Release v1.1.3 scored WebGPU evidence

Adds browser state-drift scoring and release evidence, refreshes the pinned wllama artifacts, and keeps DFlash inspection-only and disabled by default.

.gitattributes

@@ -1 +1,3 @@
 public/wasm/*.wasm filter=lfs diff=lfs merge=lfs -text
+dist/wasm/wllama-compat.wasm filter=lfs diff=lfs merge=lfs -text
+dist/wasm/wllama.wasm filter=lfs diff=lfs merge=lfs -text

README.md

@@ -71,8 +71,29 @@ deep-link keeps the benchmark executable under the same COOP/COEP headers. It
 loads nothing until **Run benchmark** is pressed, then records first-load and
 verified-cache reload wall time, TTFT,
 prefill/decode throughput, graph placement, device limits, engine revisions,
-the exact runtime artifact hash, and a fixed temperature-0 prompt. The report
-exports as JSON; generated model text is deliberately excluded.
+the exact runtime artifact hash, and a fixed temperature-0 prompt. Schema v3
+exports the exact messages, sampling controls, raw generated text, pinned model
+source/shard hashes, and an opt-in sampled-token trace. Normal chat does not
+request logprobs and returns no token-id trace.
+
+The default `field-core-v1` workload remains short. The separate
+`state-drift-1k-v1` evidence workload locks greedy sampling and refuses to
+export unless all 1,024 sampled token IDs and the 64/128/256/512/768/1,024
+uint32-LE prefix hashes are present. Its diagnostic-only 27B WebGPU path also
+validates the pinned native CPU prompt/reference fixture, then teacher-forces
+that fixed 1,024-token sequence through raw token-ID prefixes. Schema v3 keeps
+the forced reference logprob, natural top five, mean NLL, and perplexity
+separate from the natural browser generation. This scoring method is not
+available to normal chat and does not activate DFlash.
+
+On the pinned Apple `metal-3` release path, the final JSPI runtime completed the
+natural 1,024-token 27B run at 11.3 tok/s with one graph, zero CPU ops, and
+65/65 GPU layers. Exact CPU/browser token parity remains `NO-GO` because the
+first natural mismatch is a reciprocal 13/198 top-one/top-two flip at position
+30. The separate numerical gate passes: browser and native CPU mean NLL are
+0.0456380321 and 0.0455182718 (absolute delta 0.0001197604 against a 0.01
+limit), with near-tie margins 0.006506 and 0.003605. This supports a bounded
+GPU floating-point tolerance; it is not a bit-exact parity claim.
 
 `n_batch`, `n_ubatch`, Flash Attention, paired K/V cache types, and the pinned
 JSPI/compat WASM flavor are available there as explicitly experimental
@@ -83,14 +104,21 @@ and Q4_0 KV: every row retained one graph, 29/29 GPU layers, zero CPU ops, and
 the same short deterministic output hash. Observed KV allocation fell from
 112 MiB for F16 to 59.5 MiB for Q8_0 and 31.5 MiB for Q4_0.
 
+The separate final-artifact 27B >8K smoke used an 8,448-token context with
+8,314 prompt tokens, Flash auto, and Q4_0/Q4_0 KV. It measured 7.55 prefill
+tok/s and 6.42 decode tok/s, retained one graph with zero CPU ops and 65/65
+GPU layers, and exported all 8 requested top-five token records. This is one
+local Apple `metal-3` result; it does not establish a cross-device default.
+
 A same-Chrome A/B with one warmup and three alternating measured runs per arm
 also proved explicit JSPI/compat switching. Median warm load was 2,257.7 ms for
 JSPI and 2,799.1 ms for compat; all eight outputs and graph tripwires matched.
 These are experimental single-device results, not release recommendations;
 normal chat keeps Flash off, F16/F16, and automatic WASM selection.
 
-The public prewarmed-cache baseline below is one Apple `metal-3` / headless
-Chrome sample, not a cross-device study and not a first-download claim:
+The historical public v1.1.2 prewarmed-cache baseline below is one Apple
+`metal-3` / headless Chrome sample, not a cross-device study and not a
+first-download claim:
 
 | Tier | Backend | Load | TTFT | Decode |
 | --- | --- | ---: | ---: | ---: |
@@ -102,10 +130,11 @@ Chrome sample, not a cross-device study and not a first-download claim:
 ## DFlash research lane
 
 DFlash is visible only as an inspection-only experiment. Its BF16 source and
-Q8_0 native GGUF reference are pinned in the manifest. Shape, tokenizer, target
-taps, conversion, and a 32-token temperature-0 native Bonsai-27B parity smoke
-pass. The custom WASM contains the DFlash llama.cpp architecture, but the
-browser load API does not yet plumb `spec_type=draft-dflash`; the Q8
+Q8_0 native GGUF reference are pinned in the manifest. The strict D1 audit
+passes all 39 shape, tokenizer, target-tap, conversion, revision, and evidence
+checks, including a 32-token temperature-0 native Bonsai-27B parity smoke. The
+custom WASM contains the DFlash llama.cpp architecture, but the browser load
+API does not yet plumb `spec_type=draft-dflash`; the Q8
 `fc.weight` also exceeds the 128 MiB baseline WebGPU binding limit. It remains
 disabled and unavailable until a lower-bit candidate, browser parity,
 acceptance, and positive speed gates all pass.
@@ -125,10 +154,10 @@ stay in the browser; only model/static asset downloads contact Hugging Face.
 - Dawn native WebGPU validation: `18eb229ef5f707c1464cc581252e7603c73a3ef0`
 - custom wllama source: `912c18b75d4358c1405a64646b8dbe43a205943b`
 - custom nested llama.cpp: `00fa7cb284cbf133fc426733bd64238a3588a33e`
-- custom patch set: `ad8786a295eaeb75ee03752d7f84d30f14ea373407844d5c589f7ab46800bb4b`
-- custom ESM: `24bcef8aea8e27fb7b7e2d9e6ea94ba8ced7bfbffd09a0675821b9eb1b4a4c9f`
-- JSPI WASM: `dd71f58c75f32c677a64eeb2e56efab7dedbe7395f730bb08bc80021bd182caa`
-- compat WASM: `eca9754d0d8a490b5c7b58cb384d67e31aa8ab26dd330e4ad1460982df6dd82e`
+- custom patch set: `27b96fa48c6fd66b44fe70e83cc775902e29ec5877261c7eb04fecabc529d2f9`
+- custom ESM: `10c3811776e34092a225632929b97046af4215a4c91156169105c886be5cba5d`
+- JSPI WASM: `b292bf670a25e1ecf02c71dd92b49196f3336842135f1f85efd5a3fd5f36ea8a`
+- compat WASM: `ea0623d2ca1758287a569169b4ad5a69c22e41bce3e5d022c7d3e4fb5c37918c`
 - compat worker: `0b667db536815fb9e5dc9ffbfbd6ba29affdda38a4ba892d5372ddb41ce9a8c8`
 
 The vendored runtime exposes the validated token-embedding WebGPU placement
@@ -139,9 +168,11 @@ backend tripwire disagrees. Its exact JS/WASM hashes are checked during
 `THIRD_PARTY_NOTICES.md` and `public/licenses/`.
 
 The pinned WebGPU source implements Flash Attention with Q4_0/Q8_0 K/V. The
-benchmark lane has short deterministic-output, memory, and graph-placement
-evidence for those combinations, but long-state, 27B, and cross-device gates
-remain open. Flash Attention therefore remains off for the release path.
+benchmark lane has short 1.7B deterministic-output and memory evidence for
+both quantized types, a 27B long-state result only for Flash-off F16, and a
+27B >8K smoke only for Flash-auto Q4_0. No 27B Q8_0 long-state result is
+claimed. Repeated and cross-device gates remain open, so Flash Attention
+remains off and F16/F16 remains the release default.
 
 ## Licenses and attribution
 

dist/fixtures/state-drift-27b-cpu-reference.json

@@ -0,0 +1 @@
+{"schemaVersion":1,"kind":"bonsai-state-drift-reference","workload":{"id":"state-drift-1k-v1","messages":[{"role":"user","content":"Output consecutive integers starting at 1, separated only by a comma and one space. Continue without explanation until you reach 500."}]},"provenance":{"sourceEvidence":"results/space-model/27b-native-cpu-state-drift-1024-b32-u16.json","engineRevision":"00fa7cb284cbf133fc426733bd64238a3588a33e","nativeBinary":{"bytes":65392,"sha256":"e8f23d8612ed27739f21a6aca1df618699da902f6fc42cbdb3f9c978f6da04fb"},"model":{"file":"Bonsai-27B-Q1_0.gguf","bytes":3803452480,"sha256":"17ef842e47450caeb8eaa3ebfbbab5d2f2278b62b79be107985fb69a2f819aa0"},"renderedPromptSha256":"7c46d0de443e6f1235ddaa7c0a55c9da710eb73958d5671247b6f0e3b7d189c7","execution":{"backend":"cpu","contextSize":2048,"batchSize":32,"microBatchSize":16}},"tokenEncoding":"uint32-le","promptTokenIds":[248045,846,198,4935,23173,24959,5732,506,220,16,11,18101,1132,539,264,30644,321,799,3433,13,14569,1973,15673,2980,488,5372,220,20,15,15,13,248046,198,248045,74455,198,248068,198],"promptTokenIdsSha256":"e7af7d66c96f24cb148575bd45cef079920669eb0d193fdc830ce61dfec85451","referenceTokenIds":[8160,579,264,7047,1817,25,271,16,13,220,2972,2014,53983,2570,5396,64700,198,256,471,2972,6065,64700,8984,23173,24959,5732,506,220,16,13,198,256,471,2972,3925,64700,32001,639,1132,539,264,30644,321,799,3433,318,72,1673,2487,328,16,11,220,17,11,220,18,11,2423,1764,198,256,471,2972,17355,64700,14569,1973,15673,2980,18176,220,20,15,15,13,198,256,471,2972,57545,64700,561,2468,1220,381,264,3074,18677,886,20296,314,4947,494,220,16,310,220,20,15,15,11,22405,6681,430,5024,13,271,17,13,220,2972,27382,1386,5141,84457,64700,198,256,471,4980,506,220,16,198,256,471,3839,506,220,20,15,15,198,256,471,75604,25,3544,328,318,43871,478,3433,8,198,256,471,2233,39492,11,1066,279,4947,198,256,471,1160,82819,24959,271,18,13,220,2972,35,23544,8984,14606,64700,198,256,561,2468,1220,1353,1040,25,198,256,1510,16,11,220,17,11,220,18,11,220,19,11,220,20,11,58317,220,19,24,24,11,220,20,15,15,63,271,19,13,220,2972,30097,279,27929,64700,198,256,353,1144,310,6707,4947,494,220,16,310,220,20,15,15,11,10542,539,3544,5748,198,256,8439,353,2688,449,14791,11,353,628,6707,411,53624,6804,303,821,1965,13,353,3172,8972,424,15060,13,271,256,6558,579,9845,279,4581,3443,25,328,16,11,220,17,11,220,18,11,58317,220,20,15,15,1,198,256,353,3172,6707,424,5774,13,271,256,6817,25,353,1144,310,1236,2617,1017,579,874,4799,1414,11,874,48794,35358,421,2493,38556,11,1066,279,6891,1414,430,10897,13,561,9640,2640,328,323,47930,1132,539,264,30644,321,799,3433,487,748,353,3172,24660,1732,421,13,271,256,353,3172,6707,279,2400,8240,1381,13,271,256,27929,25,220,16,11,220,17,11,220,18,11,220,19,11,220,20,11,220,21,11,220,22,11,220,23,11,220,24,11,220,16,15,11,220,16,16,11,220,16,17,11,220,16,18,11,220,16,19,11,220,16,20,11,220,16,21,11,220,16,22,11,220,16,23,11,220,16,24,11,220,17,15,11,220,17,16,11,220,17,17,11,220,17,18,11,220,17,19,11,220,17,20,11,220,17,21,11,220,17,22,11,220,17,23,11,220,17,24,11,220,18,15,11,220,18,16,11,220,18,17,11,220,18,18,11,220,18,19,11,220,18,20,11,220,18,21,11,220,18,22,11,220,18,23,11,220,18,24,11,220,19,15,11,220,19,16,11,220,19,17,11,220,19,18,11,220,19,19,11,220,19,20,11,220,19,21,11,220,19,22,11,220,19,23,11,220,19,24,11,220,20,15,11,220,20,16,11,220,20,17,11,220,20,18,11,220,20,19,11,220,20,20,11,220,20,21,11,220,20,22,11,220,20,23,11,220,20,24,11,220,21,15,11,220,21,16,11,220,21,17,11,220,21,18,11,220,21,19,11,220,21,20,11,220,21,21,11,220,21,22,11,220,21,23,11,220,21,24,11,220,22,15,11,220,22,16,11,220,22,17,11,220,22,18,11,220,22,19,11,220,22,20,11,220,22,21,11,220,22,22,11,220,22,23,11,220,22,24,11,220,23,15,11,220,23,16,11,220,23,17,11,220,23,18,11,220,23,19,11,220,23,20,11,220,23,21,11,220,23,22,11,220,23,23,11,220,23,24,11,220,24,15,11,220,24,16,11,220,24,17,11,220,24,18,11,220,24,19,11,220,24,20,11,220,24,21,11,220,24,22,11,220,24,23,11,220,24,24,11,220,16,15,15,11,220,16,15,16,11,220,16,15,17,11,220,16,15,18,11,220,16,15,19,11,220,16,15,20,11,220,16,15,21,11,220,16,15,22,11,220,16,15,23,11,220,16,15,24,11,220,16,16,15,11,220,16,16,16,11,220,16,16,17,11,220,16,16,18,11,220,16,16,19,11,220,16,16,20,11,220,16,16,21,11,220,16,16,22,11,220,16,16,23,11,220,16,16,24,11,220,16,17,15,11,220,16,17,16,11,220,16,17,17,11,220,16,17,18,11,220,16,17,19,11,220,16,17,20,11,220,16,17,21,11,220,16,17,22,11,220,16,17,23,11,220,16,17,24,11,220,16,18,15,11,220,16,18,16,11,220,16,18,17,11,220,16,18,18,11,220,16,18,19,11,220,16,18,20,11,220,16,18,21,11,220,16,18,22,11,220,16,18,23,11,220,16,18,24,11,220,16,19,15,11,220,16,19,16,11,220,16,19,17,11,220,16,19,18,11,220,16,19,19,11,220,16,19,20,11,220,16,19,21,11,220,16,19,22,11,220,16,19,23,11,220,16,19,24,11,220,16,20,15,11,220,16,20,16,11,220,16,20,17,11,220,16,20,18,11,220,16,20,19,11,220,16,20,20,11,220,16],"referenceTokenIdsSha256":"c503b2db0dcf10daecfda4f19a5f466d1699b31688076d6c32f7c29daefcef9b","checkpointPrefixes":[{"tokens":64,"sha256":"f6398ce3b2c935435615b8ba33d0416c03228e4a1ab0a9ff3f4229418558a130"},{"tokens":128,"sha256":"0efb5a50141c6b4140846e69df0e3ed3b86491040fa52f3cc4fb1ab561c153a5"},{"tokens":256,"sha256":"f4e94ebe35a375348e5156732df7fc858060cd350b4e2f366e6c249758db7b9e"},{"tokens":512,"sha256":"11acf787b85658d6a8db7a4d872faf9145684291aba790b39afd184ecf48e7b9"},{"tokens":768,"sha256":"faa10cea44ae4d461f355ed31f1599b5b67c6ac121b9296c2ed24af8af92aec7"},{"tokens":1024,"sha256":"c503b2db0dcf10daecfda4f19a5f466d1699b31688076d6c32f7c29daefcef9b"}]}

dist/index.html

@@ -10,7 +10,7 @@
       content="Bonsai local inference field station — private WebGPU language models in your browser."
     />
     <title>Bonsai Field Station</title>
-    <script type="module" crossorigin src="/assets/index-jtL5gUVS.js"></script>
+    <script type="module" crossorigin src="/assets/index-DIzN0g8r.js"></script>
     <link rel="stylesheet" crossorigin href="/assets/index-C-2tm91j.css">
   </head>
   <body>

dist/wasm/wllama-compat.wasm

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eca9754d0d8a490b5c7b58cb384d67e31aa8ab26dd330e4ad1460982df6dd82e
-size 14864188
+oid sha256:ea0623d2ca1758287a569169b4ad5a69c22e41bce3e5d022c7d3e4fb5c37918c
+size 14865338

dist/wasm/wllama.wasm

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dd71f58c75f32c677a64eeb2e56efab7dedbe7395f730bb08bc80021bd182caa
-size 8017169
+oid sha256:b292bf670a25e1ecf02c71dd92b49196f3336842135f1f85efd5a3fd5f36ea8a
+size 8017784

public/fixtures/state-drift-27b-cpu-reference.json

@@ -0,0 +1 @@
+{"schemaVersion":1,"kind":"bonsai-state-drift-reference","workload":{"id":"state-drift-1k-v1","messages":[{"role":"user","content":"Output consecutive integers starting at 1, separated only by a comma and one space. Continue without explanation until you reach 500."}]},"provenance":{"sourceEvidence":"results/space-model/27b-native-cpu-state-drift-1024-b32-u16.json","engineRevision":"00fa7cb284cbf133fc426733bd64238a3588a33e","nativeBinary":{"bytes":65392,"sha256":"e8f23d8612ed27739f21a6aca1df618699da902f6fc42cbdb3f9c978f6da04fb"},"model":{"file":"Bonsai-27B-Q1_0.gguf","bytes":3803452480,"sha256":"17ef842e47450caeb8eaa3ebfbbab5d2f2278b62b79be107985fb69a2f819aa0"},"renderedPromptSha256":"7c46d0de443e6f1235ddaa7c0a55c9da710eb73958d5671247b6f0e3b7d189c7","execution":{"backend":"cpu","contextSize":2048,"batchSize":32,"microBatchSize":16}},"tokenEncoding":"uint32-le","promptTokenIds":[248045,846,198,4935,23173,24959,5732,506,220,16,11,18101,1132,539,264,30644,321,799,3433,13,14569,1973,15673,2980,488,5372,220,20,15,15,13,248046,198,248045,74455,198,248068,198],"promptTokenIdsSha256":"e7af7d66c96f24cb148575bd45cef079920669eb0d193fdc830ce61dfec85451","referenceTokenIds":[8160,579,264,7047,1817,25,271,16,13,220,2972,2014,53983,2570,5396,64700,198,256,471,2972,6065,64700,8984,23173,24959,5732,506,220,16,13,198,256,471,2972,3925,64700,32001,639,1132,539,264,30644,321,799,3433,318,72,1673,2487,328,16,11,220,17,11,220,18,11,2423,1764,198,256,471,2972,17355,64700,14569,1973,15673,2980,18176,220,20,15,15,13,198,256,471,2972,57545,64700,561,2468,1220,381,264,3074,18677,886,20296,314,4947,494,220,16,310,220,20,15,15,11,22405,6681,430,5024,13,271,17,13,220,2972,27382,1386,5141,84457,64700,198,256,471,4980,506,220,16,198,256,471,3839,506,220,20,15,15,198,256,471,75604,25,3544,328,318,43871,478,3433,8,198,256,471,2233,39492,11,1066,279,4947,198,256,471,1160,82819,24959,271,18,13,220,2972,35,23544,8984,14606,64700,198,256,561,2468,1220,1353,1040,25,198,256,1510,16,11,220,17,11,220,18,11,220,19,11,220,20,11,58317,220,19,24,24,11,220,20,15,15,63,271,19,13,220,2972,30097,279,27929,64700,198,256,353,1144,310,6707,4947,494,220,16,310,220,20,15,15,11,10542,539,3544,5748,198,256,8439,353,2688,449,14791,11,353,628,6707,411,53624,6804,303,821,1965,13,353,3172,8972,424,15060,13,271,256,6558,579,9845,279,4581,3443,25,328,16,11,220,17,11,220,18,11,58317,220,20,15,15,1,198,256,353,3172,6707,424,5774,13,271,256,6817,25,353,1144,310,1236,2617,1017,579,874,4799,1414,11,874,48794,35358,421,2493,38556,11,1066,279,6891,1414,430,10897,13,561,9640,2640,328,323,47930,1132,539,264,30644,321,799,3433,487,748,353,3172,24660,1732,421,13,271,256,353,3172,6707,279,2400,8240,1381,13,271,256,27929,25,220,16,11,220,17,11,220,18,11,220,19,11,220,20,11,220,21,11,220,22,11,220,23,11,220,24,11,220,16,15,11,220,16,16,11,220,16,17,11,220,16,18,11,220,16,19,11,220,16,20,11,220,16,21,11,220,16,22,11,220,16,23,11,220,16,24,11,220,17,15,11,220,17,16,11,220,17,17,11,220,17,18,11,220,17,19,11,220,17,20,11,220,17,21,11,220,17,22,11,220,17,23,11,220,17,24,11,220,18,15,11,220,18,16,11,220,18,17,11,220,18,18,11,220,18,19,11,220,18,20,11,220,18,21,11,220,18,22,11,220,18,23,11,220,18,24,11,220,19,15,11,220,19,16,11,220,19,17,11,220,19,18,11,220,19,19,11,220,19,20,11,220,19,21,11,220,19,22,11,220,19,23,11,220,19,24,11,220,20,15,11,220,20,16,11,220,20,17,11,220,20,18,11,220,20,19,11,220,20,20,11,220,20,21,11,220,20,22,11,220,20,23,11,220,20,24,11,220,21,15,11,220,21,16,11,220,21,17,11,220,21,18,11,220,21,19,11,220,21,20,11,220,21,21,11,220,21,22,11,220,21,23,11,220,21,24,11,220,22,15,11,220,22,16,11,220,22,17,11,220,22,18,11,220,22,19,11,220,22,20,11,220,22,21,11,220,22,22,11,220,22,23,11,220,22,24,11,220,23,15,11,220,23,16,11,220,23,17,11,220,23,18,11,220,23,19,11,220,23,20,11,220,23,21,11,220,23,22,11,220,23,23,11,220,23,24,11,220,24,15,11,220,24,16,11,220,24,17,11,220,24,18,11,220,24,19,11,220,24,20,11,220,24,21,11,220,24,22,11,220,24,23,11,220,24,24,11,220,16,15,15,11,220,16,15,16,11,220,16,15,17,11,220,16,15,18,11,220,16,15,19,11,220,16,15,20,11,220,16,15,21,11,220,16,15,22,11,220,16,15,23,11,220,16,15,24,11,220,16,16,15,11,220,16,16,16,11,220,16,16,17,11,220,16,16,18,11,220,16,16,19,11,220,16,16,20,11,220,16,16,21,11,220,16,16,22,11,220,16,16,23,11,220,16,16,24,11,220,16,17,15,11,220,16,17,16,11,220,16,17,17,11,220,16,17,18,11,220,16,17,19,11,220,16,17,20,11,220,16,17,21,11,220,16,17,22,11,220,16,17,23,11,220,16,17,24,11,220,16,18,15,11,220,16,18,16,11,220,16,18,17,11,220,16,18,18,11,220,16,18,19,11,220,16,18,20,11,220,16,18,21,11,220,16,18,22,11,220,16,18,23,11,220,16,18,24,11,220,16,19,15,11,220,16,19,16,11,220,16,19,17,11,220,16,19,18,11,220,16,19,19,11,220,16,19,20,11,220,16,19,21,11,220,16,19,22,11,220,16,19,23,11,220,16,19,24,11,220,16,20,15,11,220,16,20,16,11,220,16,20,17,11,220,16,20,18,11,220,16,20,19,11,220,16,20,20,11,220,16],"referenceTokenIdsSha256":"c503b2db0dcf10daecfda4f19a5f466d1699b31688076d6c32f7c29daefcef9b","checkpointPrefixes":[{"tokens":64,"sha256":"f6398ce3b2c935435615b8ba33d0416c03228e4a1ab0a9ff3f4229418558a130"},{"tokens":128,"sha256":"0efb5a50141c6b4140846e69df0e3ed3b86491040fa52f3cc4fb1ab561c153a5"},{"tokens":256,"sha256":"f4e94ebe35a375348e5156732df7fc858060cd350b4e2f366e6c249758db7b9e"},{"tokens":512,"sha256":"11acf787b85658d6a8db7a4d872faf9145684291aba790b39afd184ecf48e7b9"},{"tokens":768,"sha256":"faa10cea44ae4d461f355ed31f1599b5b67c6ac121b9296c2ed24af8af92aec7"},{"tokens":1024,"sha256":"c503b2db0dcf10daecfda4f19a5f466d1699b31688076d6c32f7c29daefcef9b"}]}

public/wasm/wllama-compat.wasm

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eca9754d0d8a490b5c7b58cb384d67e31aa8ab26dd330e4ad1460982df6dd82e
-size 14864188
+oid sha256:ea0623d2ca1758287a569169b4ad5a69c22e41bce3e5d022c7d3e4fb5c37918c
+size 14865338

public/wasm/wllama.wasm

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dd71f58c75f32c677a64eeb2e56efab7dedbe7395f730bb08bc80021bd182caa
-size 8017169
+oid sha256:b292bf670a25e1ecf02c71dd92b49196f3336842135f1f85efd5a3fd5f36ea8a
+size 8017784

scripts/verify-wllama-assets.mjs

@@ -44,6 +44,10 @@ const wllamaDeclarations = await readFile(resolve(root, 'vendor/wllama-bonsai/es
 if (!wllamaDeclarations.includes('forceCompat?: boolean')) {
   throw new Error('Vendored Bonsai wllama declarations lack forceCompat');
 }
+const oaiDeclarations = await readFile(resolve(root, 'vendor/wllama-bonsai/esm/types/oai-compat.d.ts'), 'utf8');
+if (!oaiDeclarations.includes('export interface ChatCompletionLogprob {\n    id: number;')) {
+  throw new Error('Vendored Bonsai wllama declarations lack sampled token ids');
+}
 const moduleSource = await readFile(resolve(root, 'vendor/wllama-bonsai/esm/index.js'), 'utf8');
 if (!moduleSource.includes('n_ubatch: params.n_ubatch')) {
   throw new Error('Vendored Bonsai wllama module does not forward n_ubatch');

src/bench/BenchApp.tsx

@@ -2,25 +2,33 @@ import { useEffect, useMemo, useRef, useState } from 'react';
 import {
   BrowserEngineClient,
   EngineClientError,
+  evaluateModelContextPolicy,
   evaluateModelGate,
   loadModelManifestV2,
   type BenchmarkFlashMode,
   type BenchmarkKvCacheType,
   type BenchmarkWasmFlavor,
   type EngineCapabilities,
+  type GenerateParams,
   type ModelManifestV2,
   type ModelTierId,
   type RequestedBackend,
 } from '../engine';
 import { formatBytes } from '../lib/format';
 import {
-  BENCHMARK_PROMPT,
-  BENCHMARK_PROMPT_ID,
+  BENCHMARK_WORKLOADS,
+  DEFAULT_BENCHMARK_WORKLOAD_ID,
   benchmarkReportFilename,
   buildBenchmarkReport,
   serializeBenchmarkReport,
   type BenchmarkReport,
+  type BenchmarkWorkloadId,
+  type TeacherForcedReferenceScoreEvidence,
 } from './report';
+import {
+  STATE_DRIFT_REFERENCE_PATH,
+  validateStateDriftReferenceFixture,
+} from './state-drift-reference';
 
 const MANIFEST_PATH = 'manifest/models.json';
 const DEFAULT_MODEL: ModelTierId = '1_7b';
@@ -43,6 +51,13 @@ function manifestUrl(): string {
   return new URL(`${import.meta.env.BASE_URL}${MANIFEST_PATH}`, document.baseURI).href;
 }
 
+function stateDriftReferenceUrl(): string {
+  return new URL(
+    `${import.meta.env.BASE_URL}${STATE_DRIFT_REFERENCE_PATH}`,
+    document.baseURI,
+  ).href;
+}
+
 function errorMessage(error: unknown): string {
   if (error instanceof EngineClientError) return `${error.code}: ${error.message}`;
   return error instanceof Error ? error.message : String(error);
@@ -89,6 +104,7 @@ export function BenchApp() {
   const [client] = useState(getBenchClient);
   const [manifest, setManifest] = useState<ModelManifestV2 | null>(null);
   const [capabilities, setCapabilities] = useState<EngineCapabilities | null>(null);
+  const [workloadId, setWorkloadId] = useState<BenchmarkWorkloadId>(DEFAULT_BENCHMARK_WORKLOAD_ID);
   const [modelId, setModelId] = useState<ModelTierId>(DEFAULT_MODEL);
   const [backend, setBackend] = useState<RequestedBackend>('auto');
   const [contextSize, setContextSize] = useState(4_096);
@@ -146,6 +162,7 @@ export function BenchApp() {
     () => manifest?.models.find((candidate) => candidate.id === modelId) ?? null,
     [manifest, modelId],
   );
+  const workload = BENCHMARK_WORKLOADS[workloadId];
   const deviceGate = useMemo(() => {
     if (!model || !capabilities) return null;
     return evaluateModelGate(model, backend, capabilities.webgpu, {
@@ -159,9 +176,15 @@ export function BenchApp() {
     return evaluateModelGate(model, backend, capabilities.webgpu, capabilities.storage);
   }, [backend, capabilities, model]);
 
-  const contextLimit = model?.id === '27b'
-    ? model.defaultContext
-    : model?.contextLength ?? 4_096;
+  const contextPolicy = model
+    ? evaluateModelContextPolicy(model, contextSize, {
+        tuningScope: 'benchmark',
+        requestedBackend: backend,
+        flashMode,
+        kvCacheType,
+      })
+    : null;
+  const contextLimit = contextPolicy?.limit ?? 4_096;
   const nBatch = parseOptionalInteger(nBatchInput);
   const nUbatch = parseOptionalInteger(nUbatchInput);
   const batchTuningError = nBatchInput.trim() !== '' && nBatch === null
@@ -179,13 +202,27 @@ export function BenchApp() {
   const wasmTuningError = wasmFlavor === 'jspi' && capabilities?.runtime.wasmFlavor === 'compat'
     ? 'JSPI is unavailable in this browser; use auto or compat.'
     : null;
-  const configurationTuningError = batchTuningError ?? runtimeTuningError ?? wasmTuningError;
-  const configurationValid = Number.isSafeInteger(contextSize)
-    && contextSize > 0
-    && contextSize <= contextLimit
+  const workloadError = workload.requiredMaxTokens !== null && maxTokens !== workload.requiredMaxTokens
+    ? `${workload.id} requires exactly ${workload.requiredMaxTokens.toLocaleString()} max tokens.`
+    : null;
+  const stateDriftReferenceError = workload.id === 'state-drift-1k-v1' && (
+    model?.id !== '27b'
+    || backend !== 'webgpu'
+    || contextSize !== 2_048
+    || nBatch !== 32
+    || nUbatch !== 16
+  )
+    ? 'state-drift-1k-v1 reference scoring requires 27B, explicit WebGPU, context 2,048, nBatch 32, and nUbatch 16.'
+    : null;
+  const configurationTuningError = batchTuningError
+    ?? runtimeTuningError
+    ?? wasmTuningError
+    ?? workloadError
+    ?? stateDriftReferenceError;
+  const configurationValid = contextPolicy?.allowed === true
     && Number.isSafeInteger(maxTokens)
     && maxTokens >= 8
-    && maxTokens <= 512
+    && maxTokens <= 1_024
     && configurationTuningError === null;
   const canRun = Boolean(manifest && capabilities && model && deviceGate?.allowed && configurationValid);
 
@@ -299,9 +336,8 @@ export function BenchApp() {
       let firstTokenAt: number | null = null;
       let streamedTokenEvents = 0;
       const generationStarted = performance.now();
-      setProgress({ label: `Generating fixed prompt · ${BENCHMARK_PROMPT_ID}`, loadedBytes: 0, totalBytes: maxTokens });
-      const generationResult = await client.generate({
-        messages: [{ role: 'user', content: BENCHMARK_PROMPT }],
+      const generationRequest = {
+        messages: workload.messages.map((message) => ({ ...message })),
         maxTokens,
         temperature: 0,
         topK: 1,
@@ -309,7 +345,10 @@ export function BenchApp() {
         seed: 42,
         toolChoice: 'none',
         cachePrompt: false,
-      }, {
+        returnTokenIds: true,
+      } satisfies GenerateParams;
+      setProgress({ label: `Generating fixed workload · ${workload.id}`, loadedBytes: 0, totalBytes: maxTokens });
+      const generationResult = await client.generate(generationRequest, {
         signal: controller.signal,
         onToken: (text, reasoningDelta) => {
           if (!text && !reasoningDelta) return;
@@ -324,6 +363,40 @@ export function BenchApp() {
       });
       const generationCompleted = performance.now();
       throwIfAborted(controller.signal);
+      let teacherForcedReferenceScore: TeacherForcedReferenceScoreEvidence | null = null;
+      if (workload.id === 'state-drift-1k-v1') {
+        setProgress({
+          label: 'Loading and validating pinned CPU reference',
+          loadedBytes: 0,
+          totalBytes: 1,
+        });
+        const fixtureResponse = await fetch(stateDriftReferenceUrl(), {
+          signal: controller.signal,
+          cache: 'no-cache',
+        });
+        if (!fixtureResponse.ok) {
+          throw new Error(
+            `CPU reference fixture request failed with HTTP ${fixtureResponse.status}.`,
+          );
+        }
+        const fixture = validateStateDriftReferenceFixture(
+          await fixtureResponse.json() as unknown,
+          { model, capabilities, loadResult: warmLoad },
+        );
+        throwIfAborted(controller.signal);
+        setProgress({
+          label: 'Teacher-forcing the 1,024-token CPU reference',
+          loadedBytes: 0,
+          totalBytes: fixture.referenceTokenIds.length,
+        });
+        const score = await client.scoreSequence({
+          promptTokenIds: fixture.promptTokenIds,
+          referenceTokenIds: fixture.referenceTokenIds,
+          topK: 5,
+        }, { signal: controller.signal });
+        teacherForcedReferenceScore = { fixture, score };
+        throwIfAborted(controller.signal);
+      }
       const backendReport = await client.backendReport();
       throwIfAborted(controller.signal);
 
@@ -335,7 +408,8 @@ export function BenchApp() {
         capabilities,
         requestedBackend: backend,
         contextSize: warmLoad.context.size,
-        maxTokens,
+        workloadId,
+        generationRequest,
         coldLoadMs,
         warmLoadMs,
         coldCachedBytes,
@@ -346,9 +420,10 @@ export function BenchApp() {
         timeToFirstTokenMs: firstTokenAt === null ? null : firstTokenAt - generationStarted,
         streamedTokenEvents,
         backendReport,
+        teacherForcedReferenceScore,
       });
       setReport(nextReport);
-      setOutput(generationResult.text);
+      setOutput(generationResult.text || generationResult.reasoningText);
       setProgress({
         label: 'Benchmark complete · JSON report is ready',
         loadedBytes: 1,
@@ -398,6 +473,26 @@ export function BenchApp() {
           <p className="bench-control-intro">The fixed prompt and temperature stay locked so exported runs remain comparable.</p>
 
           <form onSubmit={(event) => { event.preventDefault(); runBenchmark(); }}>
+            <label className="bench-field">
+              <span>Workload</span>
+              <select
+                value={workloadId}
+                disabled={running}
+                onChange={(event) => {
+                  const nextId = event.target.value as BenchmarkWorkloadId;
+                  const next = BENCHMARK_WORKLOADS[nextId];
+                  setWorkloadId(nextId);
+                  setMaxTokens(next.requiredMaxTokens ?? 64);
+                  clearPreviousRun();
+                }}
+              >
+                {Object.values(BENCHMARK_WORKLOADS).map((candidate) => (
+                  <option key={candidate.id} value={candidate.id}>{candidate.label}</option>
+                ))}
+              </select>
+              <small>{workload.description}</small>
+            </label>
+
             <label className="bench-field">
               <span>Model tier</span>
               <select
@@ -451,7 +546,7 @@ export function BenchApp() {
                 <input
                   type="number"
                   min="8"
-                  max="512"
+                  max="1024"
                   step="8"
                   value={maxTokens}
                   disabled={running}
@@ -460,7 +555,7 @@ export function BenchApp() {
                     clearPreviousRun();
                   }}
                 />
-                <small>8–512</small>
+                <small>8–1,024</small>
               </label>
             </div>
 
@@ -642,6 +737,8 @@ export function BenchApp() {
                 <div><dt>Decode</dt><dd>{formatRate(report?.generation.decodeTokensPerSecond)}</dd></div>
                 <div><dt>Token count</dt><dd>{report?.generation.completionTokens ?? '—'}</dd></div>
                 <div><dt>Finish</dt><dd>{report?.generation.finishReason ?? '—'}</dd></div>
+                <div><dt>CPU-ref PPL</dt><dd>{report?.generation.teacherForcedReferenceScore?.score.summary.perplexity.toFixed(4) ?? '—'}</dd></div>
+                <div><dt>CPU-ref mean NLL</dt><dd>{report?.generation.teacherForcedReferenceScore?.score.summary.meanNll.toFixed(6) ?? '—'}</dd></div>
               </dl>
             </article>
             <article className="bench-data-card">
@@ -677,11 +774,11 @@ export function BenchApp() {
           <section className="bench-output">
             <div>
               <p className="eyebrow ink">Fixed specimen</p>
-              <h3>{BENCHMARK_PROMPT_ID}</h3>
-              <p>{BENCHMARK_PROMPT}</p>
+              <h3>{workload.id}</h3>
+              <p>{workload.messages.map((message) => message.content ?? '').join('\n')}</p>
             </div>
             <div>
-              <p className="eyebrow ink">Generated text · excluded from export</p>
+              <p className="eyebrow ink">Generated output · visible text, else reasoning trace</p>
               <p data-testid="bench-generated-output">{output || 'Run the assay to inspect the generated specimen.'}</p>
             </div>
           </section>

src/bench/report.test.ts

@@ -2,16 +2,21 @@ import { describe, expect, it } from 'vitest';
 import type {
   BackendReport,
   EngineCapabilities,
+  GenerateParams,
   LoadModelResult,
   ManifestModelV2,
   ModelManifestV2,
 } from '../engine';
 import {
+  BENCHMARK_WORKLOADS,
+  DEFAULT_BENCHMARK_WORKLOAD_ID,
+  TOKEN_ID_CHECKPOINTS,
   benchmarkReportFilename,
   buildBenchmarkReport,
   classifyCacheState,
   serializeBenchmarkReport,
   type BenchmarkObservation,
+  type TeacherForcedReferenceScoreEvidence,
 } from './report';
 
 const JSPI_WASM_SHA256 = 'd'.repeat(64);
@@ -33,6 +38,18 @@ const model = {
   id: '8b',
   displayName: 'Bonsai 8B',
   architecture: 'qwen3next',
+  source: {
+    repo: 'prism-ml/Bonsai-8B-gguf',
+    revision: '0123456789abcdef0123456789abcdef01234567',
+    file: 'Bonsai-8B-Q1_0.gguf',
+    bytes: 1_000,
+    sha256: 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+  },
+  files: [{
+    path: '8b/Bonsai-8B-Q1_0.gguf',
+    bytes: 1_000,
+    sha256: 'bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb',
+  }],
   downloadBytes: 1_000,
 } as ManifestModelV2;
 
@@ -150,6 +167,96 @@ const loadResult = {
   backendReport,
 } satisfies LoadModelResult;
 
+const fieldCoreGenerationRequest: GenerateParams = {
+  messages: BENCHMARK_WORKLOADS[DEFAULT_BENCHMARK_WORKLOAD_ID].messages.map((message) => ({
+    ...message,
+  })),
+  maxTokens: 64,
+  temperature: 0,
+  topP: 1,
+  topK: 1,
+  seed: 42,
+  toolChoice: 'none',
+  cachePrompt: false,
+  returnTokenIds: true,
+};
+
+function sampledTokenTrace(tokenIds: readonly number[]) {
+  return tokenIds.map((id) => ({
+    selected: { id, logprob: -0.1 },
+    topCandidates: [
+      { id, logprob: -0.1 },
+      { id: id + 10_001, logprob: -1 },
+      { id: id + 10_002, logprob: -2 },
+      { id: id + 10_003, logprob: -3 },
+      { id: id + 10_004, logprob: -4 },
+    ],
+  }));
+}
+
+function teacherForcedReferenceScore(
+  referenceTokenIds: readonly number[],
+): TeacherForcedReferenceScoreEvidence {
+  const entries = referenceTokenIds.map((id, index) => ({
+    index,
+    selectedReference: { id, logprob: -0.1 },
+    naturalTop1: { id, logprob: -0.1 },
+    topCandidates: [
+      { id, logprob: -0.1 },
+      { id: id + 10_001, logprob: -1 },
+      { id: id + 10_002, logprob: -2 },
+      { id: id + 10_003, logprob: -3 },
+      { id: id + 10_004, logprob: -4 },
+    ],
+    referenceRankInTopCandidatesZeroBased: 0,
+    top1Top2Margin: 0.9,
+  }));
+  return {
+    fixture: {
+      schemaVersion: 1,
+      kind: 'bonsai-state-drift-reference',
+      workload: {
+        id: 'state-drift-1k-v1',
+        messages: BENCHMARK_WORKLOADS['state-drift-1k-v1'].messages.map((message) => ({
+          role: 'user',
+          content: message.content ?? '',
+        })),
+      },
+      provenance: {
+        sourceEvidence: 'results/space-model/fixture.json',
+        engineRevision: capabilities.runtime.llamaCppRevision,
+        nativeBinary: { bytes: 1, sha256: '1'.repeat(64) },
+        model: { file: model.source.file, bytes: model.source.bytes, sha256: model.source.sha256 },
+        renderedPromptSha256: '2'.repeat(64),
+        execution: { backend: 'cpu', contextSize: 4_096, batchSize: 2_048, microBatchSize: 512 },
+      },
+      tokenEncoding: 'uint32-le',
+      promptTokenIds: Array.from({ length: 38 }, (_, index) => index + 1),
+      promptTokenIdsSha256: '3'.repeat(64),
+      referenceTokenIds: [...referenceTokenIds],
+      referenceTokenIdsSha256: '4'.repeat(64),
+      checkpointPrefixes: TOKEN_ID_CHECKPOINTS.map((tokens) => ({
+        tokens,
+        sha256: '5'.repeat(64),
+      })),
+    },
+    score: {
+      method: {
+        promptMode: 'raw-token-id-prefix',
+        maxTokensPerStep: 1,
+        temperature: 0,
+        topK: 1,
+        reportedTopLogprobs: 5,
+        logitBias: 1_000,
+        cachePromptFirst: false,
+        cachePromptSubsequent: true,
+      },
+      entries,
+      summary: { tokenCount: 1_024, meanNll: 0.1, perplexity: Math.exp(0.1) },
+    },
+  };
+}
+
 function observation(overrides: Partial<BenchmarkObservation> = {}): BenchmarkObservation {
   return {
     startedAt: '2026-07-15T10:00:00.000Z',
@@ -159,7 +266,11 @@ function observation(overrides: Partial<BenchmarkObservation> = {}): BenchmarkOb
     capabilities,
     requestedBackend: 'auto',
     contextSize: 4_096,
-    maxTokens: 64,
+    workloadId: DEFAULT_BENCHMARK_WORKLOAD_ID,
+    generationRequest: {
+      ...fieldCoreGenerationRequest,
+      messages: fieldCoreGenerationRequest.messages.map((message) => ({ ...message })),
+    },
     coldLoadMs: 1_234.56,
     warmLoadMs: 234.54,
     coldCachedBytes: 250,
@@ -167,6 +278,9 @@ function observation(overrides: Partial<BenchmarkObservation> = {}): BenchmarkOb
     loadResult,
     generationResult: {
       text: 'fixture',
+      reasoningText: '',
+      tokenIds: [101, 102, 103, 104, 105, 106, 107, 108],
+      tokenTrace: sampledTokenTrace([101, 102, 103, 104, 105, 106, 107, 108]),
       finishReason: 'stop',
       toolCalls: [],
       usage: { promptTokens: 20, completionTokens: 8, totalTokens: 28 },
@@ -180,6 +294,87 @@ function observation(overrides: Partial<BenchmarkObservation> = {}): BenchmarkOb
   };
 }
 
+function longContextObservation(): BenchmarkObservation {
+  const longContextBackendReport = {
+    ...backendReport,
+    layersGpu: { offloaded: 65, total: 65 },
+    flashAttention: true,
+    cacheTypeK: 'q4_0',
+    cacheTypeV: 'q4_0',
+    webgpuKvBufferBytes: 155_713_536,
+  } satisfies BackendReport;
+  const longContextLoadResult = {
+    ...loadResult,
+    modelId: '27b',
+    gate: {
+      ...loadResult.gate,
+      requestedBackend: 'webgpu',
+      selectedBackend: 'webgpu',
+    },
+    context: { ...loadResult.context, size: 8_448, layerCount: 65 },
+    tuning: {
+      scope: 'benchmark',
+      requested: {
+        nBatch: 32,
+        nUbatch: 16,
+        flashMode: 'auto',
+        cacheTypeK: 'q4_0',
+        cacheTypeV: 'q4_0',
+        wasmFlavor: 'auto',
+      },
+      observed: {
+        nBatch: 32,
+        nUbatch: 16,
+        flashAttention: true,
+        cacheTypeK: 'q4_0',
+        cacheTypeV: 'q4_0',
+        kvBufferBytes: 155_713_536,
+        wasmFlavor: 'jspi',
+        wasmSha256: JSPI_WASM_SHA256,
+        compatWorkerSha256: null,
+      },
+      applied: true,
+    },
+    backendReport: longContextBackendReport,
+  } satisfies LoadModelResult;
+
+  return observation({
+    model: {
+      ...model,
+      id: '27b',
+      displayName: 'Bonsai 27B',
+      architecture: 'qwen35',
+    } as ManifestModelV2,
+    requestedBackend: 'webgpu',
+    contextSize: 8_448,
+    workloadId: 'context-prefill-8k-v1',
+    generationRequest: {
+      messages: BENCHMARK_WORKLOADS['context-prefill-8k-v1'].messages.map((message) => ({ ...message })),
+      maxTokens: 8,
+      temperature: 0,
+      topP: 1,
+      topK: 1,
+      seed: 42,
+      toolChoice: 'none',
+      cachePrompt: false,
+      returnTokenIds: true,
+    },
+    loadResult: longContextLoadResult,
+    backendReport: longContextBackendReport,
+    streamedTokenEvents: 8,
+    generationResult: {
+      text: 'fixture',
+      reasoningText: '',
+      tokenIds: [201, 202, 203, 204, 205, 206, 207, 208],
+      tokenTrace: sampledTokenTrace([201, 202, 203, 204, 205, 206, 207, 208]),
+      finishReason: 'length',
+      toolCalls: [],
+      usage: { promptTokens: 8_314, completionTokens: 8, totalTokens: 8_322 },
+      timings: { promptTokensPerSecond: 20, predictedTokensPerSecond: 30 },
+    },
+  });
+}
+
 describe('benchmark report', () => {
   it('classifies the shard cache observed at the first load boundary', () => {
     expect(classifyCacheState(null, 1_000)).toBe('unknown');
@@ -191,6 +386,11 @@ describe('benchmark report', () => {
   it('builds a shareable report with pinned runtime evidence and fixed-safe policy', () => {
     const report = buildBenchmarkReport(observation());
 
+    expect(report.schemaVersion).toBe(3);
+    expect(report.model).toMatchObject({
+      source: model.source,
+      shards: model.files,
+    });
     expect(report.load).toMatchObject({
       tuning: {
         scope: 'release-defaults',
@@ -219,12 +419,38 @@ describe('benchmark report', () => {
       warm: { durationMs: 234.5 },
     });
     expect(report.generation).toMatchObject({
+      workload: {
+        id: 'field-core-v1',
+        messages: BENCHMARK_WORKLOADS['field-core-v1'].messages,
+      },
+      sampling: {
+        maxTokens: 64,
+        temperature: 0,
+        topP: 1,
+        topK: 1,
+        minP: null,
+        seed: 42,
+        toolChoice: 'none',
+        cachePrompt: false,
+      },
+      tokenTrace: {
+        returnTokenIds: true,
+        logprobs: true,
+        topLogprobs: 5,
+        topAlternatives: 4,
+        encoding: 'uint32-le',
+        tokenIds: [101, 102, 103, 104, 105, 106, 107, 108],
+        entries: sampledTokenTrace([101, 102, 103, 104, 105, 106, 107, 108]),
+        checkpointPrefixes: [],
+      },
+      rawText: 'fixture',
+      rawReasoningText: '',
       promptTokensPerSecond: 123.46,
       decodeTokensPerSecond: 45.68,
       engineCompletionTokens: 8,
       streamedTokenEvents: 5,
       completionTokens: 8,
-      tokenCountSource: 'engine-usage',
+      tokenCountSource: 'token-id-trace',
     });
     expect(report.execution).toMatchObject({
       selectedBackend: 'webgpu',
@@ -345,6 +571,17 @@ describe('benchmark report', () => {
     });
   });
 
+  it('exports reasoning-only output separately from visible content', () => {
+    const input = observation();
+    input.generationResult.text = '';
+    input.generationResult.reasoningText = 'reasoning-only fixture';
+
+    const report = buildBenchmarkReport(input);
+
+    expect(report.generation.rawText).toBe('');
+    expect(report.generation.rawReasoningText).toBe('reasoning-only fixture');
+  });
+
   it('exports the selected model DFlash pairing for unsupported 1.7B and conditional 27B', () => {
     const reportFor = (modelId: ManifestModelV2['id']) => buildBenchmarkReport(observation({
       model: {
@@ -378,11 +615,14 @@ describe('benchmark report', () => {
     expect(exportedPairing.targetLayerIdsZeroBased).not.toBe(dflash27bPairing.targetLayerIdsZeroBased);
   });
 
-  it('uses streamed events when final engine usage is present but incomplete', () => {
+  it('uses the exact token-id trace when final engine usage is present but incomplete', () => {
     const report = buildBenchmarkReport(observation({
       streamedTokenEvents: 64,
       generationResult: {
         text: 'fixture',
+        reasoningText: '',
+        tokenIds: Array.from({ length: 64 }, (_, index) => index + 1),
+        tokenTrace: sampledTokenTrace(Array.from({ length: 64 }, (_, index) => index + 1)),
         finishReason: 'length',
         toolCalls: [],
         usage: { promptTokens: 5, completionTokens: 0, totalTokens: 5 },
@@ -396,16 +636,22 @@ describe('benchmark report', () => {
       completionTokens: 64,
       engineTotalTokens: 5,
       totalTokens: 69,
-      tokenCountSource: 'stream-events-fallback',
+      tokenCountSource: 'token-id-trace',
     });
+    expect(report.generation.tokenTrace.checkpointPrefixes).toEqual([
+      { tokens: 64, sha256: '0c8f462927e331f28e3f1a6d342957cd27118febc309bd3b2f646e2dfbaeec32' },
+    ]);
   });
 
-  it('falls back to streamed events without emitting non-finite JSON metrics', () => {
+  it('uses the trace without emitting non-finite JSON metrics', () => {
     const report = buildBenchmarkReport(observation({
       generationElapsedMs: Number.NaN,
       timeToFirstTokenMs: Number.POSITIVE_INFINITY,
       generationResult: {
         text: 'fixture',
+        reasoningText: '',
+        tokenIds: [1, 2, 3, 4, 5],
+        tokenTrace: sampledTokenTrace([1, 2, 3, 4, 5]),
         finishReason: 'length',
         toolCalls: [],
         usage: null,
@@ -418,12 +664,228 @@ describe('benchmark report', () => {
       timeToFirstTokenMs: null,
       decodeTokensPerSecond: null,
       completionTokens: 5,
-      tokenCountSource: 'stream-events-fallback',
+      tokenCountSource: 'token-id-trace',
     });
     expect(serializeBenchmarkReport(report)).not.toContain('NaN');
     expect(serializeBenchmarkReport(report)).toMatch(/\n$/);
   });
 
+  it('records all locked state-drift checkpoints from the exact token-id prefixes', () => {
+    const tokenIds = Array.from({ length: 1_024 }, (_, index) => index + 1);
+    const report = buildBenchmarkReport(observation({
+      workloadId: 'state-drift-1k-v1',
+      generationRequest: {
+        messages: BENCHMARK_WORKLOADS['state-drift-1k-v1'].messages.map((message) => ({ ...message })),
+        maxTokens: 1_024,
+        temperature: 0,
+        topP: 1,
+        topK: 1,
+        seed: 42,
+        toolChoice: 'none',
+        cachePrompt: false,
+        returnTokenIds: true,
+      },
+      streamedTokenEvents: 1_024,
+      generationResult: {
+        text: '1, 2, 3, …',
+        reasoningText: '',
+        tokenIds,
+        tokenTrace: sampledTokenTrace(tokenIds),
+        finishReason: 'length',
+        toolCalls: [],
+        usage: { promptTokens: 20, completionTokens: 1_024, totalTokens: 1_044 },
+        timings: { promptTokensPerSecond: 20, predictedTokensPerSecond: 30 },
+      },
+      teacherForcedReferenceScore: teacherForcedReferenceScore(tokenIds),
+    }));
+
+    expect(report.generation.workload).toEqual({
+      id: 'state-drift-1k-v1',
+      messages: BENCHMARK_WORKLOADS['state-drift-1k-v1'].messages,
+    });
+    expect(report.generation.rawText).toBe('1, 2, 3, …');
+    expect(report.generation.rawReasoningText).toBe('');
+    expect(report.generation.tokenTrace.tokenIds).toEqual(tokenIds);
+    expect(report.generation.tokenTrace.entries).toHaveLength(1_024);
+    expect(report.generation.teacherForcedReferenceScore).toMatchObject({
+      fixture: { kind: 'bonsai-state-drift-reference' },
+      score: { summary: { tokenCount: 1_024, meanNll: 0.1 } },
+    });
+    expect(report.generation.tokenTrace.checkpointPrefixes).toEqual([
+      { tokens: 64, sha256: '0c8f462927e331f28e3f1a6d342957cd27118febc309bd3b2f646e2dfbaeec32' },
+      { tokens: 128, sha256: '24f9ac547baae524ba0ea5220692d48f7526cdb1df5e99edcbb1f32239a8d5f5' },
+      { tokens: 256, sha256: '80fa0f6d1caca9aad2b012051399b33bcd1976b145f3f3eea0f7ba10637761b0' },
+      { tokens: 512, sha256: '8ab9b2cf36ec9e9d68711df73334731d2fee0552f5d95d76538b1d2cdcefd564' },
+      { tokens: 768, sha256: '23905fc64dc47867f49672959cddf676ca257967f5786eefaf8583ddf7ddf3e9' },
+      { tokens: 1_024, sha256: '6b8b6bd30ff821daf5db90cc071525f5696e366efd1aeb30d0bf60f785a3c26d' },
+    ]);
+  });
+
+  it('exports null teacher-forced scoring for non-state-drift workloads', () => {
+    expect(buildBenchmarkReport(observation()).generation.teacherForcedReferenceScore).toBeNull();
+  });
+
+  it('rejects incomplete state-drift traces instead of exporting partial evidence', () => {
+    expect(() => buildBenchmarkReport(observation({
+      workloadId: 'state-drift-1k-v1',
+      generationRequest: {
+        messages: BENCHMARK_WORKLOADS['state-drift-1k-v1'].messages.map((message) => ({ ...message })),
+        maxTokens: 1_024,
+        temperature: 0,
+        topP: 1,
+        topK: 1,
+        seed: 42,
+        toolChoice: 'none',
+        cachePrompt: false,
+        returnTokenIds: true,
+      },
+      generationResult: {
+        text: 'partial fixture',
+        reasoningText: '',
+        tokenIds: Array.from({ length: 1_023 }, (_, index) => index + 1),
+        tokenTrace: sampledTokenTrace(Array.from({ length: 1_023 }, (_, index) => index + 1)),
+        finishReason: 'length',
+        toolCalls: [],
+        usage: { promptTokens: 20, completionTokens: 1_023, totalTokens: 1_043 },
+        timings: null,
+      },
+    }))).toThrow('requires exactly 1,024 sampled token ids');
+  });
+
+  it('rejects state-drift traces that stop before exhausting the locked budget', () => {
+    const input = observation({
+      workloadId: 'state-drift-1k-v1',
+      generationRequest: {
+        messages: BENCHMARK_WORKLOADS['state-drift-1k-v1'].messages.map((message) => ({ ...message })),
+        maxTokens: 1_024,
+        temperature: 0,
+        topP: 1,
+        topK: 1,
+        seed: 42,
+        toolChoice: 'none',
+        cachePrompt: false,
+        returnTokenIds: true,
+      },
+      generationResult: {
+        text: 'early stop',
+        reasoningText: '',
+        tokenIds: Array.from({ length: 1_024 }, (_, index) => index + 1),
+        tokenTrace: sampledTokenTrace(Array.from({ length: 1_024 }, (_, index) => index + 1)),
+        finishReason: 'stop',
+        toolCalls: [],
+        usage: { promptTokens: 20, completionTokens: 1_024, totalTokens: 1_044 },
+        timings: null,
+      },
+    });
+
+    expect(() => buildBenchmarkReport(input)).toThrow('requires finishReason=length');
+  });
+
+  it('exports long-context evidence only after processing more than 8K prompt tokens', () => {
+    const report = buildBenchmarkReport(longContextObservation());
+
+    expect(BENCHMARK_WORKLOADS['context-prefill-8k-v1'].messages).toEqual([{
+      role: 'user',
+      content: `Continue the pattern.${' x'.repeat(8_300)}`,
+    }]);
+    expect(report).toMatchObject({
+      model: { id: '27b', contextSize: 8_448 },
+      generation: {
+        workload: { id: 'context-prefill-8k-v1' },
+        sampling: { maxTokens: 8 },
+        promptTokens: 8_314,
+        completionTokens: 8,
+      },
+      load: {
+        tuning: {
+          scope: 'benchmark',
+          requested: { flashMode: 'auto', cacheTypeK: 'q4_0', cacheTypeV: 'q4_0' },
+          observed: { flashAttention: true, cacheTypeK: 'q4_0', cacheTypeV: 'q4_0' },
+          applied: true,
+        },
+      },
+      execution: { requestedBackend: 'webgpu', selectedBackend: 'webgpu' },
+    });
+    expect(report.generation.tokenTrace.tokenIds).toHaveLength(8);
+  });
+
+  it('rejects long-context evidence at or below 8,192 engine-counted prompt tokens', () => {
+    const input = longContextObservation();
+    input.generationResult.usage = { promptTokens: 8_192, completionTokens: 8, totalTokens: 8_200 };
+
+    expect(() => buildBenchmarkReport(input)).toThrow(
+      'requires more than 8,192 engine-counted prompt tokens',
+    );
+  });
+
+  it('rejects a long-context trace that stops before exhausting the 8-token budget', () => {
+    const input = longContextObservation();
+    input.generationResult.finishReason = 'stop';
+
+    expect(() => buildBenchmarkReport(input)).toThrow('requires finishReason=length');
+  });
+
+  it('rejects long-context evidence with the wrong context or runtime tuning', () => {
+    const wrongContext = longContextObservation();
+    wrongContext.contextSize = 8_447;
+    wrongContext.loadResult = {
+      ...wrongContext.loadResult,
+      context: { ...wrongContext.loadResult.context, size: 8_447 },
+    };
+    expect(() => buildBenchmarkReport(wrongContext)).toThrow(
+      'requires Bonsai 27B at context 8,448',
+    );
+
+    const wrongTuning = longContextObservation();
+    wrongTuning.loadResult = {
+      ...wrongTuning.loadResult,
+      tuning: {
+        ...wrongTuning.loadResult.tuning,
+        requested: {
+          ...wrongTuning.loadResult.tuning.requested,
+          cacheTypeK: 'q8_0',
+          cacheTypeV: 'q8_0',
+        },
+        observed: {
+          ...wrongTuning.loadResult.tuning.observed,
+          cacheTypeK: 'q8_0',
+          cacheTypeV: 'q8_0',
+        },
+      },
+    };
+    expect(() => buildBenchmarkReport(wrongTuning)).toThrow(
+      'requires Bonsai 27B at context 8,448',
+    );
+  });
+
+  it('rejects missing and malformed token-id traces', () => {
+    const missingTrace = observation();
+    missingTrace.generationResult.tokenIds = null;
+    expect(() => buildBenchmarkReport(missingTrace)).toThrow('without a token-id trace');
+
+    const malformedTrace = observation();
+    malformedTrace.generationResult.tokenIds = [1, 2.5];
+    expect(() => buildBenchmarkReport(malformedTrace)).toThrow('token id 1 is invalid');
+  });
+
+  it('rejects missing, mismatched, and unsorted sampled-token logprob traces', () => {
+    const missingTrace = observation();
+    missingTrace.generationResult.tokenTrace = null;
+    expect(() => buildBenchmarkReport(missingTrace)).toThrow('without a sampled-token logprob trace');
+
+    const mismatchedTrace = observation();
+    mismatchedTrace.generationResult.tokenTrace?.pop();
+    expect(() => buildBenchmarkReport(mismatchedTrace)).toThrow(
+      'sampled-token trace length 7 does not match token-id length 8',
+    );
+
+    const unsortedTrace = observation();
+    const candidates = unsortedTrace.generationResult.tokenTrace?.[0]?.topCandidates;
+    if (!candidates) throw new Error('Unsorted trace fixture requires top candidates.');
+    [candidates[0], candidates[1]] = [candidates[1]!, candidates[0]!];
+    expect(() => buildBenchmarkReport(unsortedTrace)).toThrow('candidates are not sorted');
+  });
+
   it('uses the model, selected backend, and UTC completion time in the export name', () => {
     const report = buildBenchmarkReport(observation());
     expect(benchmarkReportFilename(report)).toBe(

src/bench/report.ts

@@ -1,15 +1,65 @@
+import { sha256 } from '@noble/hashes/sha2.js';
+import { bytesToHex } from '@noble/hashes/utils.js';
 import type {
   BackendReport,
   EngineCapabilities,
+  EngineChatMessage,
+  EngineSampledTokenTraceEntry,
+  GenerateParams,
   GenerateResult,
   LoadModelResult,
   ManifestModelV2,
   ModelManifestV2,
   RequestedBackend,
+  ScoreSequenceResult,
 } from '../engine';
+import type { StateDriftReferenceFixture } from './state-drift-reference';
 
-export const BENCHMARK_PROMPT_ID = 'bonsai-field-core-v1';
-export const BENCHMARK_PROMPT = 'Describe how a tree survives one dry season in exactly eight short, factual sentences.';
+export const BENCHMARK_WORKLOADS = {
+  'field-core-v1': {
+    id: 'field-core-v1',
+    label: 'Field core · short',
+    description: 'Stable throughput workload; completion length remains operator-selected.',
+    requiredMaxTokens: null,
+    messages: [{
+      role: 'user',
+      content: 'Describe how a tree survives one dry season in exactly eight short, factual sentences.',
+    }],
+  },
+  'state-drift-1k-v1': {
+    id: 'state-drift-1k-v1',
+    label: 'State drift · 1K',
+    description: 'P2.5 recurrent-state evidence; requires an exact 1,024-token trace.',
+    requiredMaxTokens: 1_024,
+    messages: [{
+      role: 'user',
+      content: 'Output consecutive integers starting at 1, separated only by a comma and one space. Continue without explanation until you reach 500.',
+    }],
+  },
+  'context-prefill-8k-v1': {
+    id: 'context-prefill-8k-v1',
+    label: 'Context prefill · >8K',
+    description: 'P2.5 long-context evidence; requires more than 8,192 engine-counted prompt tokens.',
+    requiredMaxTokens: 8,
+    messages: [{
+      role: 'user',
+      content: `Continue the pattern.${' x'.repeat(8_300)}`,
+    }],
+  },
+} as const satisfies Record<string, {
+  id: string;
+  label: string;
+  description: string;
+  requiredMaxTokens: number | null;
+  messages: readonly EngineChatMessage[];
+}>;
+
+export type BenchmarkWorkloadId = keyof typeof BENCHMARK_WORKLOADS;
+export const DEFAULT_BENCHMARK_WORKLOAD_ID: BenchmarkWorkloadId = 'field-core-v1';
+export const BENCHMARK_PROMPT_ID = DEFAULT_BENCHMARK_WORKLOAD_ID;
+export const BENCHMARK_PROMPT = BENCHMARK_WORKLOADS[DEFAULT_BENCHMARK_WORKLOAD_ID].messages[0].content;
+export const TOKEN_ID_CHECKPOINTS = [64, 128, 256, 512, 768, 1_024] as const;
+export type TokenIdCheckpointSize = typeof TOKEN_ID_CHECKPOINTS[number];
 
 export type CacheState = 'empty' | 'partial' | 'cached' | 'unknown';
 
@@ -21,7 +71,8 @@ export interface BenchmarkObservation {
   capabilities: EngineCapabilities;
   requestedBackend: RequestedBackend;
   contextSize: number;
-  maxTokens: number;
+  workloadId: BenchmarkWorkloadId;
+  generationRequest: GenerateParams;
   coldLoadMs: number;
   warmLoadMs: number;
   coldCachedBytes: number | null;
@@ -32,10 +83,16 @@ export interface BenchmarkObservation {
   timeToFirstTokenMs: number | null;
   streamedTokenEvents: number;
   backendReport: BackendReport;
+  teacherForcedReferenceScore?: TeacherForcedReferenceScoreEvidence | null;
+}
+
+export interface TeacherForcedReferenceScoreEvidence {
+  fixture: StateDriftReferenceFixture;
+  score: ScoreSequenceResult;
 }
 
 export interface BenchmarkReport {
-  schemaVersion: 2;
+  schemaVersion: 3;
   kind: 'bonsai-browser-benchmark';
   startedAt: string;
   completedAt: string;
@@ -43,6 +100,8 @@ export interface BenchmarkReport {
     id: ManifestModelV2['id'];
     displayName: string;
     architecture: string;
+    source: ManifestModelV2['source'];
+    shards: Array<Pick<ManifestModelV2['files'][number], 'path' | 'bytes' | 'sha256'>>;
     downloadBytes: number;
     contextSize: number;
   };
@@ -60,11 +119,36 @@ export interface BenchmarkReport {
     };
   };
   generation: {
-    promptId: typeof BENCHMARK_PROMPT_ID;
-    prompt: typeof BENCHMARK_PROMPT;
-    maxTokens: number;
-    temperature: 0;
-    seed: 42;
+    workload: {
+      id: BenchmarkWorkloadId;
+      messages: EngineChatMessage[];
+    };
+    sampling: {
+      maxTokens: number;
+      temperature: number;
+      topP: number | null;
+      topK: number | null;
+      minP: number | null;
+      seed: number;
+      toolChoice: GenerateParams['toolChoice'] | null;
+      cachePrompt: boolean;
+    };
+    tokenTrace: {
+      returnTokenIds: true;
+      logprobs: true;
+      topLogprobs: 5;
+      topAlternatives: 4;
+      encoding: 'uint32-le';
+      tokenIds: number[];
+      entries: EngineSampledTokenTraceEntry[];
+      checkpointPrefixes: Array<{
+        tokens: TokenIdCheckpointSize;
+        sha256: string;
+      }>;
+    };
+    teacherForcedReferenceScore: TeacherForcedReferenceScoreEvidence | null;
+    rawText: string;
+    rawReasoningText: string;
     elapsedMs: number | null;
     timeToFirstTokenMs: number | null;
     promptTokensPerSecond: number | null;
@@ -75,7 +159,7 @@ export interface BenchmarkReport {
     completionTokens: number;
     engineTotalTokens: number | null;
     totalTokens: number | null;
-    tokenCountSource: 'engine-usage' | 'stream-events-fallback';
+    tokenCountSource: 'token-id-trace';
     finishReason: GenerateResult['finishReason'];
   };
   execution: {
@@ -141,8 +225,289 @@ export function classifyCacheState(observedBytes: number | null, totalBytes: num
   return 'partial';
 }
 
+function cloneMessages(messages: readonly EngineChatMessage[]): EngineChatMessage[] {
+  return messages.map((message) => message.role === 'assistant' && message.tool_calls
+    ? {
+        ...message,
+        tool_calls: message.tool_calls.map((call) => ({
+          ...call,
+          function: { ...call.function },
+        })),
+      }
+    : { ...message });
+}
+
+function checkpointPrefixSha256(tokenIds: readonly number[], tokens: number): string {
+  const bytes = new Uint8Array(tokens * Uint32Array.BYTES_PER_ELEMENT);
+  const view = new DataView(bytes.buffer);
+  for (let index = 0; index < tokens; index += 1) {
+    view.setUint32(index * Uint32Array.BYTES_PER_ELEMENT, tokenIds[index]!, true);
+  }
+  return bytesToHex(sha256(bytes));
+}
+
+function buildCheckpointPrefixes(tokenIds: readonly number[]): BenchmarkReport['generation']['tokenTrace']['checkpointPrefixes'] {
+  return TOKEN_ID_CHECKPOINTS
+    .filter((tokens) => tokens <= tokenIds.length)
+    .map((tokens) => ({ tokens, sha256: checkpointPrefixSha256(tokenIds, tokens) }));
+}
+
+function cloneValidatedTokenTrace(
+  value: GenerateResult['tokenTrace'],
+  tokenIds: readonly number[],
+): EngineSampledTokenTraceEntry[] {
+  if (value === null) {
+    throw new Error('Benchmark generation completed without a sampled-token logprob trace.');
+  }
+  if (value.length !== tokenIds.length) {
+    throw new Error(
+      `Benchmark sampled-token trace length ${value.length} does not match token-id length ${tokenIds.length}.`,
+    );
+  }
+  return value.map((entry, index) => {
+    const selected = { ...entry.selected };
+    if (selected.id !== tokenIds[index]) {
+      throw new Error(`Benchmark sampled-token trace entry ${index} does not match its token id.`);
+    }
+    if (typeof selected.logprob !== 'number' || !Number.isFinite(selected.logprob)) {
+      throw new Error(`Benchmark sampled-token trace entry ${index} has an invalid selected logprob.`);
+    }
+    if (!Array.isArray(entry.topCandidates) || entry.topCandidates.length !== 5) {
+      throw new Error(`Benchmark sampled-token trace entry ${index} must contain exactly five candidates.`);
+    }
+    const seenIds = new Set<number>();
+    const topCandidates = entry.topCandidates.map((candidate, candidateIndex) => {
+      if (!Number.isSafeInteger(candidate.id) || candidate.id < 0 || candidate.id > 0xffff_ffff) {
+        throw new Error(
+          `Benchmark sampled-token trace entry ${index} candidate ${candidateIndex} has an invalid id.`,
+        );
+      }
+      if (seenIds.has(candidate.id)) {
+        throw new Error(`Benchmark sampled-token trace entry ${index} has duplicate candidate ids.`);
+      }
+      seenIds.add(candidate.id);
+      if (typeof candidate.logprob !== 'number' || !Number.isFinite(candidate.logprob)) {
+        throw new Error(
+          `Benchmark sampled-token trace entry ${index} candidate ${candidateIndex} has an invalid logprob.`,
+        );
+      }
+      const previous = entry.topCandidates[candidateIndex - 1];
+      if (previous && (
+        previous.logprob < candidate.logprob
+        || (previous.logprob === candidate.logprob && previous.id > candidate.id)
+      )) {
+        throw new Error(`Benchmark sampled-token trace entry ${index} candidates are not sorted.`);
+      }
+      return { ...candidate };
+    });
+    const selectedCandidate = topCandidates.find((candidate) => candidate.id === selected.id);
+    if (!selectedCandidate || selectedCandidate.logprob !== selected.logprob) {
+      throw new Error(
+        `Benchmark sampled-token trace entry ${index} does not include its selected token and logprob.`,
+      );
+    }
+    return { selected, topCandidates };
+  });
+}
+
+function cloneTeacherForcedReferenceScore(
+  value: TeacherForcedReferenceScoreEvidence | null | undefined,
+  workloadId: BenchmarkWorkloadId,
+): TeacherForcedReferenceScoreEvidence | null {
+  if (workloadId !== 'state-drift-1k-v1') {
+    if (value != null) {
+      throw new Error('Teacher-forced reference scoring is only valid for state-drift-1k-v1.');
+    }
+    return null;
+  }
+  if (value == null) {
+    throw new Error('state-drift-1k-v1 requires a complete teacher-forced CPU reference score.');
+  }
+  const { fixture, score } = value;
+  if (
+    fixture.schemaVersion !== 1
+    || fixture.kind !== 'bonsai-state-drift-reference'
+    || fixture.workload.id !== 'state-drift-1k-v1'
+    || fixture.tokenEncoding !== 'uint32-le'
+    || fixture.promptTokenIds.length !== 38
+    || fixture.referenceTokenIds.length !== 1_024
+  ) {
+    throw new Error('state-drift-1k-v1 received an invalid CPU reference fixture.');
+  }
+  const expectedMethod: ScoreSequenceResult['method'] = {
+    promptMode: 'raw-token-id-prefix',
+    maxTokensPerStep: 1,
+    temperature: 0,
+    topK: 1,
+    reportedTopLogprobs: 5,
+    logitBias: 1_000,
+    cachePromptFirst: false,
+    cachePromptSubsequent: true,
+  };
+  if (JSON.stringify(score.method) !== JSON.stringify(expectedMethod)) {
+    throw new Error('state-drift-1k-v1 teacher-forced method metadata is not locked.');
+  }
+  if (score.entries.length !== 1_024 || score.summary.tokenCount !== 1_024) {
+    throw new Error('state-drift-1k-v1 requires exactly 1,024 teacher-forced score entries.');
+  }
+  const entries = score.entries.map((entry, index) => {
+    const referenceTokenId = fixture.referenceTokenIds[index];
+    if (entry.index !== index || entry.selectedReference.id !== referenceTokenId) {
+      throw new Error(`Teacher-forced score entry ${index} does not select its CPU reference token.`);
+    }
+    if (!Number.isFinite(entry.selectedReference.logprob)) {
+      throw new Error(`Teacher-forced score entry ${index} has a non-finite reference logprob.`);
+    }
+    if (!Array.isArray(entry.topCandidates) || entry.topCandidates.length !== 5) {
+      throw new Error(`Teacher-forced score entry ${index} must contain five natural candidates.`);
+    }
+    const seenIds = new Set<number>();
+    const topCandidates = entry.topCandidates.map((candidate, candidateIndex) => {
+      if (
+        !Number.isSafeInteger(candidate.id)
+        || candidate.id < 0
+        || !Number.isFinite(candidate.logprob)
+        || seenIds.has(candidate.id)
+      ) {
+        throw new Error(`Teacher-forced score entry ${index} candidate ${candidateIndex} is invalid.`);
+      }
+      seenIds.add(candidate.id);
+      const previous = entry.topCandidates[candidateIndex - 1];
+      if (previous && (
+        previous.logprob < candidate.logprob
+        || (previous.logprob === candidate.logprob && previous.id > candidate.id)
+      )) {
+        throw new Error(`Teacher-forced score entry ${index} candidates are not sorted.`);
+      }
+      return { ...candidate };
+    });
+    if (
+      entry.naturalTop1.id !== topCandidates[0]?.id
+      || entry.naturalTop1.logprob !== topCandidates[0]?.logprob
+    ) {
+      throw new Error(`Teacher-forced score entry ${index} has inconsistent natural top-1 data.`);
+    }
+    const referenceRank = topCandidates.findIndex((candidate) => candidate.id === referenceTokenId);
+    const normalizedReferenceRank = referenceRank === -1 ? null : referenceRank;
+    if (entry.referenceRankInTopCandidatesZeroBased !== normalizedReferenceRank) {
+      throw new Error(`Teacher-forced score entry ${index} has an invalid reference rank.`);
+    }
+    const margin = topCandidates[0]!.logprob - topCandidates[1]!.logprob;
+    if (!Number.isFinite(entry.top1Top2Margin) || Math.abs(entry.top1Top2Margin - margin) > 1e-12) {
+      throw new Error(`Teacher-forced score entry ${index} has an invalid top-1/top-2 margin.`);
+    }
+    return {
+      index,
+      selectedReference: { ...entry.selectedReference },
+      naturalTop1: { ...entry.naturalTop1 },
+      topCandidates,
+      referenceRankInTopCandidatesZeroBased: normalizedReferenceRank,
+      top1Top2Margin: entry.top1Top2Margin,
+    };
+  });
+  const meanNll = -entries.reduce(
+    (sum, entry) => sum + entry.selectedReference.logprob,
+    0,
+  ) / entries.length;
+  const perplexity = Math.exp(meanNll);
+  if (
+    !Number.isFinite(score.summary.meanNll)
+    || !Number.isFinite(score.summary.perplexity)
+    || Math.abs(score.summary.meanNll - meanNll) > 1e-12
+    || Math.abs(score.summary.perplexity - perplexity) > Math.max(1, perplexity) * 1e-12
+  ) {
+    throw new Error('Teacher-forced score summary does not match its reference logprobs.');
+  }
+
+  return {
+    fixture: {
+      schemaVersion: 1,
+      kind: 'bonsai-state-drift-reference',
+      workload: {
+        id: 'state-drift-1k-v1',
+        messages: fixture.workload.messages.map((message) => ({ ...message })),
+      },
+      provenance: {
+        sourceEvidence: fixture.provenance.sourceEvidence,
+        engineRevision: fixture.provenance.engineRevision,
+        nativeBinary: { ...fixture.provenance.nativeBinary },
+        model: { ...fixture.provenance.model },
+        renderedPromptSha256: fixture.provenance.renderedPromptSha256,
+        execution: { ...fixture.provenance.execution },
+      },
+      tokenEncoding: 'uint32-le',
+      promptTokenIds: [...fixture.promptTokenIds],
+      promptTokenIdsSha256: fixture.promptTokenIdsSha256,
+      referenceTokenIds: [...fixture.referenceTokenIds],
+      referenceTokenIdsSha256: fixture.referenceTokenIdsSha256,
+      checkpointPrefixes: fixture.checkpointPrefixes.map((checkpoint) => ({ ...checkpoint })),
+    },
+    score: {
+      method: { ...expectedMethod },
+      entries,
+      summary: {
+        tokenCount: 1_024,
+        meanNll: score.summary.meanNll,
+        perplexity: score.summary.perplexity,
+      },
+    },
+  };
+}
+
 export function buildBenchmarkReport(observation: BenchmarkObservation): BenchmarkReport {
   const { generationResult, capabilities, backendReport, loadResult } = observation;
+  const workload = BENCHMARK_WORKLOADS[observation.workloadId];
+  const generationRequest = observation.generationRequest;
+  const maxTokens = generationRequest.maxTokens ?? 512;
+  const temperature = generationRequest.temperature ?? 0;
+  const topP = generationRequest.topP ?? null;
+  const topK = generationRequest.topK ?? 1;
+  const minP = generationRequest.minP ?? null;
+  const seed = generationRequest.seed ?? 42;
+  const toolChoice = generationRequest.toolChoice ?? null;
+  const cachePrompt = generationRequest.cachePrompt ?? true;
+  if (generationRequest.returnTokenIds !== true) {
+    throw new Error('Benchmark reports require returnTokenIds=true.');
+  }
+  if (JSON.stringify(generationRequest.messages) !== JSON.stringify(workload.messages)) {
+    throw new Error(`Benchmark messages do not match workload ${workload.id}.`);
+  }
+  if (generationResult.tokenIds === null) {
+    throw new Error('Benchmark generation completed without a token-id trace.');
+  }
+  const tokenIds = [...generationResult.tokenIds];
+  for (const [index, tokenId] of tokenIds.entries()) {
+    if (!Number.isSafeInteger(tokenId) || tokenId < 0 || tokenId > 0xffff_ffff) {
+      throw new Error(`Benchmark token id ${index} is invalid.`);
+    }
+  }
+  const tokenTrace = cloneValidatedTokenTrace(generationResult.tokenTrace, tokenIds);
+  if (typeof generationResult.reasoningText !== 'string') {
+    throw new Error('Benchmark generation completed without a reasoning-text channel.');
+  }
+  const checkpointPrefixes = buildCheckpointPrefixes(tokenIds);
+  if (workload.requiredMaxTokens !== null) {
+    if (
+      maxTokens !== workload.requiredMaxTokens
+      || temperature !== 0
+      || topP !== 1
+      || topK !== 1
+      || minP !== null
+      || seed !== 42
+      || toolChoice !== 'none'
+      || cachePrompt !== false
+    ) {
+      throw new Error(`${workload.id} requires its locked greedy sampling configuration.`);
+    }
+  }
+  if (workload.id === 'state-drift-1k-v1') {
+    if (tokenIds.length !== workload.requiredMaxTokens || checkpointPrefixes.length !== TOKEN_ID_CHECKPOINTS.length) {
+      throw new Error(`${workload.id} requires exactly 1,024 sampled token ids and all checkpoint hashes.`);
+    }
+    if (generationResult.finishReason !== 'length') {
+      throw new Error(`${workload.id} requires finishReason=length after exhausting its 1,024-token budget.`);
+    }
+  }
   const usage = generationResult.usage;
   const engineCompletionTokens = usage
     && Number.isFinite(usage.completionTokens)
@@ -152,10 +517,47 @@ export function buildBenchmarkReport(observation: BenchmarkObservation): Benchma
   const streamedTokenEvents = Number.isFinite(observation.streamedTokenEvents)
     ? Math.floor(Math.max(0, observation.streamedTokenEvents))
     : 0;
-  const completionTokens = Math.max(engineCompletionTokens ?? 0, streamedTokenEvents);
+  const completionTokens = tokenIds.length;
   const promptTokens = usage && Number.isFinite(usage.promptTokens) && usage.promptTokens >= 0
     ? Math.floor(usage.promptTokens)
     : null;
+  if (workload.id === 'context-prefill-8k-v1') {
+    const tuning = loadResult.tuning;
+    const exactRuntimePolicy = observation.model.id === '27b'
+      && loadResult.modelId === '27b'
+      && observation.contextSize === 8_448
+      && loadResult.context.size === 8_448
+      && observation.requestedBackend === 'webgpu'
+      && loadResult.backend === 'webgpu'
+      && tuning.scope === 'benchmark'
+      && tuning.requested.flashMode === 'auto'
+      && tuning.requested.cacheTypeK === 'q4_0'
+      && tuning.requested.cacheTypeV === 'q4_0'
+      && tuning.observed.flashAttention === true
+      && tuning.observed.cacheTypeK === 'q4_0'
+      && tuning.observed.cacheTypeV === 'q4_0'
+      && tuning.observed.kvBufferBytes !== null
+      && tuning.observed.kvBufferBytes > 0
+      && tuning.applied;
+    if (!exactRuntimePolicy) {
+      throw new Error(
+        `${workload.id} requires Bonsai 27B at context 8,448 with explicit WebGPU benchmark Flash auto and applied Q4_0 K/V.`,
+      );
+    }
+    if (tokenIds.length !== 8) {
+      throw new Error(`${workload.id} requires exactly 8 sampled token ids.`);
+    }
+    if (generationResult.finishReason !== 'length') {
+      throw new Error(`${workload.id} requires finishReason=length after exhausting its 8-token budget.`);
+    }
+    if (promptTokens === null || promptTokens <= 8_192) {
+      throw new Error(`${workload.id} requires more than 8,192 engine-counted prompt tokens.`);
+    }
+  }
+  const teacherForcedReferenceScore = cloneTeacherForcedReferenceScore(
+    observation.teacherForcedReferenceScore,
+    workload.id,
+  );
   const engineTotalTokens = usage && Number.isFinite(usage.totalTokens) && usage.totalTokens >= 0
     ? Math.floor(usage.totalTokens)
     : null;
@@ -172,7 +574,7 @@ export function buildBenchmarkReport(observation: BenchmarkObservation): Benchma
   const dflashPairing = observation.manifest.dflash.pairings[observation.model.id];
 
   return {
-    schemaVersion: 2,
+    schemaVersion: 3,
     kind: 'bonsai-browser-benchmark',
     startedAt: observation.startedAt,
     completedAt: observation.completedAt,
@@ -180,6 +582,8 @@ export function buildBenchmarkReport(observation: BenchmarkObservation): Benchma
       id: observation.model.id,
       displayName: observation.model.displayName,
       architecture: observation.model.architecture,
+      source: { ...observation.model.source },
+      shards: observation.model.files.map(({ path, bytes, sha256 }) => ({ path, bytes, sha256 })),
       downloadBytes: observation.model.downloadBytes,
       contextSize: observation.contextSize,
     },
@@ -202,11 +606,33 @@ export function buildBenchmarkReport(observation: BenchmarkObservation): Benchma
       },
     },
     generation: {
-      promptId: BENCHMARK_PROMPT_ID,
-      prompt: BENCHMARK_PROMPT,
-      maxTokens: observation.maxTokens,
-      temperature: 0,
-      seed: 42,
+      workload: {
+        id: workload.id,
+        messages: cloneMessages(generationRequest.messages),
+      },
+      sampling: {
+        maxTokens,
+        temperature,
+        topP,
+        topK,
+        minP,
+        seed,
+        toolChoice,
+        cachePrompt,
+      },
+      tokenTrace: {
+        returnTokenIds: true,
+        logprobs: true,
+        topLogprobs: 5,
+        topAlternatives: 4,
+        encoding: 'uint32-le',
+        tokenIds,
+        entries: tokenTrace,
+        checkpointPrefixes,
+      },
+      teacherForcedReferenceScore,
+      rawText: generationResult.text,
+      rawReasoningText: generationResult.reasoningText,
       elapsedMs: roundedNonNegative(observation.generationElapsedMs, 1),
       timeToFirstTokenMs: roundedNonNegative(observation.timeToFirstTokenMs, 1),
       promptTokensPerSecond: roundedNonNegative(
@@ -220,9 +646,7 @@ export function buildBenchmarkReport(observation: BenchmarkObservation): Benchma
       completionTokens,
       engineTotalTokens,
       totalTokens,
-      tokenCountSource: engineCompletionTokens !== null && engineCompletionTokens >= streamedTokenEvents
-        ? 'engine-usage'
-        : 'stream-events-fallback',
+      tokenCountSource: 'token-id-trace',
       finishReason: generationResult.finishReason,
     },
     execution: {

src/bench/state-drift-reference.test.ts

@@ -0,0 +1,81 @@
+import { describe, expect, it } from 'vitest';
+import fixtureJson from '../../public/fixtures/state-drift-27b-cpu-reference.json';
+import type {
+  EngineCapabilities,
+  LoadModelResult,
+  ManifestModelV2,
+} from '../engine';
+import { validateStateDriftReferenceFixture } from './state-drift-reference';
+
+const fixture: unknown = fixtureJson;
+
+const model = {
+  id: '27b',
+  source: {
+    repo: 'WaveCut/Bonsai-web-GGUF',
+    revision: 'd85382fa09fe868c0242d81488dfc2edd8d3729b',
+    file: 'Bonsai-27B-Q1_0.gguf',
+    bytes: 3_803_452_480,
+    sha256: '17ef842e47450caeb8eaa3ebfbbab5d2f2278b62b79be107985fb69a2f819aa0',
+  },
+} as ManifestModelV2;
+
+const capabilities = {
+  runtime: {
+    llamaCppRevision: '00fa7cb284cbf133fc426733bd64238a3588a33e',
+  },
+} as EngineCapabilities;
+
+const loadResult = {
+  modelId: '27b',
+  backend: 'webgpu',
+  gate: { requestedBackend: 'webgpu' },
+  context: {
+    size: 2_048,
+    batchSize: 32,
+    microBatchSize: 16,
+    vocabularySize: 248_320,
+  },
+  tuning: {
+    scope: 'benchmark',
+    requested: { nBatch: 32, nUbatch: 16 },
+    observed: { nBatch: 32, nUbatch: 16 },
+    applied: true,
+  },
+} as LoadModelResult;
+
+describe('state-drift CPU reference fixture', () => {
+  it('validates the pinned fixture against the loaded model, runtime, context, and batching', () => {
+    const result = validateStateDriftReferenceFixture(fixture, {
+      model,
+      capabilities,
+      loadResult,
+    });
+
+    expect(result.promptTokenIds).toHaveLength(38);
+    expect(result.referenceTokenIds).toHaveLength(1_024);
+    expect(result.referenceTokenIdsSha256).toBe(
+      'c503b2db0dcf10daecfda4f19a5f466d1699b31688076d6c32f7c29daefcef9b',
+    );
+    expect(result.checkpointPrefixes).toHaveLength(6);
+  });
+
+  it('rejects provenance or token-hash drift before scoring', () => {
+    const wrongModel = structuredClone(fixture) as Record<string, unknown>;
+    const provenance = wrongModel.provenance as { model: { sha256: string } };
+    provenance.model.sha256 = '0'.repeat(64);
+    expect(() => validateStateDriftReferenceFixture(wrongModel, {
+      model,
+      capabilities,
+      loadResult,
+    })).toThrow('model source sha256');
+
+    const wrongTokens = structuredClone(fixture) as { referenceTokenIds: number[] };
+    wrongTokens.referenceTokenIds[29] = wrongTokens.referenceTokenIds[29]! + 1;
+    expect(() => validateStateDriftReferenceFixture(wrongTokens, {
+      model,
+      capabilities,
+      loadResult,
+    })).toThrow('referenceTokenIdsSha256');
+  });
+});

src/bench/state-drift-reference.ts

@@ -0,0 +1,272 @@
+import { sha256 } from '@noble/hashes/sha2.js';
+import { bytesToHex } from '@noble/hashes/utils.js';
+import type {
+  EngineCapabilities,
+  LoadModelResult,
+  ManifestModelV2,
+} from '../engine';
+import { BENCHMARK_WORKLOADS, TOKEN_ID_CHECKPOINTS } from './report';
+
+export const STATE_DRIFT_REFERENCE_PATH = 'fixtures/state-drift-27b-cpu-reference.json';
+const STATE_DRIFT_SOURCE_EVIDENCE = 'results/space-model/27b-native-cpu-state-drift-1024-b32-u16.json';
+const STATE_DRIFT_RENDERED_PROMPT_SHA256 = '7c46d0de443e6f1235ddaa7c0a55c9da710eb73958d5671247b6f0e3b7d189c7';
+const STATE_DRIFT_PROMPT_TOKEN_IDS_SHA256 = 'e7af7d66c96f24cb148575bd45cef079920669eb0d193fdc830ce61dfec85451';
+const STATE_DRIFT_REFERENCE_TOKEN_IDS_SHA256 = 'c503b2db0dcf10daecfda4f19a5f466d1699b31688076d6c32f7c29daefcef9b';
+
+export interface StateDriftReferenceFixture {
+  schemaVersion: 1;
+  kind: 'bonsai-state-drift-reference';
+  workload: {
+    id: 'state-drift-1k-v1';
+    messages: Array<{ role: 'user'; content: string }>;
+  };
+  provenance: {
+    sourceEvidence: string;
+    engineRevision: string;
+    nativeBinary: { bytes: number; sha256: string };
+    model: { file: string; bytes: number; sha256: string };
+    renderedPromptSha256: string;
+    execution: {
+      backend: 'cpu';
+      contextSize: number;
+      batchSize: number;
+      microBatchSize: number;
+    };
+  };
+  tokenEncoding: 'uint32-le';
+  promptTokenIds: number[];
+  promptTokenIdsSha256: string;
+  referenceTokenIds: number[];
+  referenceTokenIdsSha256: string;
+  checkpointPrefixes: Array<{ tokens: number; sha256: string }>;
+}
+
+export interface StateDriftReferenceContext {
+  model: ManifestModelV2;
+  capabilities: EngineCapabilities;
+  loadResult: LoadModelResult;
+}
+
+function fail(message: string): never {
+  throw new Error(`Invalid state-drift CPU reference: ${message}`);
+}
+
+function record(value: unknown, label: string): Record<string, unknown> {
+  if (typeof value !== 'object' || value === null || Array.isArray(value)) {
+    fail(`${label} must be an object.`);
+  }
+  return value as Record<string, unknown>;
+}
+
+function exact<T>(value: unknown, expected: T, label: string): T {
+  if (!Object.is(value, expected)) {
+    fail(`${label} must be ${JSON.stringify(expected)}; received ${JSON.stringify(value)}.`);
+  }
+  return expected;
+}
+
+function nonEmptyString(value: unknown, label: string): string {
+  if (typeof value !== 'string' || value.length === 0) fail(`${label} must be a non-empty string.`);
+  return value;
+}
+
+function positiveInteger(value: unknown, label: string): number {
+  if (!Number.isSafeInteger(value) || (value as number) <= 0) {
+    fail(`${label} must be a positive safe integer.`);
+  }
+  return value as number;
+}
+
+function sha256Digest(value: unknown, label: string): string {
+  const digest = nonEmptyString(value, label).toLowerCase();
+  if (!/^[0-9a-f]{64}$/.test(digest)) fail(`${label} must be a 64-character SHA256 digest.`);
+  return digest;
+}
+
+function gitRevision(value: unknown, label: string): string {
+  const revision = nonEmptyString(value, label).toLowerCase();
+  if (!/^[0-9a-f]{40}$/.test(revision)) fail(`${label} must be a pinned 40-character revision.`);
+  return revision;
+}
+
+function tokenIds(
+  value: unknown,
+  expectedLength: number,
+  vocabularySize: number,
+  label: string,
+): number[] {
+  if (!Array.isArray(value) || value.length !== expectedLength) {
+    fail(`${label} must contain exactly ${expectedLength} token ids.`);
+  }
+  return value.map((tokenId, index) => {
+    if (!Number.isSafeInteger(tokenId) || tokenId < 0 || tokenId >= vocabularySize) {
+      fail(`${label}[${index}] is outside the loaded vocabulary.`);
+    }
+    return tokenId;
+  });
+}
+
+function tokenIdsSha256(values: readonly number[], count = values.length): string {
+  const bytes = new Uint8Array(count * Uint32Array.BYTES_PER_ELEMENT);
+  const view = new DataView(bytes.buffer);
+  for (let index = 0; index < count; index += 1) {
+    view.setUint32(index * Uint32Array.BYTES_PER_ELEMENT, values[index]!, true);
+  }
+  return bytesToHex(sha256(bytes));
+}
+
+export function validateStateDriftReferenceFixture(
+  value: unknown,
+  context: StateDriftReferenceContext,
+): StateDriftReferenceFixture {
+  const root = record(value, 'root');
+  exact(root.schemaVersion, 1, 'schemaVersion');
+  exact(root.kind, 'bonsai-state-drift-reference', 'kind');
+
+  const workload = record(root.workload, 'workload');
+  exact(workload.id, 'state-drift-1k-v1', 'workload.id');
+  const expectedMessages = BENCHMARK_WORKLOADS['state-drift-1k-v1'].messages;
+  if (JSON.stringify(workload.messages) !== JSON.stringify(expectedMessages)) {
+    fail('workload.messages do not match the locked state-drift prompt.');
+  }
+
+  const { model, capabilities, loadResult } = context;
+  if (
+    model.id !== '27b'
+    || loadResult.modelId !== '27b'
+    || loadResult.backend !== 'webgpu'
+    || loadResult.gate.requestedBackend !== 'webgpu'
+    || loadResult.tuning.scope !== 'benchmark'
+    || !loadResult.tuning.applied
+  ) {
+    fail('the loaded runtime is not the explicit 27B WebGPU benchmark path.');
+  }
+
+  const provenance = record(root.provenance, 'provenance');
+  const sourceEvidence = nonEmptyString(provenance.sourceEvidence, 'provenance.sourceEvidence');
+  exact(sourceEvidence, STATE_DRIFT_SOURCE_EVIDENCE, 'provenance.sourceEvidence');
+  const engineRevision = gitRevision(provenance.engineRevision, 'provenance.engineRevision');
+  exact(engineRevision, capabilities.runtime.llamaCppRevision, 'runtime llama.cpp revision');
+
+  const nativeBinary = record(provenance.nativeBinary, 'provenance.nativeBinary');
+  const nativeBinaryBytes = positiveInteger(nativeBinary.bytes, 'provenance.nativeBinary.bytes');
+  const nativeBinarySha256 = sha256Digest(
+    nativeBinary.sha256,
+    'provenance.nativeBinary.sha256',
+  );
+
+  const fixtureModel = record(provenance.model, 'provenance.model');
+  const modelFile = nonEmptyString(fixtureModel.file, 'provenance.model.file');
+  const modelBytes = positiveInteger(fixtureModel.bytes, 'provenance.model.bytes');
+  const modelSha256 = sha256Digest(fixtureModel.sha256, 'provenance.model.sha256');
+  exact(modelFile, model.source.file, 'model source file');
+  exact(modelBytes, model.source.bytes, 'model source bytes');
+  exact(modelSha256, model.source.sha256, 'model source sha256');
+
+  const renderedPromptSha256 = sha256Digest(
+    provenance.renderedPromptSha256,
+    'provenance.renderedPromptSha256',
+  );
+  exact(
+    renderedPromptSha256,
+    STATE_DRIFT_RENDERED_PROMPT_SHA256,
+    'provenance.renderedPromptSha256',
+  );
+  const execution = record(provenance.execution, 'provenance.execution');
+  exact(execution.backend, 'cpu', 'provenance.execution.backend');
+  const contextSize = positiveInteger(execution.contextSize, 'provenance.execution.contextSize');
+  const batchSize = positiveInteger(execution.batchSize, 'provenance.execution.batchSize');
+  const microBatchSize = positiveInteger(
+    execution.microBatchSize,
+    'provenance.execution.microBatchSize',
+  );
+  exact(contextSize, loadResult.context.size, 'reference/loaded context size');
+  exact(batchSize, loadResult.context.batchSize, 'reference/loaded batch size');
+  exact(microBatchSize, loadResult.context.microBatchSize, 'reference/loaded micro-batch size');
+  exact(loadResult.tuning.requested.nBatch, batchSize, 'requested reference batch size');
+  exact(loadResult.tuning.requested.nUbatch, microBatchSize, 'requested reference micro-batch size');
+  exact(loadResult.tuning.observed.nBatch, batchSize, 'observed reference batch size');
+  exact(loadResult.tuning.observed.nUbatch, microBatchSize, 'observed reference micro-batch size');
+
+  exact(root.tokenEncoding, 'uint32-le', 'tokenEncoding');
+  const promptTokenIds = tokenIds(
+    root.promptTokenIds,
+    38,
+    loadResult.context.vocabularySize,
+    'promptTokenIds',
+  );
+  const promptTokenIdsSha256 = sha256Digest(root.promptTokenIdsSha256, 'promptTokenIdsSha256');
+  exact(promptTokenIdsSha256, tokenIdsSha256(promptTokenIds), 'promptTokenIdsSha256');
+  exact(
+    promptTokenIdsSha256,
+    STATE_DRIFT_PROMPT_TOKEN_IDS_SHA256,
+    'pinned promptTokenIdsSha256',
+  );
+  const referenceTokenIds = tokenIds(
+    root.referenceTokenIds,
+    1_024,
+    loadResult.context.vocabularySize,
+    'referenceTokenIds',
+  );
+  const referenceTokenIdsSha256 = sha256Digest(
+    root.referenceTokenIdsSha256,
+    'referenceTokenIdsSha256',
+  );
+  exact(referenceTokenIdsSha256, tokenIdsSha256(referenceTokenIds), 'referenceTokenIdsSha256');
+  exact(
+    referenceTokenIdsSha256,
+    STATE_DRIFT_REFERENCE_TOKEN_IDS_SHA256,
+    'pinned referenceTokenIdsSha256',
+  );
+
+  if (!Array.isArray(root.checkpointPrefixes)
+    || root.checkpointPrefixes.length !== TOKEN_ID_CHECKPOINTS.length) {
+    fail(`checkpointPrefixes must contain exactly ${TOKEN_ID_CHECKPOINTS.length} entries.`);
+  }
+  const checkpointPrefixes = root.checkpointPrefixes.map((rawCheckpoint, index) => {
+    const checkpoint = record(rawCheckpoint, `checkpointPrefixes[${index}]`);
+    const tokens = TOKEN_ID_CHECKPOINTS[index]!;
+    exact(checkpoint.tokens, tokens, `checkpointPrefixes[${index}].tokens`);
+    const embeddedSha256 = sha256Digest(
+      checkpoint.sha256,
+      `checkpointPrefixes[${index}].sha256`,
+    );
+    exact(
+      embeddedSha256,
+      tokenIdsSha256(referenceTokenIds, tokens),
+      `checkpointPrefixes[${index}].sha256`,
+    );
+    return { tokens, sha256: embeddedSha256 };
+  });
+
+  return {
+    schemaVersion: 1,
+    kind: 'bonsai-state-drift-reference',
+    workload: {
+      id: 'state-drift-1k-v1',
+      messages: expectedMessages.map((message) => ({ ...message })) as Array<{
+        role: 'user';
+        content: string;
+      }>,
+    },
+    provenance: {
+      sourceEvidence,
+      engineRevision,
+      nativeBinary: { bytes: nativeBinaryBytes, sha256: nativeBinarySha256 },
+      model: { file: modelFile, bytes: modelBytes, sha256: modelSha256 },
+      renderedPromptSha256,
+      execution: {
+        backend: 'cpu',
+        contextSize,
+        batchSize,
+        microBatchSize,
+      },
+    },
+    tokenEncoding: 'uint32-le',
+    promptTokenIds,
+    promptTokenIdsSha256,
+    referenceTokenIds,
+    referenceTokenIdsSha256,
+    checkpointPrefixes,
+  };
+}

src/engine/client.test.ts

@@ -231,6 +231,50 @@ describe('BrowserEngineClient worker lifecycle', () => {
     client.close();
   });
 
+  it('routes and softly aborts diagnostic sequence scoring', async () => {
+    const client = new BrowserEngineClient();
+    const worker = latestWorker();
+    const controller = new AbortController();
+    const scoring = client.scoreSequence({
+      promptTokenIds: Array.from({ length: 38 }, (_, index) => index + 1),
+      referenceTokenIds: Array.from({ length: 1_024 }, (_, index) => index + 1_000),
+      topK: 5,
+    }, {
+      requestId: 'soft-score-sequence',
+      signal: controller.signal,
+    });
+    const scoringFailure = expect(scoring).rejects.toMatchObject({ code: 'ABORTED' });
+
+    expect(worker.postMessage).toHaveBeenCalledWith(expect.objectContaining({
+      requestId: 'soft-score-sequence',
+      method: 'scoreSequence',
+    }));
+    controller.abort();
+    const abortRequest = worker.postMessage.mock.calls
+      .map((call) => call[0] as { requestId: string; method: string })
+      .find((request) => request.method === 'abort');
+    expect(abortRequest).toBeTruthy();
+    worker.send({
+      type: 'response',
+      requestId: 'soft-score-sequence',
+      method: 'scoreSequence',
+      ok: false,
+      error: { code: 'ABORTED', message: 'Sequence scoring stopped.' },
+    });
+    if (!abortRequest) throw new Error('Expected abort request.');
+    worker.send({
+      type: 'response',
+      requestId: abortRequest.requestId,
+      method: 'abort',
+      ok: true,
+      result: { targetRequestId: 'soft-score-sequence', aborted: true },
+    });
+
+    await scoringFailure;
+    expect(worker.terminate).not.toHaveBeenCalled();
+    client.close();
+  });
+
   it('invalidates the model when generation ignores abort beyond the grace period', async () => {
     vi.useFakeTimers();
     try {

src/engine/client.ts

@@ -9,6 +9,7 @@ import type {
   EngineWorkerMessage,
   GenerateParams,
   LoadModelParams,
+  ScoreSequenceParams,
 } from './protocol';
 
 type ProgressEvent = Extract<EngineEvent, { event: 'progress' }>;
@@ -134,16 +135,20 @@ export class BrowserEngineClient {
                   ));
                 }
               }, LOAD_ABORT_CLEANUP_TIMEOUT_MS);
-            } else if (method === 'generate') {
+            } else if (method === 'generate' || method === 'scoreSequence') {
+              const operationLabel = method === 'generate' ? 'Generation' : 'Sequence scoring';
+              const cause = method === 'generate'
+                ? 'generation-abort-timeout'
+                : 'score-sequence-abort-timeout';
               pending.abortFallbackTimer = setTimeout(() => {
                 if (this.pending.has(requestId)) {
                   this.restart(new EngineClientError({
                     code: 'ENGINE_WORKER_FAILED',
-                    message: 'Generation did not stop within 5 seconds. The browser engine worker was restarted and the loaded model was invalidated.',
+                    message: `${operationLabel} did not stop within 5 seconds. The browser engine worker was restarted and the loaded model was invalidated.`,
                     details: {
                       recoverable: true,
                       nextAction: 'reload-model',
-                      cause: 'generation-abort-timeout',
+                      cause,
                       graceMs: GENERATION_ABORT_GRACE_MS,
                     },
                   }));
@@ -176,6 +181,10 @@ export class BrowserEngineClient {
     return this.request('generate', params, options);
   }
 
+  scoreSequence(params: ScoreSequenceParams, options?: EngineRequestOptions) {
+    return this.request('scoreSequence', params, options);
+  }
+
   abort(targetRequestId: string) {
     return this.request('abort', { targetRequestId });
   }

src/engine/device-gate.test.ts

@@ -1,6 +1,7 @@
 import { afterEach, describe, expect, it, vi } from 'vitest';
 import {
   assertBackendPolicy,
+  evaluateModelContextPolicy,
   evaluateModelGate,
   inspectWebGpuAdapter,
   supportsWllamaWebGpuRuntime,
@@ -164,6 +165,52 @@ describe('evaluateModelGate', () => {
   });
 });
 
+describe('evaluateModelContextPolicy', () => {
+  const model27b = model('27b', false);
+  const longContextExperiment = {
+    tuningScope: 'benchmark',
+    requestedBackend: 'webgpu',
+    flashMode: 'auto',
+    kvCacheType: 'q4_0',
+  } as const;
+
+  it('allows the bounded 8,448-token 27B benchmark experiment', () => {
+    expect(evaluateModelContextPolicy(model27b, 8_448, longContextExperiment)).toEqual({
+      allowed: true,
+      limit: 8_448,
+    });
+  });
+
+  it('rejects 8,449 tokens even for the exact long-context experiment', () => {
+    expect(evaluateModelContextPolicy(model27b, 8_449, longContextExperiment)).toEqual({
+      allowed: false,
+      limit: 8_448,
+    });
+  });
+
+  it.each([
+    ['q8 KV', { ...longContextExperiment, kvCacheType: 'q8_0' as const }],
+    ['f16 KV', { ...longContextExperiment, kvCacheType: 'f16' as const }],
+    ['auto backend', { ...longContextExperiment, requestedBackend: 'auto' as const }],
+    ['release defaults', { ...longContextExperiment, tuningScope: 'release-defaults' as const }],
+  ])('keeps 27B at 2,048 for %s', (_label, input) => {
+    expect(evaluateModelContextPolicy(model27b, 8_448, input)).toEqual({
+      allowed: false,
+      limit: 2_048,
+    });
+    expect(evaluateModelContextPolicy(model27b, 2_048, input).allowed).toBe(true);
+  });
+
+  it('leaves dense-tier manifest limits unchanged', () => {
+    expect(evaluateModelContextPolicy(model('8b', true), 32_768, {
+      tuningScope: 'release-defaults',
+      requestedBackend: 'auto',
+      flashMode: 'off',
+      kvCacheType: 'f16',
+    })).toEqual({ allowed: true, limit: 32_768 });
+  });
+});
+
 describe('assertBackendPolicy', () => {
   const strictModel = model('27b', false);
   const baseReport = {

src/engine/device-gate.ts

@@ -1,5 +1,7 @@
 import type { ManifestModelV2 } from './manifest';
 import type {
+  BenchmarkFlashMode,
+  BenchmarkKvCacheType,
   GateDecision,
   RequestedBackend,
   RuntimeBackend,
@@ -8,6 +10,21 @@ import type {
 } from './protocol';
 import type { BackendReport } from './native-log';
 
+export const MODEL_27B_RELEASE_CONTEXT_LIMIT = 2_048;
+export const MODEL_27B_BENCHMARK_CONTEXT_LIMIT = 8_448;
+
+export interface ModelContextPolicyInput {
+  tuningScope: 'release-defaults' | 'benchmark';
+  requestedBackend: RequestedBackend;
+  flashMode: BenchmarkFlashMode;
+  kvCacheType: BenchmarkKvCacheType;
+}
+
+export interface ModelContextPolicyDecision {
+  allowed: boolean;
+  limit: number;
+}
+
 const KNOWN_LIMITS = [
   'maxBindGroups',
   'maxBindingsPerBindGroup',
@@ -161,6 +178,30 @@ function availableStorageBytes(storage: StorageEstimate, cachedModelBytes: numbe
   return Math.max(0, storage.quotaBytes - storage.usageBytes + cachedModelBytes);
 }
 
+export function evaluateModelContextPolicy(
+  model: ManifestModelV2,
+  contextSize: number,
+  input: ModelContextPolicyInput,
+): ModelContextPolicyDecision {
+  const longContextExperiment = model.id === '27b'
+    && input.tuningScope === 'benchmark'
+    && input.requestedBackend === 'webgpu'
+    && input.flashMode === 'auto'
+    && input.kvCacheType === 'q4_0';
+  const limit = model.id === '27b'
+    ? Math.min(
+        model.contextLength,
+        longContextExperiment
+          ? MODEL_27B_BENCHMARK_CONTEXT_LIMIT
+          : MODEL_27B_RELEASE_CONTEXT_LIMIT,
+      )
+    : model.contextLength;
+  return {
+    allowed: Number.isSafeInteger(contextSize) && contextSize > 0 && contextSize <= limit,
+    limit,
+  };
+}
+
 function webGpuFailures(model: ManifestModelV2, adapter: WebGpuAdapterSnapshot): string[] {
   if (!adapter.available) {
     return ['WebGPU adapter is unavailable.'];

src/engine/errors.ts

@@ -1,7 +1,7 @@
 export interface WebGpuDeviceLostDetails {
   recoverable: true;
   nextAction: 'reload-model';
-  stage: 'load' | 'generate' | 'backend-report';
+  stage: 'load' | 'generate' | 'score-sequence' | 'backend-report';
   modelId: string;
   cpuFallbackAvailable: boolean;
   signal: {

src/engine/protocol.ts

@@ -133,6 +133,7 @@ export interface GenerateParams {
   tools?: EngineChatTool[];
   toolChoice?: 'none' | 'auto' | 'required';
   cachePrompt?: boolean;
+  returnTokenIds?: boolean;
 }
 
 export interface LoadModelResult {
@@ -180,8 +181,21 @@ export interface LoadModelResult {
   backendReport: BackendReport;
 }
 
+export interface EngineTokenLogprob {
+  id: number;
+  logprob: number;
+}
+
+export interface EngineSampledTokenTraceEntry {
+  selected: EngineTokenLogprob;
+  topCandidates: EngineTokenLogprob[];
+}
+
 export interface GenerateResult {
   text: string;
+  reasoningText: string;
+  tokenIds: number[] | null;
+  tokenTrace: EngineSampledTokenTraceEntry[] | null;
   finishReason: 'stop' | 'length' | 'tool_calls' | 'content_filter' | null;
   toolCalls: EngineChatToolCall[];
   usage: {
@@ -195,6 +209,40 @@ export interface GenerateResult {
   } | null;
 }
 
+export interface ScoreSequenceParams {
+  promptTokenIds: number[];
+  referenceTokenIds: number[];
+  topK: 5;
+}
+
+export interface EngineTeacherForcedScoreEntry {
+  index: number;
+  selectedReference: EngineTokenLogprob;
+  naturalTop1: EngineTokenLogprob;
+  topCandidates: EngineTokenLogprob[];
+  referenceRankInTopCandidatesZeroBased: number | null;
+  top1Top2Margin: number;
+}
+
+export interface ScoreSequenceResult {
+  method: {
+    promptMode: 'raw-token-id-prefix';
+    maxTokensPerStep: 1;
+    temperature: 0;
+    topK: 1;
+    reportedTopLogprobs: 5;
+    logitBias: 1_000;
+    cachePromptFirst: false;
+    cachePromptSubsequent: true;
+  };
+  entries: EngineTeacherForcedScoreEntry[];
+  summary: {
+    tokenCount: 1_024;
+    meanNll: number;
+    perplexity: number;
+  };
+}
+
 export interface EmptyParams {
   readonly _empty?: never;
 }
@@ -203,6 +251,7 @@ export interface EngineRequestMap {
   capabilities: EmptyParams;
   loadModel: LoadModelParams;
   generate: GenerateParams;
+  scoreSequence: ScoreSequenceParams;
   abort: { targetRequestId: string };
   unload: EmptyParams;
   backendReport: EmptyParams;
@@ -215,6 +264,7 @@ export interface EngineResponseMap {
   capabilities: EngineCapabilities;
   loadModel: LoadModelResult;
   generate: GenerateResult;
+  scoreSequence: ScoreSequenceResult;
   abort: { targetRequestId: string; aborted: boolean };
   unload: { unloaded: true };
   backendReport: BackendReport;

src/engine/runtime-generate.test.ts

@@ -2,12 +2,16 @@ import { describe, expect, it } from 'vitest';
 import type { ChatCompletionChunk } from '../../vendor/wllama-bonsai/esm/index.js';
 import { BrowserEngineRuntime } from './runtime';
 
+interface CompletionOptions {
+  onData(chunk: ChatCompletionChunk): void;
+  logprobs?: boolean;
+  top_logprobs?: number;
+}
+
 interface RuntimeInternals {
   wllama: {
     isModelLoaded(): boolean;
-    createChatCompletion(options: {
-      onData(chunk: ChatCompletionChunk): void;
-    }): Promise<void>;
+    createChatCompletion(options: CompletionOptions): Promise<void>;
   } | null;
   loaded: {
     manifest: unknown;
@@ -44,6 +48,46 @@ function chunk(
   };
 }
 
+function tracedChunk(
+  content: string,
+  id: number,
+  metadata: Partial<Pick<ChatCompletionChunk, 'usage' | 'timings'>> = {},
+): ChatCompletionChunk {
+  const value = chunk(content, metadata);
+  const choice = value.choices[0];
+  if (!choice) throw new Error('Chunk fixture requires one choice.');
+  choice.logprobs = {
+    content: [{
+      id,
+      token: content,
+      logprob: -0.1,
+      bytes: null,
+      top_logprobs: [
+        { id: id + 1_004, token: 'fifth', logprob: -4, bytes: null },
+        { id: id + 1_002, token: 'third', logprob: -2, bytes: null },
+        { id, token: content, logprob: -0.1, bytes: null },
+        { id: id + 1_003, token: 'fourth', logprob: -3, bytes: null },
+        { id: id + 1_001, token: 'second', logprob: -1, bytes: null },
+      ],
+    }],
+    refusal: null,
+  };
+  return value;
+}
+
+function reasoningTracedChunk(
+  reasoningContent: string,
+  id: number,
+  metadata: Partial<Pick<ChatCompletionChunk, 'usage' | 'timings'>> = {},
+): ChatCompletionChunk {
+  const value = tracedChunk('', id, metadata);
+  const choice = value.choices[0];
+  if (!choice) throw new Error('Chunk fixture requires one choice.');
+  choice.finish_reason = null;
+  (choice.delta as unknown as Record<string, unknown>).reasoning_content = reasoningContent;
+  return value;
+}
+
 describe('BrowserEngineRuntime generation telemetry', () => {
   it('retains usage and timings when a trailing stream chunk omits metadata', async () => {
     const runtime = new BrowserEngineRuntime();
@@ -62,10 +106,12 @@ describe('BrowserEngineRuntime generation telemetry', () => {
         },
       },
     };
+    let completionOptions: CompletionOptions | null = null;
     internals.wllama = {
       isModelLoaded: () => true,
-      createChatCompletion: async ({ onData }) => {
-        onData(chunk('Ready', {
+      createChatCompletion: async (options) => {
+        completionOptions = options;
+        options.onData(chunk('Ready', {
           usage: { prompt_tokens: 9, completion_tokens: 1, total_tokens: 10 },
           timings: {
             cache_n: 0,
@@ -79,7 +125,7 @@ describe('BrowserEngineRuntime generation telemetry', () => {
             predicted_per_second: 50,
           },
         }));
-        onData(chunk(''));
+        options.onData(chunk(''));
       },
     };
 
@@ -92,6 +138,11 @@ describe('BrowserEngineRuntime generation telemetry', () => {
 
     expect(result.usage).toEqual({ promptTokens: 9, completionTokens: 1, totalTokens: 10 });
     expect(result.timings).toEqual({ promptTokensPerSecond: 100, predictedTokensPerSecond: 50 });
+    expect(result.reasoningText).toBe('');
+    expect(result.tokenIds).toBeNull();
+    expect(result.tokenTrace).toBeNull();
+    expect(completionOptions).not.toHaveProperty('logprobs');
+    expect(completionOptions).not.toHaveProperty('top_logprobs');
   });
 
   it('reconciles incomplete streamed usage with llama.cpp timing counts', async () => {
@@ -140,5 +191,310 @@ describe('BrowserEngineRuntime generation telemetry', () => {
     );
 
     expect(result.usage).toEqual({ promptTokens: 9, completionTokens: 64, totalTokens: 73 });
+    expect(result.tokenIds).toBeNull();
+    expect(result.tokenTrace).toBeNull();
+  });
+
+  it('returns sampled token ids with selected logprobs and sorted top-five candidates', async () => {
+    const runtime = new BrowserEngineRuntime();
+    const internals = runtime as unknown as RuntimeInternals;
+    let completionOptions: CompletionOptions | null = null;
+    internals.loaded = {
+      manifest: {},
+      backend: 'wasm',
+      model: {
+        id: '1_7b',
+        displayName: 'Fixture Bonsai',
+        cpuFallback: true,
+        runtimePolicy: {
+          flashAttention: false,
+          tokenEmbeddingOnWebGPU: true,
+          requireSingleWebGPUGraph: false,
+        },
+      },
+    };
+    internals.wllama = {
+      isModelLoaded: () => true,
+      createChatCompletion: async (options) => {
+        completionOptions = options;
+        options.onData(tracedChunk('one', 101));
+        options.onData(tracedChunk(' two', 202, {
+          usage: { prompt_tokens: 4, completion_tokens: 2, total_tokens: 6 },
+          timings: {
+            cache_n: 0,
+            prompt_n: 4,
+            prompt_ms: 40,
+            prompt_per_token_ms: 10,
+            prompt_per_second: 100,
+            predicted_n: 2,
+            predicted_ms: 40,
+            predicted_per_token_ms: 20,
+            predicted_per_second: 50,
+          },
+        }));
+        options.onData(chunk(''));
+      },
+    };
+
+    const result = await runtime.generate(
+      'request-token-ids',
+      { messages: [{ role: 'user', content: 'Count.' }], returnTokenIds: true },
+      new AbortController().signal,
+      () => undefined,
+    );
+
+    expect(result.tokenIds).toEqual([101, 202]);
+    expect(result.tokenTrace).toEqual([
+      {
+        selected: { id: 101, logprob: -0.1 },
+        topCandidates: [
+          { id: 101, logprob: -0.1 },
+          { id: 1_102, logprob: -1 },
+          { id: 1_103, logprob: -2 },
+          { id: 1_104, logprob: -3 },
+          { id: 1_105, logprob: -4 },
+        ],
+      },
+      {
+        selected: { id: 202, logprob: -0.1 },
+        topCandidates: [
+          { id: 202, logprob: -0.1 },
+          { id: 1_203, logprob: -1 },
+          { id: 1_204, logprob: -2 },
+          { id: 1_205, logprob: -3 },
+          { id: 1_206, logprob: -4 },
+        ],
+      },
+    ]);
+    expect(completionOptions).toHaveProperty('logprobs', true);
+    expect(completionOptions).toHaveProperty('top_logprobs', 5);
+  });
+
+  it('keeps reasoning-only output separate from visible text while preserving its token trace', async () => {
+    const runtime = new BrowserEngineRuntime();
+    const internals = runtime as unknown as RuntimeInternals;
+    internals.loaded = {
+      manifest: {},
+      backend: 'wasm',
+      model: {
+        id: '1_7b',
+        displayName: 'Fixture Bonsai',
+        cpuFallback: true,
+        runtimePolicy: {
+          flashAttention: false,
+          tokenEmbeddingOnWebGPU: true,
+          requireSingleWebGPUGraph: false,
+        },
+      },
+    };
+    internals.wllama = {
+      isModelLoaded: () => true,
+      createChatCompletion: async ({ onData }) => {
+        onData(reasoningTracedChunk('private reasoning', 303, {
+          usage: { prompt_tokens: 4, completion_tokens: 1, total_tokens: 5 },
+        }));
+        onData(chunk(''));
+      },
+    };
+    const streamed: Array<{ text: string; reasoningDelta?: string }> = [];
+
+    const result = await runtime.generate(
+      'request-reasoning-trace',
+      { messages: [{ role: 'user', content: 'Think.' }], returnTokenIds: true },
+      new AbortController().signal,
+      (event) => {
+        if (event.event === 'token') streamed.push(event);
+      },
+    );
+
+    expect(result.text).toBe('');
+    expect(result.reasoningText).toBe('private reasoning');
+    expect(result.tokenIds).toEqual([303]);
+    expect(result.tokenTrace).toHaveLength(1);
+    expect(streamed).toEqual([{ type: 'event', requestId: 'request-reasoning-trace', event: 'token', text: '', reasoningDelta: 'private reasoning' }]);
+  });
+
+  it('fails loudly when a sampled token id is non-integer', async () => {
+    const runtime = new BrowserEngineRuntime();
+    const internals = runtime as unknown as RuntimeInternals;
+    internals.loaded = {
+      manifest: {},
+      backend: 'wasm',
+      model: {
+        id: '1_7b',
+        displayName: 'Fixture Bonsai',
+        cpuFallback: true,
+        runtimePolicy: {
+          flashAttention: false,
+          tokenEmbeddingOnWebGPU: true,
+          requireSingleWebGPUGraph: false,
+        },
+      },
+    };
+    internals.wllama = {
+      isModelLoaded: () => true,
+      createChatCompletion: async ({ onData }) => {
+        onData(tracedChunk('bad', 1.5));
+      },
+    };
+
+    await expect(runtime.generate(
+      'request-invalid-token-id',
+      { messages: [{ role: 'user', content: 'Count.' }], returnTokenIds: true },
+      new AbortController().signal,
+      () => undefined,
+    )).rejects.toMatchObject({ code: 'INVALID_TOKEN_ID_TRACE' });
+  });
+
+  it('fails loudly when a sampled logprob entry omits its token id', async () => {
+    const runtime = new BrowserEngineRuntime();
+    const internals = runtime as unknown as RuntimeInternals;
+    internals.loaded = {
+      manifest: {},
+      backend: 'wasm',
+      model: {
+        id: '1_7b',
+        displayName: 'Fixture Bonsai',
+        cpuFallback: true,
+        runtimePolicy: {
+          flashAttention: false,
+          tokenEmbeddingOnWebGPU: true,
+          requireSingleWebGPUGraph: false,
+        },
+      },
+    };
+    internals.wllama = {
+      isModelLoaded: () => true,
+      createChatCompletion: async ({ onData }) => {
+        const malformed = tracedChunk('missing', 123);
+        const entry = malformed.choices[0]?.logprobs?.content?.[0];
+        if (!entry) throw new Error('Malformed trace fixture requires one logprob entry.');
+        Reflect.deleteProperty(entry, 'id');
+        onData(malformed);
+      },
+    };
+
+    await expect(runtime.generate(
+      'request-missing-token-id-field',
+      { messages: [{ role: 'user', content: 'Count.' }], returnTokenIds: true },
+      new AbortController().signal,
+      () => undefined,
+    )).rejects.toMatchObject({
+      code: 'INVALID_TOKEN_ID_TRACE',
+      details: { index: 0, id: undefined },
+    });
+  });
+
+  it('fails loudly when completion usage outnumbers sampled token ids', async () => {
+    const runtime = new BrowserEngineRuntime();
+    const internals = runtime as unknown as RuntimeInternals;
+    internals.loaded = {
+      manifest: {},
+      backend: 'wasm',
+      model: {
+        id: '1_7b',
+        displayName: 'Fixture Bonsai',
+        cpuFallback: true,
+        runtimePolicy: {
+          flashAttention: false,
+          tokenEmbeddingOnWebGPU: true,
+          requireSingleWebGPUGraph: false,
+        },
+      },
+    };
+    internals.wllama = {
+      isModelLoaded: () => true,
+      createChatCompletion: async ({ onData }) => {
+        onData(chunk('missing', {
+          usage: { prompt_tokens: 4, completion_tokens: 1, total_tokens: 5 },
+        }));
+      },
+    };
+
+    await expect(runtime.generate(
+      'request-missing-token-id',
+      { messages: [{ role: 'user', content: 'Count.' }], returnTokenIds: true },
+      new AbortController().signal,
+      () => undefined,
+    )).rejects.toMatchObject({
+      code: 'INCOMPLETE_TOKEN_ID_TRACE',
+      details: { expected: 1, observed: 0 },
+    });
+  });
+
+  it('fails loudly when a sampled token does not have exactly five top candidates', async () => {
+    const runtime = new BrowserEngineRuntime();
+    const internals = runtime as unknown as RuntimeInternals;
+    internals.loaded = {
+      manifest: {},
+      backend: 'wasm',
+      model: {
+        id: '1_7b',
+        displayName: 'Fixture Bonsai',
+        cpuFallback: true,
+        runtimePolicy: {
+          flashAttention: false,
+          tokenEmbeddingOnWebGPU: true,
+          requireSingleWebGPUGraph: false,
+        },
+      },
+    };
+    internals.wllama = {
+      isModelLoaded: () => true,
+      createChatCompletion: async ({ onData }) => {
+        const malformed = tracedChunk('short', 404);
+        malformed.choices[0]?.logprobs?.content?.[0]?.top_logprobs.pop();
+        onData(malformed);
+      },
+    };
+
+    await expect(runtime.generate(
+      'request-short-top-logprobs',
+      { messages: [{ role: 'user', content: 'Count.' }], returnTokenIds: true },
+      new AbortController().signal,
+      () => undefined,
+    )).rejects.toMatchObject({
+      code: 'INVALID_TOKEN_LOGPROB_TRACE',
+      details: { index: 0, expected: 5, observed: 4 },
+    });
+  });
+
+  it('fails loudly when a top candidate has an invalid logprob', async () => {
+    const runtime = new BrowserEngineRuntime();
+    const internals = runtime as unknown as RuntimeInternals;
+    internals.loaded = {
+      manifest: {},
+      backend: 'wasm',
+      model: {
+        id: '1_7b',
+        displayName: 'Fixture Bonsai',
+        cpuFallback: true,
+        runtimePolicy: {
+          flashAttention: false,
+          tokenEmbeddingOnWebGPU: true,
+          requireSingleWebGPUGraph: false,
+        },
+      },
+    };
+    internals.wllama = {
+      isModelLoaded: () => true,
+      createChatCompletion: async ({ onData }) => {
+        const malformed = tracedChunk('invalid', 505);
+        const candidate = malformed.choices[0]?.logprobs?.content?.[0]?.top_logprobs[1];
+        if (!candidate) throw new Error('Malformed trace fixture requires a top candidate.');
+        candidate.logprob = Number.NaN;
+        onData(malformed);
+      },
+    };
+
+    await expect(runtime.generate(
+      'request-invalid-top-logprob',
+      { messages: [{ role: 'user', content: 'Count.' }], returnTokenIds: true },
+      new AbortController().signal,
+      () => undefined,
+    )).rejects.toMatchObject({
+      code: 'INVALID_TOKEN_LOGPROB_TRACE',
+      details: { index: 0, field: 'topCandidates[1].logprob' },
+    });
   });
 });

src/engine/runtime-score-sequence.test.ts

@@ -0,0 +1,225 @@
+import { describe, expect, it, vi } from 'vitest';
+import type { BackendReport } from './native-log';
+import { BrowserEngineRuntime } from './runtime';
+
+interface RawCompletionOptions {
+  prompt: number[];
+  max_tokens: number;
+  temperature: number;
+  top_k: number;
+  logprobs: number;
+  logit_bias: Record<string, number>;
+  cache_prompt: boolean;
+  post_sampling_probs: boolean;
+  abortSignal: AbortSignal;
+}
+
+interface RuntimeInternals {
+  wllama: {
+    isModelLoaded(): boolean;
+    createCompletion(options: RawCompletionOptions): Promise<unknown>;
+  } | null;
+  loaded: {
+    manifest: unknown;
+    backend: 'webgpu';
+    tuningScope: 'benchmark';
+    contextSize: number;
+    batchSize: number;
+    microBatchSize: number;
+    vocabularySize: number;
+    model: {
+      id: '27b';
+      displayName: string;
+      cpuFallback: false;
+      runtimePolicy: {
+        flashAttention: false;
+        tokenEmbeddingOnWebGPU: true;
+        requireSingleWebGPUGraph: true;
+      };
+    };
+  } | null;
+  nativeLog: {
+    report(): BackendReport;
+  };
+}
+
+const backendReport: BackendReport = {
+  backends: ['WebGPU'],
+  nGraphSplits: 1,
+  opsOnCpu: 0,
+  layersGpu: { offloaded: 65, total: 65 },
+  flashAttention: false,
+  cacheTypeK: 'f16',
+  cacheTypeV: 'f16',
+  webgpuKvBufferBytes: 128 * 1024 ** 2,
+};
+
+function configuredRuntime(createCompletion: (options: RawCompletionOptions) => Promise<unknown>) {
+  const runtime = new BrowserEngineRuntime();
+  const internals = runtime as unknown as RuntimeInternals;
+  internals.loaded = {
+    manifest: {},
+    backend: 'webgpu',
+    tuningScope: 'benchmark',
+    contextSize: 2_048,
+    batchSize: 32,
+    microBatchSize: 16,
+    vocabularySize: 248_320,
+    model: {
+      id: '27b',
+      displayName: 'Fixture Bonsai 27B',
+      cpuFallback: false,
+      runtimePolicy: {
+        flashAttention: false,
+        tokenEmbeddingOnWebGPU: true,
+        requireSingleWebGPUGraph: true,
+      },
+    },
+  };
+  internals.wllama = { isModelLoaded: () => true, createCompletion };
+  vi.spyOn(internals.nativeLog, 'report').mockReturnValue(backendReport);
+  return runtime;
+}
+
+describe('BrowserEngineRuntime teacher-forced scoring', () => {
+  it('scores the exact CPU sequence with raw token prefixes and keeps natural top-1 separate', async () => {
+    const calls: RawCompletionOptions[] = [];
+    const createCompletion = vi.fn(async (options: RawCompletionOptions) => {
+      calls.push(options);
+      const index = options.prompt.length - 38;
+      const referenceId = Number(Object.keys(options.logit_bias)[0]);
+      const naturalTop1Id = index === 29 ? referenceId + 10 : referenceId;
+      const candidates = index === 29
+        ? [
+            { id: naturalTop1Id, token: 'natural', logprob: -0.01, bytes: null },
+            { id: referenceId, token: 'reference', logprob: -0.02, bytes: null },
+            { id: referenceId + 20, token: 'third', logprob: -1, bytes: null },
+            { id: referenceId + 21, token: 'fourth', logprob: -2, bytes: null },
+            { id: referenceId + 22, token: 'fifth', logprob: -3, bytes: null },
+          ]
+        : [
+            { id: referenceId, token: 'reference', logprob: -0.01, bytes: null },
+            { id: referenceId + 10, token: 'second', logprob: -0.02, bytes: null },
+            { id: referenceId + 20, token: 'third', logprob: -1, bytes: null },
+            { id: referenceId + 21, token: 'fourth', logprob: -2, bytes: null },
+            { id: referenceId + 22, token: 'fifth', logprob: -3, bytes: null },
+          ];
+      return {
+        choices: [{
+          text: 'forced',
+          finish_reason: 'length',
+          logprobs: {
+            content: [{
+              id: referenceId,
+              token: 'reference',
+              logprob: index === 29 ? -0.02 : -0.01,
+              bytes: null,
+              top_logprobs: candidates,
+            }],
+          },
+        }],
+      };
+    });
+    const runtime = configuredRuntime(createCompletion);
+    const promptTokenIds = Array.from({ length: 38 }, (_, index) => index + 1);
+    const referenceTokenIds = Array.from({ length: 1_024 }, (_, index) => index + 1_000);
+
+    const result = await runtime.scoreSequence({
+      promptTokenIds,
+      referenceTokenIds,
+      topK: 5,
+    }, new AbortController().signal);
+
+    expect(createCompletion).toHaveBeenCalledTimes(1_024);
+    expect(calls[0]).toMatchObject({
+      prompt: promptTokenIds,
+      max_tokens: 1,
+      temperature: 0,
+      top_k: 1,
+      logprobs: 5,
+      logit_bias: { '1000': 1_000 },
+      cache_prompt: false,
+      post_sampling_probs: false,
+    });
+    expect(calls[1]).toMatchObject({
+      prompt: [...promptTokenIds, 1_000],
+      cache_prompt: true,
+    });
+    expect(calls.at(-1)?.prompt).toEqual([
+      ...promptTokenIds,
+      ...referenceTokenIds.slice(0, -1),
+    ]);
+    expect(result.entries[29]).toMatchObject({
+      index: 29,
+      selectedReference: { id: 1_029, logprob: -0.02 },
+      naturalTop1: { id: 1_039, logprob: -0.01 },
+      referenceRankInTopCandidatesZeroBased: 1,
+      top1Top2Margin: 0.01,
+    });
+    expect(result.summary.tokenCount).toBe(1_024);
+    expect(result.summary.meanNll).toBeCloseTo((1_023 * 0.01 + 0.02) / 1_024, 12);
+    expect(result.summary.perplexity).toBeCloseTo(Math.exp(result.summary.meanNll), 12);
+  });
+
+  it('honors an already-aborted diagnostic request before the first raw completion', async () => {
+    const createCompletion = vi.fn(async () => ({}));
+    const runtime = configuredRuntime(createCompletion);
+    const controller = new AbortController();
+    controller.abort();
+
+    await expect(runtime.scoreSequence({
+      promptTokenIds: Array.from({ length: 38 }, (_, index) => index + 1),
+      referenceTokenIds: Array.from({ length: 1_024 }, (_, index) => index + 1_000),
+      topK: 5,
+    }, controller.signal)).rejects.toMatchObject({ name: 'AbortError' });
+    expect(createCompletion).not.toHaveBeenCalled();
+  });
+
+  it('fails loudly when logit bias does not return the fixed reference token', async () => {
+    const runtime = configuredRuntime(async (options) => {
+      const referenceId = Number(Object.keys(options.logit_bias)[0]);
+      const selectedId = referenceId + 1;
+      return {
+        choices: [{
+          logprobs: {
+            content: [{
+              id: selectedId,
+              logprob: -0.01,
+              top_logprobs: [
+                { id: selectedId, logprob: -0.01 },
+                { id: referenceId, logprob: -0.02 },
+                { id: referenceId + 2, logprob: -1 },
+                { id: referenceId + 3, logprob: -2 },
+                { id: referenceId + 4, logprob: -3 },
+              ],
+            }],
+          },
+        }],
+      };
+    });
+
+    await expect(runtime.scoreSequence({
+      promptTokenIds: Array.from({ length: 38 }, (_, index) => index + 1),
+      referenceTokenIds: Array.from({ length: 1_024 }, (_, index) => index + 1_000),
+      topK: 5,
+    }, new AbortController().signal)).rejects.toMatchObject({
+      code: 'INVALID_SCORE_SEQUENCE_RESPONSE',
+      details: { index: 0, referenceTokenId: 1_000 },
+    });
+  });
+
+  it('rejects scoring outside the loaded 27B WebGPU benchmark path', async () => {
+    const runtime = configuredRuntime(async () => ({}));
+    const internals = runtime as unknown as RuntimeInternals;
+    if (!internals.loaded) throw new Error('Expected loaded fixture state.');
+    (internals.loaded as { tuningScope: string }).tuningScope = 'release-defaults';
+
+    await expect(runtime.scoreSequence({
+      promptTokenIds: Array.from({ length: 38 }, (_, index) => index + 1),
+      referenceTokenIds: Array.from({ length: 1_024 }, (_, index) => index + 1_000),
+      topK: 5,
+    }, new AbortController().signal)).rejects.toMatchObject({
+      code: 'SCORE_SEQUENCE_UNAVAILABLE',
+    });
+  });
+});

src/engine/runtime.ts

@@ -10,6 +10,7 @@ import runtimeSource from '../../vendor/wllama-bonsai/SOURCE.json';
 import {
   assertBackendPolicy,
   estimateStorage,
+  evaluateModelContextPolicy,
   evaluateModelGate,
   inspectWebGpuAdapter,
   persistStorage,
@@ -36,11 +37,14 @@ import type {
   BenchmarkWasmFlavor,
   EngineCapabilities,
   EngineEvent,
+  EngineSampledTokenTraceEntry,
   GenerateParams,
   GenerateResult,
   LoadModelParams,
   LoadModelResult,
   RuntimeBackend,
+  ScoreSequenceParams,
+  ScoreSequenceResult,
   ShardDownloadFailureDetails,
   StorageEstimate,
 } from './protocol';
@@ -52,6 +56,10 @@ const WLLAMA_COMPAT_WASM_PATH = '/wasm/wllama-compat.wasm';
 const WLLAMA_COMPAT_WORKER_PATH = '/wasm/wllama-compat.js';
 const MAX_GLUE_INT = 2_147_483_647;
 const DEVICE_LOST_EXIT_GRACE_MS = 100;
+const TOKEN_TRACE_TOP_LOGPROBS = 5;
+const STATE_DRIFT_REFERENCE_TOKENS = 1_024;
+const STATE_DRIFT_PROMPT_TOKENS = 38;
+const TEACHER_FORCE_LOGIT_BIAS = 1_000;
 
 export interface ResolvedLoadTuning {
   scope: 'release-defaults' | 'benchmark';
@@ -183,6 +191,11 @@ interface LoadedModelState {
   manifest: ModelManifestV2;
   model: ManifestModelV2;
   backend: RuntimeBackend;
+  tuningScope: ResolvedLoadTuning['scope'];
+  contextSize: number;
+  batchSize: number;
+  microBatchSize: number;
+  vocabularySize: number;
 }
 
 interface CachedShardState {
@@ -312,6 +325,244 @@ function mapTimings(timings: ResultTimings | undefined): GenerateResult['timings
   };
 }
 
+function traceTokenId(value: unknown, index: number, field: string): number {
+  if (!Number.isSafeInteger(value) || (value as number) < 0 || (value as number) > 0xffff_ffff) {
+    throw new EngineRuntimeError(
+      'INVALID_TOKEN_ID_TRACE',
+      'The native runtime returned a missing or invalid token id in the sampled-token trace.',
+      { index, field, id: value },
+    );
+  }
+  return value as number;
+}
+
+function traceLogprob(value: unknown, index: number, field: string): number {
+  if (typeof value !== 'number' || !Number.isFinite(value)) {
+    throw new EngineRuntimeError(
+      'INVALID_TOKEN_LOGPROB_TRACE',
+      'The native runtime returned a missing or invalid logprob in the sampled-token trace.',
+      { index, field, logprob: value },
+    );
+  }
+  return value;
+}
+
+function appendSampledTokenTrace(
+  target: EngineSampledTokenTraceEntry[],
+  chunk: ChatCompletionChunk,
+): void {
+  const entries = chunk.choices[0]?.logprobs?.content;
+  if (!entries) return;
+  for (const entry of entries) {
+    const index = target.length;
+    const selected = {
+      id: traceTokenId(entry.id, index, 'selected.id'),
+      logprob: traceLogprob(entry.logprob, index, 'selected.logprob'),
+    };
+    if (!Array.isArray(entry.top_logprobs) || entry.top_logprobs.length !== TOKEN_TRACE_TOP_LOGPROBS) {
+      throw new EngineRuntimeError(
+        'INVALID_TOKEN_LOGPROB_TRACE',
+        `The native runtime must return exactly ${TOKEN_TRACE_TOP_LOGPROBS} top logprob candidates per sampled token.`,
+        {
+          index,
+          expected: TOKEN_TRACE_TOP_LOGPROBS,
+          observed: Array.isArray(entry.top_logprobs) ? entry.top_logprobs.length : null,
+        },
+      );
+    }
+    const seenIds = new Set<number>();
+    const topCandidates = entry.top_logprobs.map((candidate, candidateIndex) => {
+      const id = traceTokenId(candidate.id, index, `topCandidates[${candidateIndex}].id`);
+      if (seenIds.has(id)) {
+        throw new EngineRuntimeError(
+          'INVALID_TOKEN_LOGPROB_TRACE',
+          'The native runtime returned duplicate ids in a top-logprob candidate list.',
+          { index, candidateIndex, id },
+        );
+      }
+      seenIds.add(id);
+      return {
+        id,
+        logprob: traceLogprob(
+          candidate.logprob,
+          index,
+          `topCandidates[${candidateIndex}].logprob`,
+        ),
+      };
+    }).sort((left, right) => right.logprob - left.logprob || left.id - right.id);
+    const selectedCandidate = topCandidates.find((candidate) => candidate.id === selected.id);
+    if (!selectedCandidate || selectedCandidate.logprob !== selected.logprob) {
+      throw new EngineRuntimeError(
+        'INVALID_TOKEN_LOGPROB_TRACE',
+        'The sampled token must appear exactly once in its top-logprob candidates with the same logprob.',
+        { index, selected, selectedCandidate: selectedCandidate ?? null },
+      );
+    }
+    target.push({ selected, topCandidates });
+  }
+}
+
+function assertCompleteTokenTrace(
+  tokenTrace: EngineSampledTokenTraceEntry[] | null,
+  usage: GenerateResult['usage'],
+): void {
+  if (tokenTrace === null) return;
+  if (usage === null || tokenTrace.length !== usage.completionTokens) {
+    throw new EngineRuntimeError(
+      'INCOMPLETE_TOKEN_ID_TRACE',
+      'The native runtime did not return exactly one complete sampled-token trace per completion token.',
+      {
+        expected: usage?.completionTokens ?? null,
+        observed: tokenTrace.length,
+      },
+    );
+  }
+}
+
+function scoreRecord(value: unknown, index: number, field: string): Record<string, unknown> {
+  if (typeof value !== 'object' || value === null || Array.isArray(value)) {
+    throw new EngineRuntimeError(
+      'INVALID_SCORE_SEQUENCE_RESPONSE',
+      `Teacher-forced response ${index} has an invalid ${field}.`,
+      { index, field },
+    );
+  }
+  return value as Record<string, unknown>;
+}
+
+function scoreTokenId(
+  value: unknown,
+  index: number,
+  field: string,
+  vocabularySize: number,
+): number {
+  if (!Number.isSafeInteger(value) || (value as number) < 0 || (value as number) >= vocabularySize) {
+    throw new EngineRuntimeError(
+      'INVALID_SCORE_SEQUENCE_RESPONSE',
+      `Teacher-forced response ${index} has an invalid ${field}.`,
+      { index, field, id: value, vocabularySize },
+    );
+  }
+  return value as number;
+}
+
+function scoreLogprob(value: unknown, index: number, field: string): number {
+  if (typeof value !== 'number' || !Number.isFinite(value)) {
+    throw new EngineRuntimeError(
+      'INVALID_SCORE_SEQUENCE_RESPONSE',
+      `Teacher-forced response ${index} has an invalid ${field}.`,
+      { index, field, logprob: value },
+    );
+  }
+  return value;
+}
+
+function parseTeacherForcedResponse(
+  response: unknown,
+  index: number,
+  referenceTokenId: number,
+  vocabularySize: number,
+): ScoreSequenceResult['entries'][number] {
+  const root = scoreRecord(response, index, 'root');
+  if (!Array.isArray(root.choices) || root.choices.length !== 1) {
+    throw new EngineRuntimeError(
+      'INVALID_SCORE_SEQUENCE_RESPONSE',
+      `Teacher-forced response ${index} must contain exactly one completion choice.`,
+      { index, observed: Array.isArray(root.choices) ? root.choices.length : null },
+    );
+  }
+  const choice = scoreRecord(root.choices[0], index, 'choices[0]');
+  const logprobs = scoreRecord(choice.logprobs, index, 'choices[0].logprobs');
+  if (!Array.isArray(logprobs.content) || logprobs.content.length !== 1) {
+    throw new EngineRuntimeError(
+      'INVALID_SCORE_SEQUENCE_RESPONSE',
+      `Teacher-forced response ${index} must contain one selected-token logprob entry.`,
+      { index, observed: Array.isArray(logprobs.content) ? logprobs.content.length : null },
+    );
+  }
+  const selectedEntry = scoreRecord(logprobs.content[0], index, 'logprobs.content[0]');
+  const selectedReference = {
+    id: scoreTokenId(selectedEntry.id, index, 'selectedReference.id', vocabularySize),
+    logprob: scoreLogprob(selectedEntry.logprob, index, 'selectedReference.logprob'),
+  };
+  if (selectedReference.id !== referenceTokenId) {
+    throw new EngineRuntimeError(
+      'INVALID_SCORE_SEQUENCE_RESPONSE',
+      `Teacher forcing selected token ${selectedReference.id} instead of reference token ${referenceTokenId} at position ${index + 1}.`,
+      { index, selectedTokenId: selectedReference.id, referenceTokenId },
+    );
+  }
+  const rawTopLogprobs = selectedEntry.top_logprobs;
+  if (!Array.isArray(rawTopLogprobs)
+    || rawTopLogprobs.length !== TOKEN_TRACE_TOP_LOGPROBS) {
+    throw new EngineRuntimeError(
+      'INVALID_SCORE_SEQUENCE_RESPONSE',
+      `Teacher-forced response ${index} must contain exactly ${TOKEN_TRACE_TOP_LOGPROBS} natural top candidates.`,
+      {
+        index,
+        observed: Array.isArray(rawTopLogprobs)
+          ? rawTopLogprobs.length
+          : null,
+      },
+    );
+  }
+  const seenIds = new Set<number>();
+  const topCandidates = rawTopLogprobs.map((rawCandidate, candidateIndex) => {
+    const candidate = scoreRecord(
+      rawCandidate,
+      index,
+      `topCandidates[${candidateIndex}]`,
+    );
+    const parsed = {
+      id: scoreTokenId(
+        candidate.id,
+        index,
+        `topCandidates[${candidateIndex}].id`,
+        vocabularySize,
+      ),
+      logprob: scoreLogprob(
+        candidate.logprob,
+        index,
+        `topCandidates[${candidateIndex}].logprob`,
+      ),
+    };
+    if (seenIds.has(parsed.id)) {
+      throw new EngineRuntimeError(
+        'INVALID_SCORE_SEQUENCE_RESPONSE',
+        `Teacher-forced response ${index} contains duplicate natural candidate id ${parsed.id}.`,
+        { index, candidateIndex, id: parsed.id },
+      );
+    }
+    seenIds.add(parsed.id);
+    return parsed;
+  }).sort((left, right) => right.logprob - left.logprob || left.id - right.id);
+  const referenceRankInTopCandidatesZeroBased = topCandidates.findIndex(
+    (candidate) => candidate.id === referenceTokenId,
+  );
+  if (
+    referenceRankInTopCandidatesZeroBased !== -1
+    && topCandidates[referenceRankInTopCandidatesZeroBased]?.logprob !== selectedReference.logprob
+  ) {
+    throw new EngineRuntimeError(
+      'INVALID_SCORE_SEQUENCE_RESPONSE',
+      `Teacher-forced response ${index} reports inconsistent reference logprobs.`,
+      { index, selectedReference, referenceRankInTopCandidatesZeroBased },
+    );
+  }
+  const naturalTop1 = topCandidates[0]!;
+  const runnerUp = topCandidates[1]!;
+  return {
+    index,
+    selectedReference,
+    naturalTop1: { ...naturalTop1 },
+    topCandidates,
+    referenceRankInTopCandidatesZeroBased: referenceRankInTopCandidatesZeroBased === -1
+      ? null
+      : referenceRankInTopCandidatesZeroBased,
+    top1Top2Margin: naturalTop1.logprob - runnerUp.logprob,
+  };
+}
+
 function asReasoningDelta(chunk: ChatCompletionChunk): string | undefined {
   const choice = chunk.choices[0] as unknown as Record<string, unknown> | undefined;
   const delta = choice?.delta;
@@ -685,6 +936,19 @@ export class BrowserEngineRuntime {
       shardPath: null,
     });
     const model = findManifestModel(manifest, params.modelId);
+    const contextSize = params.contextSize ?? model.defaultContext;
+    const contextPolicy = evaluateModelContextPolicy(model, contextSize, {
+      tuningScope: tuning.scope,
+      requestedBackend: params.backend,
+      flashMode: tuning.flashMode,
+      kvCacheType: tuning.kvCacheType,
+    });
+    if (!contextPolicy.allowed) {
+      throw new EngineRuntimeError(
+        'INVALID_CONTEXT_SIZE',
+        `Context size must be between 1 and ${contextPolicy.limit} for this model and runtime policy.`,
+      );
+    }
     const urls = orderedShardUrls(manifest, model);
     const wllama = await this.ensureWllama(wasmFlavor);
     const cached = await this.inspectCachedShards(wllama, model, urls, signal);
@@ -713,13 +977,6 @@ export class BrowserEngineRuntime {
       shardCount: model.files.length,
       shardPath: null,
     });
-    const contextSize = params.contextSize ?? model.defaultContext;
-    if (!Number.isSafeInteger(contextSize) || contextSize <= 0 || contextSize > model.contextLength) {
-      throw new EngineRuntimeError(
-        'INVALID_CONTEXT_SIZE',
-        `Context size must be between 1 and ${model.contextLength}.`,
-      );
-    }
     const defaultThreads = Math.max(1, Math.floor((navigator.hardwareConcurrency || 1) / 2));
     const threads = params.threads ?? (gate.selectedBackend === 'wasm' ? defaultThreads : 1);
     if (!Number.isSafeInteger(threads) || threads <= 0) {
@@ -801,7 +1058,16 @@ export class BrowserEngineRuntime {
           tuningResult,
         );
       }
-      this.loaded = { manifest, model, backend: gate.selectedBackend };
+      this.loaded = {
+        manifest,
+        model,
+        backend: gate.selectedBackend,
+        tuningScope: tuning.scope,
+        contextSize: context.n_ctx,
+        batchSize: context.n_batch,
+        microBatchSize: context.n_ubatch,
+        vocabularySize: context.n_vocab,
+      };
       emitProgress(sink, requestId, {
         phase: 'load',
         loadedBytes: model.downloadBytes,
@@ -860,9 +1126,11 @@ export class BrowserEngineRuntime {
     }
     throwIfAborted(signal);
     let text = '';
+    let reasoningText = '';
     let finishReason: GenerateResult['finishReason'] = null;
     let usage: GenerateResult['usage'] = null;
     let timings: GenerateResult['timings'] = null;
+    const tokenTrace: EngineSampledTokenTraceEntry[] | null = params.returnTokenIds === true ? [] : null;
     const streamedToolCalls = new ToolCallAccumulator();
     try {
       await this.raceWebGpuDeviceLoss('generate', loaded.backend, loaded.model, wllama.createChatCompletion({
@@ -877,15 +1145,20 @@ export class BrowserEngineRuntime {
         tools: params.tools as ChatCompletionTool[] | undefined,
         tool_choice: params.toolChoice,
         cache_prompt: params.cachePrompt ?? true,
+        ...(params.returnTokenIds === true
+          ? { logprobs: true, top_logprobs: TOKEN_TRACE_TOP_LOGPROBS }
+          : {}),
         abortSignal: signal,
         timings_per_token: true,
         onData: (chunk: ChatCompletionChunk) => {
+          if (tokenTrace !== null) appendSampledTokenTrace(tokenTrace, chunk);
           streamedToolCalls.append(chunk.choices[0]?.delta.tool_calls);
           const delta = chunk.choices[0]?.delta.content;
           const textDelta = typeof delta === 'string' ? delta : '';
           const reasoningDelta = asReasoningDelta(chunk);
           if (textDelta || reasoningDelta) {
             text += textDelta;
+            reasoningText += reasoningDelta ?? '';
             sink({
               type: 'event',
               requestId,
@@ -906,6 +1179,8 @@ export class BrowserEngineRuntime {
     }
     await this.assertWebGpuAlive('generate', loaded.backend, loaded.model);
     throwIfAborted(signal);
+    assertCompleteTokenTrace(tokenTrace, usage);
+    const tokenIds = tokenTrace?.map((entry) => entry.selected.id) ?? null;
     const report = this.nativeLog.report();
     try {
       assertBackendPolicy(loaded.model, loaded.backend, report);
@@ -926,7 +1201,172 @@ export class BrowserEngineRuntime {
         error instanceof Error ? error.message : String(error),
       );
     }
-    return { text, finishReason, toolCalls, usage, timings };
+    return { text, reasoningText, tokenIds, tokenTrace, finishReason, toolCalls, usage, timings };
+  }
+
+  async scoreSequence(
+    params: ScoreSequenceParams,
+    signal: AbortSignal,
+  ): Promise<ScoreSequenceResult> {
+    const loaded = this.loaded;
+    const wllama = this.wllama;
+    if (!loaded || !wllama?.isModelLoaded()) {
+      throw new EngineRuntimeError('MODEL_NOT_LOADED', 'Load a model before scoring a sequence.');
+    }
+    if (
+      loaded.model.id !== '27b'
+      || loaded.backend !== 'webgpu'
+      || loaded.tuningScope !== 'benchmark'
+    ) {
+      throw new EngineRuntimeError(
+        'SCORE_SEQUENCE_UNAVAILABLE',
+        'Teacher-forced scoring is diagnostic-only and requires the loaded 27B WebGPU benchmark path.',
+        {
+          modelId: loaded.model.id,
+          backend: loaded.backend,
+          tuningScope: loaded.tuningScope,
+        },
+      );
+    }
+    if (params.topK !== TOKEN_TRACE_TOP_LOGPROBS) {
+      throw new EngineRuntimeError(
+        'INVALID_SCORE_SEQUENCE',
+        `Teacher-forced scoring requires topK=${TOKEN_TRACE_TOP_LOGPROBS}.`,
+      );
+    }
+    if (!Array.isArray(params.promptTokenIds)
+      || params.promptTokenIds.length !== STATE_DRIFT_PROMPT_TOKENS) {
+      throw new EngineRuntimeError(
+        'INVALID_SCORE_SEQUENCE',
+        `Teacher-forced scoring requires exactly ${STATE_DRIFT_PROMPT_TOKENS} rendered prompt token ids.`,
+      );
+    }
+    if (!Array.isArray(params.referenceTokenIds)
+      || params.referenceTokenIds.length !== STATE_DRIFT_REFERENCE_TOKENS) {
+      throw new EngineRuntimeError(
+        'INVALID_SCORE_SEQUENCE',
+        `Teacher-forced scoring requires exactly ${STATE_DRIFT_REFERENCE_TOKENS} reference token ids.`,
+      );
+    }
+    if (params.promptTokenIds.length + params.referenceTokenIds.length > loaded.contextSize) {
+      throw new EngineRuntimeError(
+        'INVALID_SCORE_SEQUENCE',
+        'The rendered prompt and fixed reference sequence exceed the loaded context.',
+        {
+          promptTokens: params.promptTokenIds.length,
+          referenceTokens: params.referenceTokenIds.length,
+          contextSize: loaded.contextSize,
+        },
+      );
+    }
+    const validateInputTokenIds = (values: readonly number[], field: string): void => {
+      for (const [index, tokenId] of values.entries()) {
+        if (!Number.isSafeInteger(tokenId) || tokenId < 0 || tokenId >= loaded.vocabularySize) {
+          throw new EngineRuntimeError(
+            'INVALID_SCORE_SEQUENCE',
+            `${field}[${index}] is outside the loaded vocabulary.`,
+            { field, index, tokenId, vocabularySize: loaded.vocabularySize },
+          );
+        }
+      }
+    };
+    validateInputTokenIds(params.promptTokenIds, 'promptTokenIds');
+    validateInputTokenIds(params.referenceTokenIds, 'referenceTokenIds');
+    await this.assertWebGpuAlive('score-sequence', loaded.backend, loaded.model);
+    throwIfAborted(signal);
+
+    const score = async (): Promise<ScoreSequenceResult> => {
+      const entries: ScoreSequenceResult['entries'] = [];
+      const createRawCompletion = wllama.createCompletion.bind(wllama) as unknown as (
+        options: Record<string, unknown>,
+      ) => Promise<unknown>;
+      for (let index = 0; index < params.referenceTokenIds.length; index += 1) {
+        throwIfAborted(signal);
+        const referenceTokenId = params.referenceTokenIds[index]!;
+        const prompt = [
+          ...params.promptTokenIds,
+          ...params.referenceTokenIds.slice(0, index),
+        ];
+        const response = await createRawCompletion({
+          // The pinned llama.cpp endpoint accepts a raw token-id prompt even though
+          // upstream wllama's OAI declaration still narrows this field to strings.
+          prompt,
+          stream: false,
+          max_tokens: 1,
+          temperature: 0,
+          top_k: 1,
+          logprobs: TOKEN_TRACE_TOP_LOGPROBS,
+          logit_bias: { [String(referenceTokenId)]: TEACHER_FORCE_LOGIT_BIAS },
+          cache_prompt: index !== 0,
+          post_sampling_probs: false,
+          abortSignal: signal,
+        });
+        throwIfAborted(signal);
+        entries.push(parseTeacherForcedResponse(
+          response,
+          index,
+          referenceTokenId,
+          loaded.vocabularySize,
+        ));
+      }
+      const meanNll = -entries.reduce(
+        (sum, entry) => sum + entry.selectedReference.logprob,
+        0,
+      ) / entries.length;
+      const perplexity = Math.exp(meanNll);
+      if (!Number.isFinite(meanNll) || !Number.isFinite(perplexity)) {
+        throw new EngineRuntimeError(
+          'INVALID_SCORE_SEQUENCE_RESPONSE',
+          'Teacher-forced scoring produced a non-finite mean NLL or perplexity.',
+          { meanNll, perplexity },
+        );
+      }
+      return {
+        method: {
+          promptMode: 'raw-token-id-prefix',
+          maxTokensPerStep: 1,
+          temperature: 0,
+          topK: 1,
+          reportedTopLogprobs: TOKEN_TRACE_TOP_LOGPROBS,
+          logitBias: TEACHER_FORCE_LOGIT_BIAS,
+          cachePromptFirst: false,
+          cachePromptSubsequent: true,
+        },
+        entries,
+        summary: {
+          tokenCount: STATE_DRIFT_REFERENCE_TOKENS,
+          meanNll,
+          perplexity,
+        },
+      };
+    };
+
+    try {
+      const result = await this.raceWebGpuDeviceLoss(
+        'score-sequence',
+        loaded.backend,
+        loaded.model,
+        score(),
+      );
+      await this.assertWebGpuAlive('score-sequence', loaded.backend, loaded.model);
+      throwIfAborted(signal);
+      const report = this.nativeLog.report();
+      try {
+        assertBackendPolicy(loaded.model, loaded.backend, report);
+      } catch (error) {
+        await this.unload();
+        throw new EngineRuntimeError(
+          'BACKEND_TRIPWIRE',
+          error instanceof Error ? error.message : String(error),
+          report,
+        );
+      }
+      return result;
+    } catch (error) {
+      if (error instanceof EngineRuntimeError && error.code === 'WEBGPU_DEVICE_LOST') throw error;
+      await this.assertWebGpuAlive('score-sequence', loaded.backend, loaded.model);
+      throw error;
+    }
   }
 
   async backendReport(): Promise<BackendReport> {

src/engine/worker.ts

@@ -137,6 +137,14 @@ async function handleRequest(request: EngineRequest): Promise<void> {
           )),
         );
         return;
+      case 'scoreSequence':
+        postSuccess(
+          request,
+          await runExclusive(request.requestId, (signal) => (
+            runtime.scoreSequence(request.params, signal)
+          )),
+        );
+        return;
       case 'abort': {
         const operation = active.get(request.params.targetRequestId);
         const aborted = operation !== undefined && !operation.controller.signal.aborted;

vendor/wllama-bonsai/PATCH.diff

@@ -23,7 +23,7 @@ index f1496a4..8883324 100644
    GLUE_FIELD(int, n_threads)
    GLUE_FIELD_NULLABLE(str, model_alias)
 diff --git a/cpp/wllama-context.h b/cpp/wllama-context.h
-index 0dafe6d..c4b46f9 100644
+index 0dafe6d..d261dfe 100644
 --- a/cpp/wllama-context.h
 +++ b/cpp/wllama-context.h
 @@ -19,6 +19,7 @@
@@ -34,7 +34,7 @@ index 0dafe6d..c4b46f9 100644
  
  #include "ggml-cpu.h"
  #include "ggml-backend.h"
-@@ -300,7 +301,7 @@ struct wllama_context
+@@ -300,16 +301,28 @@ struct wllama_context
        server_task task = server_task(SERVER_TASK_TYPE_COMPLETION);
        task.id = rd->get_new_id();
        task.index = 0;
@@ -43,7 +43,31 @@ index 0dafe6d..c4b46f9 100644
            vocab,
            params,
            meta->slot_n_ctx,
-@@ -415,6 +416,8 @@ struct wllama_context
+           meta->logit_bias_eog,
+           body);
+       task.params.res_type = res_type;
+-      task.cli_prompt = prompt;
+-      task.cli_files = files;
+-      task.cli = true;
++      if (prompt.is_array())
++      {
++        auto tokenized_prompts = tokenize_input_prompts(vocab, nullptr, prompt, true, true);
++        if (tokenized_prompts.size() != 1)
++        {
++          throw app_exception("wllama completion accepts exactly one token-id prompt");
++        }
++        task.tokens = std::move(tokenized_prompts[0]);
++      }
++      else
++      {
++        task.cli_prompt = prompt.get<std::string>();
++        task.cli_files = files;
++        task.cli = true;
++      }
+ 
+       rd->post_task({std::move(task)});
+     }
+@@ -415,6 +428,8 @@ struct wllama_context
        params.embedding = req.embeddings.value;
      if (req.n_batch.not_null())
        params.n_batch = req.n_batch.value;
@@ -52,7 +76,7 @@ index 0dafe6d..c4b46f9 100644
      if (req.n_parallel.not_null())
        params.n_parallel = req.n_parallel.value;
      if (req.pooling_type.not_null())
-@@ -568,6 +571,24 @@ struct wllama_context
+@@ -568,6 +583,24 @@ struct wllama_context
      // load model
      llama_backend_init();
      llama_numa_init(params.numa);
@@ -101,6 +125,18 @@ index 08f0d87..c1001da 100644
    n_ctx: number;
    n_threads: number;
    model_alias?: string | undefined;
+diff --git a/src/types/oai-compat.ts b/src/types/oai-compat.ts
+index dbeeaa2..fd41519 100644
+--- a/src/types/oai-compat.ts
++++ b/src/types/oai-compat.ts
+@@ -135,6 +135,7 @@ export type ChatCompletionParams = {
+ // Response types----------
+ 
+ export interface ChatCompletionLogprob {
++  id: number;
+   token: string;
+   logprob: number;
+   bytes: number[] | null;
 diff --git a/src/types/types.ts b/src/types/types.ts
 index 2b3cca7..e91b395 100644
 --- a/src/types/types.ts
@@ -235,3 +271,17 @@ index ce74b2b..1796ca1 100644
        if (!this.compat) {
          this.logger().warn(
            'Not using compat mode' +
+diff --git a/src/workers-code/llama-cpp.js b/src/workers-code/llama-cpp.js
+index be62030..9f5144a 100644
+--- a/src/workers-code/llama-cpp.js
++++ b/src/workers-code/llama-cpp.js
+@@ -468,6 +468,9 @@ onmessage = async (e) => {
+       );
+       inputBuffer.set(argEncodedMsg, 0);
+       const outputPtr = await wllamaAction(argAction, inputPtr);
++      if (!outputPtr) {
++        throw new Error(`wllama_action("${argAction}") returned a null pointer`);
++      }
+       // length of output buffer is written at the first 4 bytes of input buffer
+       const outputLen = new Uint32Array(
+         getHeapU8().buffer,

vendor/wllama-bonsai/SOURCE.json

@@ -4,19 +4,21 @@
   "wllamaRevision": "912c18b75d4358c1405a64646b8dbe43a205943b",
   "llamaCppRevision": "00fa7cb284cbf133fc426733bd64238a3588a33e",
   "license": "MIT",
-  "patch": "Expose offload_token_embedding and n_ubatch; place token_embd.weight on WebGPU whenever GPU layers are requested; fail loudly when the requested WebGPU buffer is unavailable; add an opt-in forceCompat selector for controlled benchmark A/B runs.",
+  "patch": "Expose offload_token_embedding and n_ubatch; place token_embd.weight on WebGPU whenever GPU layers are requested; fail loudly when the requested WebGPU buffer is unavailable; add an opt-in forceCompat selector for controlled benchmark A/B runs; type sampled token ids in OAI logprob payloads; accept one raw token-id completion prompt and fail loudly on null action responses.",
   "patchSet": {
     "format": "git-diff-binary",
     "path": "vendor/wllama-bonsai/PATCH.diff",
-    "bytes": 9350,
-    "sha256": "ad8786a295eaeb75ee03752d7f84d30f14ea373407844d5c589f7ab46800bb4b",
+    "bytes": 11052,
+    "sha256": "27b96fa48c6fd66b44fe70e83cc775902e29ec5877261c7eb04fecabc529d2f9",
     "files": [
       "CMakeLists.txt",
       "cpp/glue.hpp",
       "cpp/wllama-context.h",
       "src/glue/messages.ts",
+      "src/types/oai-compat.ts",
       "src/types/types.ts",
-      "src/wllama.ts"
+      "src/wllama.ts",
+      "src/workers-code/llama-cpp.js"
     ]
   },
   "build": {
@@ -28,18 +30,18 @@
   "files": [
     {
       "path": "vendor/wllama-bonsai/esm/index.js",
-      "bytes": 365397,
-      "sha256": "24bcef8aea8e27fb7b7e2d9e6ea94ba8ced7bfbffd09a0675821b9eb1b4a4c9f"
+      "bytes": 365536,
+      "sha256": "10c3811776e34092a225632929b97046af4215a4c91156169105c886be5cba5d"
     },
     {
       "path": "public/wasm/wllama.wasm",
-      "bytes": 8017169,
-      "sha256": "dd71f58c75f32c677a64eeb2e56efab7dedbe7395f730bb08bc80021bd182caa"
+      "bytes": 8017784,
+      "sha256": "b292bf670a25e1ecf02c71dd92b49196f3336842135f1f85efd5a3fd5f36ea8a"
     },
     {
       "path": "public/wasm/wllama-compat.wasm",
-      "bytes": 14864188,
-      "sha256": "eca9754d0d8a490b5c7b58cb384d67e31aa8ab26dd330e4ad1460982df6dd82e"
+      "bytes": 14865338,
+      "sha256": "ea0623d2ca1758287a569169b4ad5a69c22e41bce3e5d022c7d3e4fb5c37918c"
     },
     {
       "path": "public/wasm/wllama-compat.js",

vendor/wllama-bonsai/esm/types/oai-compat.d.ts

@@ -96,6 +96,7 @@ export type ChatCompletionParams = {
     timings_per_token?: boolean;
 } & SamplingParams;
 export interface ChatCompletionLogprob {
+    id: number;
     token: string;
     logprob: number;
     bytes: number[] | null;