–732.9 | — |
-| 4 questions, 1,477-token state (`own_L02`), direct | 1,622 | 380 | 10,026.3 | 2: 10,018.3–10,034.3 | — |
-| the same, shared prefix | 1,622 | 104 | 2,761.9 | 2: 2,752.4–2,771.5 | — |
-
-The last item ran 38 s back to back and ended at thermal state fair. Kev-4B on the phone: see the
-[Kev-4B card](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/README.md). Transcript: [`gate-kev-0.8b-iphone.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-iphone.json).
-
-### Time per decision on the Mac
-
-Swift, Release CLI, AOT asset, in one machine-wide GPU lock window (2026-10-04 00:11–00:33 JST) with no
-other GPU job and a CPU load of 4–10. A B A B: two processes, one warm-up, the median over 20 decisions
-(p10–p90):
+- **Load.** The `.aimodel` specialized on the phone (GPU preferred, `expectFrequentReshapes`) took 4.26 s cold,
+  3.60 s of it in `AIModel(contentsOf:)`, with +2,502,056,463 B of runtime cache and a peak footprint of 233 MB
+  (round 13, an earlier KevGate build). With the cache warm, the gate run below loaded it in 1.50 s.
+- **Gate** (run `20261004-140048`, KevGate built on round 15's host):
 
-| request | tokens | calls | ms | p10–p90 |
-|---|---:|---:|---:|---|
-| one question (`tv4_000`) | 94 | 6 | 112.1 | 103.2–166.5 |
-| one question (`own_j03` q0) | 380 | 24 | 472.0 | 413.9–583.1 |
-| one question (`own_L02` q0) | 1,518 | 95 | 1,595.4 | 1,567.5–1,660.8 |
-| one question (`own_L01` q2) | 1,802 | 113 | 1,915.0 | 1,875.8–2,117.4 |
-| 5 questions (`own_m01`), direct / shared prefix | 240 | 51 / 19 | 879.9 / 342.3 | 835.3–917.1 / 323.0–345.4 |
-| 8 questions (`own_m01`), direct / shared prefix | 320 | 83 / 27 | 1,454.7 / 484.5 | 1,384.9–1,555.2 / 454.9–496.6 |
-| 4 questions (`own_L02`), direct / shared prefix | 1,622 | 380 / 104 | 6,415.3 / 1,785.0 | 6,266.8–6,733.8 / 1,733.2–1,884.7 |
-
-A 16-token call took 18.6 and 17.2 ms (the two processes' medians). Loading the AOT asset with the
-runtime's cache warm: 0.61 / 0.27 s at each process's first load, 0.43 / 0.24 s at its second. The Python
-reference in the same window: 104.0 ms for `tv4_000` and 347.5 ms for five questions with the shared
-prefix. Transcript: [`gate-kev-0.8b-timing-mac.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-timing-mac.json).
+| set | questions | argmax (margin > 0.02) | near-ties agreeing | max \|Δp\| | mean of row means | bar |
+|---|---:|---:|---:|---:|---:|---|
+| fixture | 434 | 420/420 | 13/14 | 0.0110 | 0.00097 | PASS |
+| held out | 130 | 125/125 | 5/5 | 0.0059 | 0.00083 | PASS |
+
+  The shared prefix equals the direct run on all 20 multi-question records, and the re-run of the opening
+  record is bit-equal. No hidden row equals the Mac's (a different GPU): their p differ by at most 0.0026,
+  with every argmax equal (434/434). The fixture pass peaked at a 389 MB footprint.
+- **Bench** (60 s of rest before each item, one warm-up, then 5 decisions; `latency_ms` = state resets,
+  graph calls and the head). The earlier upload's graph ran in the same session:
+
+| request | tokens | ms | earlier upload's graph, ms |
+|---|---:|---:|---:|
+| one question | 94 | 37.8 | 156.7 |
+| one question | 380 | 110.6 | 627.9 |
+| one question | 1,518 | 444.1 | 2,500.9 |
+| one question | 1,802 | 559.1 | 2,977.8 |
+| five questions on one 137-token state, each row from zero / shared | 240 | 361.9 / 223.2 | 1,331.4 / 507.4 |
+| eight questions on that state, each row from zero / shared | 320 | 583.2 / 333.4 | 2,167.4 / 723.3 |
+| four questions on one 1,477-token state, each row from zero / shared | 1,622 | 1,822.2 / 613.9 | 10,061.6 / 2,769.8 |
+| the five questions on a prepared state (the prepare: 38.8 / 211.6 ms) | 240 | 183.4 | 295.5 |
+| the four questions on a prepared state (the prepare: 415.3 / 2,451.4 ms) | 1,622 | 196.5 | 342.4 |
+
+  An earlier session on the same phone, with an earlier KevGate build, gave 36.5 ms for the 94-token question
+  against 157.1 ms for the earlier upload's graph. Transcript: [`gate-kev-0.8b-iphone.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-iphone.json).
+
+Kev-4B on the phone: see the [Kev-4B card](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/README.md).
+
+## Forms measured
+
+Every form below was exported, compiled and gated the same way. The times come from different windows, so
+compare forms only within one window. Full records: [`gate-kev-0.8b-forms.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-forms.json).
+
+Mac (Swift Release CLI, the 94-token question, median of the counted processes):
+
+| form | 94-token decision, ms | window | fixture max \|Δp\| | note |
+|---|---:|---|---:|---|
+| Gated DeltaNet unrolled, S = 16 (the earlier upload) | 100.1 | round 15 | 0.0117 | |
+| unrolled, S = 32 | 139.0 | round 11 | 0.0118 | |
+| unrolled, S = 64 | — | — | 0.0100 | not timed in a lock window (68.06 ms per call on a shared GPU) |
+| in-graph chunk scan, S = 16 | — | — | 0.0135 | not timed in a lock window (10.90 ms per call on a shared GPU) |
+| in-graph chunk scan, S = 32 | — | — | — | fails: non-near-tie argmax 112/127, max \|Δp\| 0.125, non-finite rows; breaks in fp32 torch too |
+| Metal kernel, S = 16 / 32 / 64 | 63.6 / 42.7 / 41.5 | round 11 | 0.0115 / 0.0134 / 0.0115 | |
+| **Metal kernel, S = 128 (this release)** | **29.9** | round 15 | **0.0124** | |
+| Metal kernel, query length 2..512, host calls of at most 512 ids | 25.2 | round 15 | 0.0112 | not shipped: memory grows (below) |
+| the same graph, host calls of at most 256 / 128 ids | 25.1 / 25.0 | round 15 | 0.0060 / 0.0060 (130 rows) | not shipped |
+| int8, every linear (unrolled S = 16) | — | — | 0.0431 | fails the bar (Precision) |
+| any form on the Neural Engine | — | — | — | not run: the recurrence's fp32 state is not an ANE element type ([`qwen3.5-static-ane.md`](https://github.com/john-rocky/coreai-model-zoo/blob/main/knowledge/qwen3.5-static-ane.md)) |
+
+iPhone 18 Pro (KevGate bench, the 94-token question):
+
+| form | ms | session | fixture max \|Δp\| |
+|---|---:|---|---:|
+| unrolled, S = 16 (the earlier upload) | 156.7 | round 13, stage C | 0.0102 |
+| Metal kernel, S = 16 / 32 / 64 | 92.4 / 58.5 / 49.3 | round 13, stage A | 0.0110 / 0.0133 / 0.0098 |
+| **Metal kernel, S = 128 (this release)** | **37.8** | round 13, stage C | **0.0110** |
+| Metal kernel, query length 2..512, host calls of at most 512 ids | 32.6 | round 13, stage C | gate stopped by memory (below) |
+
+**Several short questions on one state.** For these, the S = 64 graph takes less time than S = 128. In
+round 11's Mac window, five questions on one 137-token state, shared, took 145.0 ms with S = 64 and
+189.1 ms with S = 128; one 94-token question took 41.5 and 30.0 ms. On the iPhone (round 13, stage A)
+the same pairs took 180.3 against 224.4 ms and 49.3 against 36.5 ms. The S = 64 bundle is not in this
+release: it would need a new export and a new gate.
+
+**Why the dynamic-length graph does not ship.** The graph whose query length is dynamic (2..512) passed the
+gate on the Mac and decided the 94-token question in 25.2 ms. A process that keeps changing its call length
+keeps growing, though. Cycling four call lengths (128, 256, 384, 512) three times added 467.1, 166.9 and
+168.5 MB per lap in Swift and 460.7, 169.5 and 169.6 MB in Python. Alternating between two lengths did not
+grow after the opening calls; bringing in a third added 19.1 MB, then 1.3 MB. On the iPhone the gate run
+reached 3,497 MB with 43 MB available after 439 calls and stopped writing. A 2-second sleep gave back part of
+the memory and loading the function again gave back none; creating the `AIModel` again gave it back. Records:
+[`gate-kev-0.8b-forms.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-forms.json); notes:
+[`knowledge/kev-port.md`](https://github.com/john-rocky/coreai-model-zoo/blob/main/knowledge/kev-port.md).
 
 ## Precision
 
-**int8 was measured and is not shipped.** int8 per block of 32 over every decoder linear
+**int8 was measured and is not shipped.** It was measured on the unrolled S = 16 graph, the earlier upload;
+the Metal-kernel graphs were not measured in int8. int8 per block of 32 over every decoder linear
 (`symmetric_with_clipping`, weights only; the embedding table, the Gated DeltaNet conv1d and every norm
 fp16) fails the bar:
 
-| decoder | `main.mlirb` bytes | fixture max \|Δp\| / mean | held out max \|Δp\| / mean | bar |
+| decoder (unrolled S = 16) | `main.mlirb` bytes | fixture max \|Δp\| / mean | held out max \|Δp\| / mean | bar |
 |---|---:|---:|---:|---|
-| fp16 (shipped) | 1,506,481,909 | 0.0117 / 0.00097 | 0.0057 / 0.00081 | PASS |
+| fp16 | 1,506,481,909 | 0.0117 / 0.00097 | 0.0057 / 0.00081 | PASS |
 | int8lin: every linear int8 | 1,040,063,350 | 0.0431 / 0.00326 | 0.0273 / 0.00294 | FAIL |
 | int8mix: int8 except layers 0–11 | 1,273,276,920 | 0.0089 / 0.00103 | 0.0043 / 0.00082 | PASS |
 
@@ -291,21 +339,15 @@ fp16, 45.5 % of the weights, 0.0141). Layers 0–11 kept fp16, exactly half and
 give 0.0046 on the bisect rows; that is int8mix. It keeps 85 % of fp16's bytes, and it was chosen on the
 fixture, so it does not ship ([`gate-kev-0.8b-int8.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-int8.json)).
 
-**What the compiled asset holds.** `--expect-frequent-reshapes`, which this dynamic-length graph needs to
-run without re-specializing, adds an fp16 copy of every linear weight to the compiled asset: the iPhone
-h19p asset is 2,505,674,625 bytes with it and 1,506,400,620 without. On a toy (an embedding and two
-linears) compiled with a fixed input length, the int8 version's `resources.bin` is 12,582,936 bytes on
-macOS and iOS, with and without the flag, the bytes of the fp16 toy compiled without it, against an int8
-`main.mlirb` of 10,623,080 bytes. An int8 `.aimodel` downloads smaller; the compiled asset is not
-smaller by the same amount
+**What the compiled asset holds.** `--expect-frequent-reshapes` adds an fp16 copy of every linear weight
+to the compiled asset: for the unrolled graph the iPhone h19p asset is 2,505,674,625 bytes with it and
+1,506,400,620 without, and this release's Mac h16c asset is 2,502,047,194 bytes against a 1,505,385,733-byte
+`main.mlirb`. On a toy (an embedding and two linears) compiled with a fixed input length, the int8
+version's `resources.bin` is 12,582,936 bytes on macOS and iOS, with and without the flag, the bytes of
+the fp16 toy compiled without it, against an int8 `main.mlirb` of 10,623,080 bytes. An int8 `.aimodel`
+downloads smaller; the compiled asset is not smaller by the same amount
 ([`../kev-4b/gate-kev-4b-int8.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/gate-kev-4b-int8.json)).
 
-**Chunk width.** S = 16, 32, 64 and 128 all pass the bar (max |Δp| 0.0117, 0.0118, 0.0100, and 0.0060 on
-130 rows). A call's time grows with S: the Gated DeltaNet layers run their S steps one after another
-inside the call. In one interleaved run over the fixture's 423 warm rows, S = 16 took 72.8 s of graph
-calls and S = 32 74.4 s, and S = 16 pads 5.1 % of the fixture against 10.4 %. S = 16 ships
-([`gate-kev-0.8b-readout-s.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-readout-s.json)).
-
 ## ⬇️ Bundle
 
 [mlboydaisuke/Kev-0.8B-CoreAI](https://huggingface.co/mlboydaisuke/Kev-0.8B-CoreAI), one folder under
@@ -313,29 +355,32 @@ calls and S = 32 74.4 s, and S = 16 pads 5.1 % of the fixture against 10.4 %. S
 
 | file | what | bytes | sha256 |
 |---|---|---:|---|
-| `kev_0_8b_decode_fp16_pf16.aimodel/main.mlirb` | the decoder, fp16 | 1,506,481,909 | `4131d028…72682d` |
-| `metadata.json` | `kind: decision-backbone`, the readout contract, the gate | 15,236 | `97b70e43…22b65b2` |
+| `kev_0_8b_decode_fp16_metal_pf128.aimodel/main.mlirb` | the decoder, fp16 | 1,505,385,733 | `19d5a480…2c3f684e` |
+| `metadata.json` | `kind: decision-backbone`, the readout contract, the gate | 12,590 | `6e50e41a…47282227` |
 | `tokenizer/tokenizer.json` | the base model's, verbatim | 12,807,196 | `fe000e3e…d50d2927` |
 | `tokenizer/tokenizer_config.json` | the base model's, verbatim | 16,712 | `e611fbcc…c47885de` |
-| `head/head.safetensors` | the pointer head: q / k weight `[256, 1024]` and bias, fp32 | 2,099,576 | `12038b02…700893a42` |
-| `head/kev_head.json` | head size, scale, temperature, delimiter ids, provenance | 2,046 | `73d51070…99d5c9d7c` |
+| `head/head.safetensors` | the pointer head: q / k weight `[256, 1024]` and bias, fp32 | 2,099,576 | `12038b02…00893a42` |
+| `head/kev_head.json` | head size, scale, temperature, delimiter ids, provenance | 2,046 | `73d51070…9d5c9d7c` |
 
 The repository root carries the base model's `config.json`, the adapter's `adapter_config.json` and the
-Apache-2.0 `LICENSE`, verbatim. `SHA256SUMS` lists every file. Uploaded 2026-10-04, Hub revision
-[`7793533`](https://huggingface.co/mlboydaisuke/Kev-0.8B-CoreAI/tree/7793533783de7c8f916c1cba1144a561fb70b321); every file verified after the upload
-(3 LFS files by the Hub's sha256 and size, 9 small files by download, `SHA256SUMS` byte-equal).
+Apache-2.0 `LICENSE`, verbatim. `SHA256SUMS` lists every file. The Hub revision of this bundle is recorded
+here after the upload.
+
+**The earlier upload.** Hub revision
+[`7793533`](https://huggingface.co/mlboydaisuke/Kev-0.8B-CoreAI/tree/7793533783de7c8f916c1cba1144a561fb70b321)
+held another graph of this port, `kev_0_8b_decode_fp16_pf16`: the Gated DeltaNet recurrence unrolled step by
+step, 16 tokens per call. It passed the same gate (fixture max |Δp| 0.0117, held out 0.0057). In the Mac
+window above it took 100.1 ms for the 94-token question against 29.9 ms for this release, and on the iPhone
+156.7 ms against 37.8 ms.
 
 No AOT asset ships: the Swift runtime specializes the `.aimodel` correctly on the Mac and on the phone (the
-JIT rows above). To compile one anyway:
+JIT rows above). To compile one for the Mac anyway:
 
 ```bash
-xcrun coreai-build compile kev_0_8b_decode_fp16_pf16.aimodel --output aot --preferred-compute gpu \
-    --platform macOS --architecture h16c --expect-frequent-reshapes     # iPhone 18 Pro: --platform iOS --architecture h19p
+xcrun coreai-build compile kev_0_8b_decode_fp16_metal_pf128.aimodel --output aot --preferred-compute gpu \
+    --platform macOS --architecture h16c --expect-frequent-reshapes
 ```
 
-`--expect-frequent-reshapes` is required (without it the runtime specializes again for every new row
-length) and makes the asset 2.5 GB. An iOS asset must not be loaded on a Mac.
-
 ## Use it
 
 Swift, with the [`Kev`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/Kev) package (macOS 27 / iOS 27; the system CoreAI framework,
@@ -344,18 +389,22 @@ Accelerate and swift-transformers' tokenizer), on a download of the repository:
 ```swift
 import Kev
 
-let bundle = URL(filePath: "Kev-0.8B-CoreAI/gpu-pipelined/kev_0_8b_decode_fp16_pf16")
+let bundle = URL(filePath: "Kev-0.8B-CoreAI/gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128")
 let kev = try await KevDecider(bundle: bundle)              // asset: nil = the .aimodel, specialized here (GPU, frequent reshapes)
-let response = try await kev.decide(requestJSON: requestData, shared: true)   // shared: the state's chunks run once
+let response = try await kev.decide(requestJSON: requestData, shared: true)   // shared: the state's whole calls run once
 print(PythonFormat.dumps(response, asciiOnly: false))
-// {"model": "kev_0_8b_decode_fp16_pf16", "answers": {"team": {"type": "choice", "choice": "billing", "confidence": …,
+// {"model": "kev_0_8b_decode_fp16_metal_pf128", "answers": {"team": {"type": "choice", "choice": "billing", "confidence": …,
 //  "probabilities": {…}}, "urgent": {"type": "noul", "noul": …}}, "usage": {"input_tokens": …, "output_tokens": …}, "latency_ms": …}
+
+// questions that arrive later, on the same state
+let prepared = try await kev.prepare(state: try JSONParser.parse(stateData))       // the state's whole calls, once
+let later = try await kev.decide(prepared: prepared, questionsJSON: questionsData)  // only the questions' rows run
 ```
 
 The same from the Mac CLI (`swift build -c release --package-path apps/Kev` builds `kev`):
 
 ```bash
-kev run --bundle Kev-0.8B-CoreAI/gpu-pipelined/kev_0_8b_decode_fp16_pf16 --asset jit \
+kev run --bundle Kev-0.8B-CoreAI/gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128 --asset jit \
     --request req.json --shared --out resp.json
 ```
 
@@ -363,7 +412,7 @@ A request, in the SystemOne-compatible request shape:
 
 ```json
 {"model": "kev-0.8b",
- "state": "I was charged twice for order 1182. Please refund one of the charges.",
+ "state": "I was charged twice for my last order. Please refund one of the charges.",
  "questions": {
    "team": {"type": "choice", "instructions": "Which team should handle this?",
             "criteria": {"billing": "Charges and refunds", "shipping": "Deliveries", "returns": "Exchanges"}},
@@ -386,7 +435,11 @@ with every flag, are in [`conversion/kev/README.md`](https://github.com/john-roc
 (cd $ZOO_WORK_ROOT/_kev/kev-src && python scripts/merge_lora_checkpoint.py \
     --lora jaredpalmer/kev-0.8b@788ddbdd65715bb03a56788c822f6c632c9a551d --out $ZOO_WORK_ROOT/_kev/merged/kev-0.8b-v1.0)
 # the decoder (--aot adds the h16c .aimodelc the Python gates load)
-python conversion/kev/export_decoder.py fp16 --prefill-chunk 16 --aot
+python conversion/kev/export_decoder.py fp16 --gdn-scan metal --prefill-chunk 128 --aot
+# the published bundle: the gated .aimodel bytes and their metadata, without exporting again
+python conversion/kev/export_decoder.py fp16 --gdn-scan metal --prefill-chunk 128 --metadata-only \
+    --from-aimodel <bundles>/kev_0_8b_decode_fp16_metal_pf128/kev_0_8b_decode_fp16_metal_pf128.aimodel \
+    --bundle-dir <ship>/kev_0_8b_decode_fp16_metal_pf128
 ```
 
 Recipe: [`recipe.toml`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/recipe.toml). Port notes: [`knowledge/kev-port.md`](https://github.com/john-rocky/coreai-model-zoo/blob/main/knowledge/kev-port.md).
@@ -425,5 +478,12 @@ and numbers.
 From the [author's card](https://huggingface.co/jaredpalmer/kev-0.8b): English only; text generation,
 chat, tool-call routing and fully automated consequential decisions about people are out of scope; the
 temperature was fitted on the author's development rows (the card explains how to measure and refit it on
-one's own data). This port adds one limit: a row holds at most 4,080 tokens, while the author's server
-accepts states up to 65,536.
+one's own data). This port adds:
+
+- A row holds at most 3,968 tokens, while the author's server accepts states up to 65,536.
+- A process's opening call is slower. On the Mac the opening 128-token call of each Python gate process
+  took 211–1,188 ms, against a median of 29.9 ms for its later calls. On the phone the opening decision
+  took 66.9 ms, against 37.8 ms in the bench of the same build.
+- If you export this decoder with a dynamic query length yourself, keep the call lengths to one or two
+  values. A process that cycles more lengths grows its memory until the `AIModel` is created again (measured
+  on the Mac and the iPhone; which runtime layer keeps the memory is not isolated).

SHA256SUMS

@@ -1,12 +1,12 @@
 50cbab8a892c5f2993b8c7351a99182507472def3b1374558308605d99b86b32  LICENSE
-d21eecda6b62741ad3eadb39cf566e76194ef6e7e76a2c0d2248e3d125072773  README.md
+3d86d2c9190033fda5f58eaa257b93069a0ee5a955c4b9bfbf56e7ac96e34df7  README.md
 748acb2cda88454cb1ba69d745ba336f3fcb5486eac349e90960c8b8d8d3e854  adapter_config.json
 b90b86f35c8e6925ef74ee04d0e758f0a845c83a42089ad82bbaa948de9b4204  config.json
-12038b02ecf7ece3ffb008b22ae97988293e521a86e7aadc8036c9c700893a42  gpu-pipelined/kev_0_8b_decode_fp16_pf16/head/head.safetensors
-73d510708790422d3c84bfc915e311e536d527bc032f0a07394e76999d5c9d7c  gpu-pipelined/kev_0_8b_decode_fp16_pf16/head/kev_head.json
-5e53b0fa31411ab780585a1183e170407d3fcd5aac0b2053ca0ef1e37df5e87c  gpu-pipelined/kev_0_8b_decode_fp16_pf16/kev_0_8b_decode_fp16_pf16.aimodel/main.hash
-4131d0284899787bab6cae6ca62a706886478b4e25c4d8cb971c5ba2ad72682d  gpu-pipelined/kev_0_8b_decode_fp16_pf16/kev_0_8b_decode_fp16_pf16.aimodel/main.mlirb
-c751b845f3c70312dab24edf1084f98321b0b8c6e89fb0ea4a8263ab51e8185c  gpu-pipelined/kev_0_8b_decode_fp16_pf16/kev_0_8b_decode_fp16_pf16.aimodel/metadata.json
-97b70e4393d549aada62381be4417a0ddb5ebf6491a5e824cb58ecb5f22b65b2  gpu-pipelined/kev_0_8b_decode_fp16_pf16/metadata.json
-fe000e3ed39ed12b8d2481d527d44f93c65d37e87645d2dcc80d1bf9d50d2927  gpu-pipelined/kev_0_8b_decode_fp16_pf16/tokenizer/tokenizer.json
-e611fbccc7c29ef3b1cafb1cb7ea548d189968632901d678fd62be68c47885de  gpu-pipelined/kev_0_8b_decode_fp16_pf16/tokenizer/tokenizer_config.json
+12038b02ecf7ece3ffb008b22ae97988293e521a86e7aadc8036c9c700893a42  gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/head/head.safetensors
+73d510708790422d3c84bfc915e311e536d527bc032f0a07394e76999d5c9d7c  gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/head/kev_head.json
+7ad0c96ad0ce3cef5315f9504b15b5dacdd8b958b2848f1600e8edbd8b762bfc  gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/kev_0_8b_decode_fp16_metal_pf128.aimodel/main.hash
+19d5a4809e348c7cdd314c17cf18646170e2af2ef62f5a32beb8948b2c3f684e  gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/kev_0_8b_decode_fp16_metal_pf128.aimodel/main.mlirb
+33d232378eb37701453e8974f12141c95bedcf856951feffa29854b24b127a8f  gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/kev_0_8b_decode_fp16_metal_pf128.aimodel/metadata.json
+6e50e41a92eebe440029974a403441cd960d2f6ef4ff00828bb2aa8f47282227  gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/metadata.json
+fe000e3ed39ed12b8d2481d527d44f93c65d37e87645d2dcc80d1bf9d50d2927  gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/tokenizer/tokenizer.json
+e611fbccc7c29ef3b1cafb1cb7ea548d189968632901d678fd62be68c47885de  gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/tokenizer/tokenizer_config.json

gpu-pipelined/kev_0_8b_decode_fp16_metal_pf128/kev_0_8b_decode_fp16_metal_pf128.aimodel/main.hash

@@ -0,0 +1 @@
+դ��4�|�1L�dap�.�/Z2����,?hN

gpu-pipelined/{kev_0_8b_decode_fp16_pf16/kev_0_8b_decode_fp16_pf16.aimodel → kev_0_8b_decode_fp16_metal_pf128/kev_0_8b_decode_fp16_metal_pf128.aimodel}/main.mlirb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4131d0284899787bab6cae6ca62a706886478b4e25c4d8cb971c5ba2ad72682d
-size 1506481909
+oid sha256:19d5a4809e348c7cdd314c17cf18646170e2af2ef62f5a32beb8948b2c3f684e
+size 1505385733

gpu-pipelined/{kev_0_8b_decode_fp16_pf16/kev_0_8b_decode_fp16_pf16.aimodel → kev_0_8b_decode_fp16_metal_pf128/kev_0_8b_decode_fp16_metal_pf128.aimodel}/metadata.json

@@ -1,5 +1,5 @@
 {
   "assetVersion" : "2.0",
   "producer" : "coreai-core 1.0.0b2",
-  "creationDate" : "20261003T073819Z"
+  "creationDate" : "20261003T234731Z"
 }

gpu-pipelined/{kev_0_8b_decode_fp16_pf16 → kev_0_8b_decode_fp16_metal_pf128}/metadata.json

@@ -1,18 +1,18 @@
 {
   "metadata_version": "0.2",
   "kind": "decision-backbone",
-  "name": "kev_0_8b_decode_fp16_pf16",
+  "name": "kev_0_8b_decode_fp16_metal_pf128",
   "assets": {
-    "main": "kev_0_8b_decode_fp16_pf16.aimodel"
+    "main": "kev_0_8b_decode_fp16_metal_pf128.aimodel"
   },
   "language": {
     "tokenizer": "Qwen/Qwen3.5-0.8B-Base",
     "vocab_size": 248320,
     "max_context_length": 4096,
     "embedded_tokenizer": true,
-    "prefill_chunk": 16,
+    "prefill_chunk": 128,
     "static_inputs": [],
-    "output": "hidden [1, 16, 1024] fp16, every position",
+    "output": "hidden [1, 128, 1024] fp16, every position",
     "function_map": {
       "main": [
         "main"
@@ -43,15 +43,14 @@
       "formula": "W + (B @ A) * alpha / r (alpha / r = 2.0) on every adapted weight, in fp32, written fp32 (12 target kinds, 186 adapted tensors); every other weight unchanged",
       "weights_sha256": "ab6bd41853ef2a45d56923202e15c8c9c3dcedaec152b1fb8e529f1ec208415f",
       "model_safetensors_sha256": "b5ebf92a9994a96d5c0c21b52f5eae23049c9e8b5a85fc625b4cea64076408db",
-      "tensors": 320,
-      "script": "scripts/merge_lora_checkpoint.py (github.com/jaredpalmer/kev tag kev-1.0)"
+      "tensors": 320
     },
     "hf_model_id": "jaredpalmer/kev-0.8b",
     "hf_revision": "bf75a6a8848ea6960ff2ed108d9ed44c2941174f",
     "tokenizer": {
       "repo": "Qwen/Qwen3.5-0.8B-Base",
       "revision": "dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68",
-      "files": {
+      "files_sha256": {
         "tokenizer.json": "fe000e3ed39ed12b8d2481d527d44f93c65d37e87645d2dcc80d1bf9d50d2927",
         "tokenizer_config.json": "e611fbccc7c29ef3b1cafb1cb7ea548d189968632901d678fd62be68c47885de"
       },
@@ -62,7 +61,7 @@
   "compression": null,
   "decision": {
     "output": "hidden [1, S, 1024] per call: the final-norm hidden state at every position (no vocabulary head in the graph)",
-    "readout": "one row per question, fresh zero states per row; a row of T ids runs as ceil(T / 16) calls of 'main' (static S = 16): call k gets ids[16k : 16k + 16] with position_ids 0..16k+15; the last call is padded with <|endoftext|> (248044) and the hidden rows of the padded positions are discarded (causal: they cannot reach a real position). The hidden rows [1, 16, 1024] of every call, concatenated and cut to T, are the backbone's final-norm last_hidden_state [T, 1024] the head reads. A row fits when ceil(T / 16) * 16 <= 4095 (the position dim's upper bound; the KV sequence dim is allocated at 4096).",
+    "readout": "one row per question, fresh zero states per row; a row of T ids runs as ceil(T / 128) calls of 'main' (static S = 128): call k gets ids[128k : 128k + 128] with position_ids 0..128k+127; the last call is padded with <|endoftext|> (248044) and the hidden rows of the padded positions are discarded (causal: they cannot reach a real position). The hidden rows [1, 128, 1024] of every call, concatenated and cut to T, are the backbone's final-norm last_hidden_state [T, 1024] the head reads. A row fits when ceil(T / 128) * 128 <= 4095 (the position dim's upper bound; the KV sequence dim is allocated at 4096).",
     "row": {
       "source": "the author's kev.model.encode + rows_of (row form, https://github.com/jaredpalmer/kev tag kev-1.0); conversion/kev/oracle_kev.py records the oracle's rows",
       "layout": "[<state>] + user_tokens(render(state)) + [<q>] + user_tokens(render(instructions)) + for each option ([<opt>] + user_tokens(option_text) + [</opt>]) + [<decide>]",
@@ -124,18 +123,7 @@
         "choice": "the criteria names, in order",
         "score": "'0'..'n-1'"
       },
-      "max_options": 255,
-      "validation": {
-        "source": "kev.api.SystemOneRequest (pydantic v2)",
-        "model": "string, default 'kev-latest'",
-        "state": "required, any JSON value (null allowed)",
-        "questions": "object of question id -> question, at least one",
-        "question": "type noul | choice | score; instructions: any JSON (optional); unknown fields ignored",
-        "noul": "criteria: object or null (optional); only 'true' and 'false' are read",
-        "choice": "criteria: object of 1..255 name -> description (any JSON; null or '' = the name alone)",
-        "score": "criteria: array of 1..255 levels (any JSON)",
-        "reject": "anything else (the author's server answers 422)"
-      }
+      "max_options": 255
     },
     "head": {
       "files": [
@@ -146,15 +134,7 @@
       "scale": 0.0625,
       "temperature": "kev_head.json temperature (2.3510958125672174, the author's calibration in head.pt)",
       "code": "kev.model.PointerHead.forward (q, k = nn.Linear(1024, 256)); weights from head.pt, fp32",
-      "dtype_of_record": "fp32 (the gate reads the graph's fp16 hidden through the fp32 head)",
-      "files_sha256": {
-        "head/head.safetensors": "12038b02ecf7ece3ffb008b22ae97988293e521a86e7aadc8036c9c700893a42",
-        "head/kev_head.json": "73d510708790422d3c84bfc915e311e536d527bc032f0a07394e76999d5c9d7c"
-      },
-      "temperature_value": 2.3510958125672174,
-      "hidden_size": 1024,
-      "head_dim": 256,
-      "host_arithmetic": "the round-5 reference host (conversion/kev/host.py) computes the head and the softmax in float64 from the fp16 hidden rows and the fp32 weights and rounds p to fp32 once; the author's PyTorch head is fp32 (on the same hidden rows they differ by fp32 rounding: max |dp| 3.0e-07 over the fixture and held-out rows)"
+      "dtype_of_record": "fp32 (the gate reads the graph's fp16 hidden through the fp32 head)"
     },
     "softmax": "per question, fp32, over that question's options, after the temperature",
     "response": {
@@ -164,17 +144,28 @@
       "score": "{type: 'score', score: sum(i * p_i), legend: {key: render(level)}, probabilities: {key: p}, confidence: max(0, 1 - E|level - mode| / D), D = mean |i - (L-1)/2| over the L levels, mode = the first most likely level (1 when L = 1)}",
       "normalize": "the confidences normalize p to sum 1 first (all zeros -> uniform)",
       "rounding": "4 decimals (kev.api.round_prob)",
-      "usage": "input_tokens = the encoded request's tokens (the author's packed form: state once + every branch); output_tokens = tokens of json.dumps(answers) (kev.api.output_tokens)",
-      "sums": "every float sum in the confidences and the score is Python 3.12's sum(): left to right with Neumaier compensation (the author's runtime); p are the fp32 probabilities as doubles; round4 = Python round(x, 4)",
-      "model": "the author's server echoes the request's model (default 'kev-latest'); the round-5 reference host writes the bundle name",
-      "latency_ms": "the graph calls + the head (state allocations included; tokenizing and the answers not), 1 decimal",
-      "output_tokens_detail": "tokens of Python's json.dumps(answers) with its defaults (', ' and ': ' separators, ensure_ascii, floats as repr), plain tokenizer (special tokens matched, no user_tokens rewrite)"
-    },
-    "graph_context": {
-      "max_context_length": 4096,
-      "row_limit": "ceil(T / 16) * 16 <= 4095 (the position axis' upper bound = max_context_length - 1), so a row holds at most 4,080 tokens; the author's serving limits (state <= 65,536, row <= 73,728) are far above it"
-    },
-    "shared_prefix": "optional, exact: with Ls state tokens (<state> included) every question's row starts with the same k = Ls // 16 chunks; run them once, copy the four states per question (KV up to k*16 on its sequence axis, the rest zero; convState / recState as they are) and run each row's remaining tokens from position k*16 in the same chunk grid: the hidden rows are bit-equal to the direct run's (round 5, every fixture and held-out row of both models)"
+      "usage": "input_tokens = the encoded request's tokens (the author's packed form: state once + every branch); output_tokens = tokens of json.dumps(answers) (kev.api.output_tokens)"
+    }
+  },
+  "gdn_scan": {
+    "form": "metal",
+    "kernel": "qwen3_5_gdn_chunk_s128",
+    "chunk_max": 128,
+    "kernel_source": "coreai_models.models.macos.qwen3_5_gdn_metal.build_gdn_chunk_kernel (overlay, unchanged)",
+    "threads_per_grid": [
+      128,
+      16,
+      1
+    ],
+    "threads_per_thread_group": [
+      128,
+      1,
+      1
+    ]
+  },
+  "compilation": {
+    "date": "2026-10-04T04:58:42.572786+00:00",
+    "targets": []
   },
   "gate": {
     "bar": {
@@ -185,95 +176,105 @@
       "finite": true
     },
     "oracle": "the author's fp32 checkpoint on the CPU (kev.model.admit + the row-form forward, threads 1)",
+    "graph": "the AOT h16c .aimodelc of this bundle's .aimodel, Python runtime, SpecializationOptions.default(), 128 ids per call",
     "fixture_434": {
       "result": "PASS",
       "rows": 434,
       "argmax_non_near_tie": "420/420",
-      "argmax_near_tie": "12/14",
-      "max_abs_dp": 0.011679,
-      "mean_of_run_mean_abs_dp": 0.0009687,
+      "argmax_near_tie": "13/14",
+      "max_abs_dp": 0.012426,
+      "mean_of_run_mean_abs_dp": 0.0009621,
       "worst_row": {
         "id": "semif_46b7029b9a704138b77a",
         "q": 0,
-        "max_abs_dp": 0.011678963899612427
+        "max_abs_dp": 0.012425720691680908
       },
       "reset_bit_equal_all_processes": true,
       "finite": true,
       "oracle_sha256": "1a72a45831eecc2af7ddfc68a86e9250f572d6bf9274362f96cc62bcb803cc6d",
-      "aimodelc_tree_sha256": "de1b6a1a12a0a7a3726bd043e9e5859b3ddea295e5c5dcec7fa87b33d25cef02",
-      "when": "2026-10-03T07:46:12+00:00",
-      "lane_file_sha256": "2c6a7846081de2e83bea6c6df792418aa8f2d5f846b96dadb7358fd0bf66219a",
-      "transcript": "coreai-model-zoo models/kev-0.8b/gate-kev-0.8b-readout-fp16_pf16.json"
+      "when": "2026-10-04T00:00:04+00:00",
+      "lane_file_sha256": "e15434cd82af6c6d2adcde98c5db3f4f9e36a94b1e83730b5a9f6996148490ab",
+      "transcript": "coreai-model-zoo models/kev-0.8b/gate-kev-0.8b-readout.json"
     },
     "heldout_130": {
       "result": "PASS",
       "rows": 130,
       "argmax_non_near_tie": "125/125",
       "argmax_near_tie": "5/5",
-      "max_abs_dp": 0.005739,
-      "mean_of_run_mean_abs_dp": 0.0008054,
+      "max_abs_dp": 0.005783,
+      "mean_of_run_mean_abs_dp": 0.0008296,
       "worst_row": {
         "id": "tv4h_11",
         "q": 0,
-        "max_abs_dp": 0.005739271640777588
+        "max_abs_dp": 0.0057830810546875
       },
       "reset_bit_equal_all_processes": true,
       "finite": true,
       "oracle_sha256": "f715653331a81bd0ff48d48d1f9d294b7b9b3343713a855fd0eed16dd1e505e6",
-      "aimodelc_tree_sha256": "de1b6a1a12a0a7a3726bd043e9e5859b3ddea295e5c5dcec7fa87b33d25cef02",
-      "when": "2026-10-03T10:00:46+00:00",
-      "lane_file_sha256": "0cbfed2c8e3a9c1cb58ecbb11de5668e7c5c2744e925be08980f589746bad603",
-      "transcript": "coreai-model-zoo models/kev-0.8b/gate-kev-0.8b-heldout.json"
+      "when": "2026-10-04T00:00:20+00:00",
+      "lane_file_sha256": "801f8420c6a0f001e62e6f848d0fef1c332373b1a0dd43cd7ccf4fe844c0276f",
+      "transcript": "coreai-model-zoo models/kev-0.8b/gate-kev-0.8b-readout.json"
     },
-    "host_round5": {
-      "test_host": {
-        "sha256": "972b6ff76ad651d001de9971094b348f0fa03a92539319263904d4e5b4b22c12",
-        "oracle_venv_file": "results/host_test_oracle_venv.json",
-        "oracle_venv_sha256": "32d7ca3bbecd115df29d7a67a91ea2c5660188d52b9b5f0b56d45b7a7dd27e73",
-        "lane_file": "results/host_test.json"
+    "red_arms_v2": {
+      "result": "RED (every arm)",
+      "transcript": "coreai-model-zoo models/kev-0.8b/gate-kev-0.8b-readout.json"
+    },
+    "swift": {
+      "result": "PASS",
+      "jit_vs_aot": {
+        "records": 10,
+        "rows": 25,
+        "hidden_sha256_equal": 25,
+        "p_bit_equal": 25
       },
-      "e2e_fixture": {
-        "sha256": "b8c3e6e3344c8899c22bf73cb8e9f51c525dbd387d3bd5b3e7245b91bc2f8851",
-        "rows": 434,
-        "hidden_bit_equal_gate": 434,
-        "torch_head_p_bit_equal_gate": 434,
-        "host_p_max_abs_dp_vs_gate": 2.980232238769531e-07,
-        "shared_p_bit_equal_direct": 434,
-        "host_bar_max_abs_dp": 0.011679,
-        "host_bar_mean": 0.0009687,
-        "lane_file": "results/e2e_kev-0.8b.json"
+      "transcript": "coreai-model-zoo models/kev-0.8b/gate-kev-0.8b-swift.json"
+    },
+    "iphone_18_pro": {
+      "run": "20261004-140048",
+      "pass": true,
+      "fixture_434": {
+        "argmax_equal_non_near_tie": 420,
+        "questions_non_near_tie": 420,
+        "argmax_equal_near_tie": 13,
+        "near_tie_questions": 14,
+        "max_abs_dp": 0.01100119948387146,
+        "mean_of_run_mean_abs_dp": 0.0009723608388956548
       },
-      "e2e_heldout": {
-        "sha256": "9a04028c1c590ca17d3a27095c185412599f6fc8016361440a66a4b51dc588ad",
-        "rows": 130,
-        "hidden_bit_equal_gate": 130,
-        "torch_head_p_bit_equal_gate": 130,
-        "host_p_max_abs_dp_vs_gate": 2.384185791015625e-07,
-        "shared_p_bit_equal_direct": 130,
-        "host_bar_max_abs_dp": 0.005739,
-        "host_bar_mean": 0.0008054,
-        "lane_file": "results/e2e_kev-0.8b_heldout.json"
+      "heldout_130": {
+        "argmax_equal_non_near_tie": 125,
+        "questions_non_near_tie": 125,
+        "argmax_equal_near_tie": 5,
+        "near_tie_questions": 5,
+        "max_abs_dp": 0.005920350551605225,
+        "mean_of_run_mean_abs_dp": 0.0008314862770240348
       },
-      "transcript": "coreai-model-zoo models/kev-0.8b/gate-kev-0.8b-host.json"
+      "transcript": "coreai-model-zoo models/kev-0.8b/gate-kev-0.8b-iphone.json"
     },
-    "aimodelc": {
-      "name": "kev_0_8b_decode_fp16_pf16.h16c.aimodelc",
-      "tree_sha256": "de1b6a1a12a0a7a3726bd043e9e5859b3ddea295e5c5dcec7fa87b33d25cef02"
+    "host": {
+      "transcript": "coreai-model-zoo models/kev-0.8b/gate-kev-0.8b-host.json"
     },
     "fixture": "coreai-model-zoo models/kev-0.8b/fixtures-kev-0.8b.json"
   },
-  "compilation": {
-    "date": "2026-10-03T07:38:20.417082+00:00",
-    "targets": []
-  },
   "metadata_of_record": {
-    "exported_metadata_json_sha256": "01251e9649bd73421a67ee0942b047d52ea81ae12f3b3c310958bd9e9bd139f0",
+    "exported_metadata_json_sha256": "a1b6e5a35ab4d744b8a469caa8c22ad574437a80b5864dd93f6bbc87979d873c",
     "edits": [
       "language.tokenizer = the base repo id (was the local id of the merged checkpoint) + tokenizer_revision",
       "source.hf_model_id / hf_revision = the Kev checkpoint (was the local HF-cache id of the merged weights)",
-      "decision.request / response / graph_context / shared_prefix and gate: added in round 5 from the reference host (metadata.candidate.json)",
-      "gate transcripts named by their zoo files"
+      "source.tokenizer.shipped_in added",
+      "gate: the published form's gate numbers and the zoo transcripts that hold them"
+    ],
+    "host_read_keys_unchanged": [
+      "name",
+      "kind",
+      "assets",
+      "language.vocab_size",
+      "language.max_context_length",
+      "language.prefill_chunk",
+      "language.query_len_range",
+      "language.query_len_call_max",
+      "language.query_len_multiple",
+      "decision"
     ],
-    "written": "2026-10-04 (round 9)"
+    "written": "2026-10-04 (round 17)"
   }
 }

gpu-pipelined/kev_0_8b_decode_fp16_pf16/kev_0_8b_decode_fp16_pf16.aimodel/main.hash

@@ -1 +0,0 @@
-A1�(H�x{�l�l�*ph�G�N%��˗[��rh-