Generated from shared/CLAIMS.md, the single source of truth for every number in launch_deck.pptx, launch_overview.mp4, technical_story.mp4 and internal_cut.mp4. If a number is not in this register, it does not go on screen. The deck carries the claim chips and the speaker notes; this register is the full manifest, shipped with the production files because the deck appendix is limited to three slides.
Labels: MEASURED (scope) we ran it · MODEL arithmetic on stated assumptions · VENDOR CLAIM someone else's published figure, attributed · ILLUSTRATION drawn to explain, not a result · METHOD a structure, not a number · PLAN a proposed next step. S, B and V8 are source-deck labels, never claim IDs. Glyphs the brand subsets lack are transliterated (>=, <=, ~, x, ->, sum of).
| C1 | cost per accepted job = ( inference + tools and sandbox + retries and failed attempts + verification + operating overhead ) ÷ accepted jobs, with the latency requirement stated. | METHOD | PROMPT.md; COST-MODEL.md §1 |
| R19 | The quality floor is a prediction, not a guarantee. | FACT | S3; V8 51 |
| A4 | Failed vs passed SQLite attempts: median 48 vs 47 core-seconds (p10–p90 44–103 vs 42–112). Failures cost real CPU. | MEASURED (scope) | pilot #outcomes |
| C11 | Cost receipt, frontier-priced: one accepted job = $0.0359 (inference $0.0326 + tools $0.0024 + retries $0.0008). | MODEL | model §4.4 |
| C13 | Retry lever: a failure rate of 20% / 33% / 50% / 67% multiplies the cost per accepted job by 1.25x / 1.49x / 2.00x / 3.03x. | MODEL | model §4.5; A14 |
| A10 | Per-agent CPU demand: SQLite 0.276 cores; CompCert 0.543 cores (stated in the source decks as "about 0.28 of a core per coding agent on EPYC"). | MEASURED (scope) | pilot #formulas; seller-deck 4; update-v8 27; battlecard 1 |
| A14 | Verified success: 4 of 5 low-load trials passed the verifier; at >=12 agents the pass rate ranged 33–69%; the heaviest loads passed about a third. | MEASURED (scope) | pilot #outcomes |
| I2 | ~250 concurrent coding agents per MI350P (Metrum tests: terminal-bench coding workloads + raw MI350P inference). | MEASURED (scope) | EVIDENCE-SEED I2 |
| A5 | Zero 429s does not mean no model-response waiting. 0 of ~4,000 calls throttled; max relay admission wait 0.0002 s; agents still spent a large share of each run waiting on the model. | MEASURED (scope) | pilot #model, #results (V4) |
| A8 | Valid whole waves above 80% busy: 5 and 8 agents on 1 core (81.9%, 83.9%); 12 agents on 2 cores (82.8%). No valid 4-core whole wave exceeded 80% (highest load tested: 24 agents). | MEASURED (scope) | pilot #cpuonly |
| A11 | Latency-safe load (secondary): estimate 2 / 5 / 9 agents on 1 / 2 / 4 cores; largest valid single waves that met the target held 3 / 7 / 14 agents at 67% / 72% / 57% busy. | MEASURED (scope) | pilot #sc-latency |
| A12 | Registered verdicts: V1 FAIL (no load met >=90% busy with >=80% on time); V2, V3, V5 INCONCLUSIVE; V4 PASS (no throttling); V6 FAIL; V6′ FAIL. | MEASURED (scope) | pilot #results |
| A7 | T1 = 184.7 s (median of the first three single-agent trials); SLO = 2 x T1 = 369.3 s (6.2 min). | MEASURED (scope) | pilot glossary |
| A1 | Scope: SQLite coding agents on pinned AMD EPYC 9R14 cores on AWS, Qwen3 Coder Next on Amazon Bedrock, 25–26 Oct 2026. It is not NPU, Halo or MI350P validation and does not establish Router uplift. | MEASURED (scope) | pilot #cpuonly, #model |
| A2 | Primary sizing question: CPU-only — the number of agents that keeps >80% of the allocated CPU busy, regardless of completion or verifier result. Latency-safe load and verified success are separate measures. | MEASURED (scope) | pilot #cpuonly |
| A3 | The 3 / 7 / 14 on-time loads were single waves, not validated sustained capacity. All six sustained studies were invalid (instrumentation faults); the held-out four-core prediction failed. | MEASURED (scope) | pilot #cpuonly, #holdout |
| A9 | Provisional per-agent estimate: 3 / 6 / 12 agents at >80% busy on 1 / 2 / 4 cores, ranges 2–5 / 4–9 / 7–17. | MODEL | pilot #cpuonly |
| A15 | Sealed 4-core prediction (SHA-256 b95bbabb...eeeb81, recorded before any 4-core run) expected 62–67% on time at 4–14 agents; measured 83–100%. Scored as registered: V6 FAIL. | MEASURED (scope) | pilot #holdout |
| R2 | vLLM Semantic Router d3 live run (11 Oct): 204/204 test prompts routed to the intended tier, 104 ms p50, fail-closed verified; shipped in v4.2. | MEASURED (scope) | EVIDENCE-SEED R2 |
| R3 | OpenJev 93% classification on a 60-prompt test set, 193 ms decision latency. | MEASURED (scope) | EVIDENCE-SEED R3 |
| R4 | 85% lower model spend vs the frontier API bill in one deployment (~25 developers). The demo's 92.5% is vs GPT-6 Astra list for one team. | MEASURED (scope) | EVIDENCE-SEED R4 |
| R5 | Per-agent key revocation is effective on the next request (~2 s across replicas). Key revocation stops model access; it is not whole-agent containment. | MEASURED (scope) | EVIDENCE-SEED R5 |
| R14 | The router pod is CPU-only and requests no GPU. | FACT | S5; V8 38 |
| R15 | 100% of routed tokens land on a usage row; three ceilings per key (daily, monthly, lifetime); budgets reserve input plus the output cap before dispatch, then reconcile. | MEASURED (scope) | V8 18 |
| R23 | Neither decision model (d3, OpenJev) has been measured on Instinct yet; both ran on one NVIDIA RTX PRO 6000. | MEASURED (scope) | S13, S14; V8 47 |
| R2b | d3 boundary: the live run was d3 26B in enforce mode, fail-closed verified when d3 was unreachable. It does not establish typical savings (1.4% on the seed set, 25% on one 8-turn take) or long-context latency. | MEASURED (scope) | V8 47; S13 |
| R3b | OpenJev boundary: live run 25 Sep, 60 prompts x 2 on one RTX PRO 6000, 0 of 60 repeat flips; spend projected from a 15-call price sample. It does not establish production accuracy, Instinct performance, or commercial rights. | MEASURED (scope) | V8 47; S13 |
| R-G | Guardrail (not a claim): the Router is a gateway, not the serving runtime. No guaranteed quality, no aggregate atomic dollar budgets, no automatic privacy from a classifier, no whole-agent containment from key revocation. Verify release/runtime/model rights and distinguish vendor claims, measured results, models and illustrations. | FACT | PROMPT.md "EVIDENCE THAT MUST SURVIVE"; EVIDENCE-SEED |
| N1 | Qwen3.6-35B-A3B on the NPU scored 58/60 on a fixed 60-case suite (cloud gpt-4o-mini: 55/60). Decode 13.5 tok/s on the NPU vs 19.8 on the iGPU (llama.cpp Vulkan). | MEASURED (scope) | NPU PDF "Same models on two chips" |
| N2a | A d3:nano decision took 0.40 s alone on the NPU. | MEASURED (scope) | NPU PDF "Decisions while a chat answer streams" |
| N2b | The same decision took 39 s when a chat model ran on the same NPU (the NPU serves one request at a time). | MEASURED (scope) | NPU PDF |
| N2c | With chat on the iGPU and decisions on the NPU, the decision took 0.45 s, all 50 decisions were correct, and chat was 9% slower. | MEASURED (scope) | NPU PDF |
| N4 | NPU 93% busy while running a 35B chat model, CPU 4%, iGPU 0% (generation run). | MEASURED (scope) | NPU PDF "Which chip does the work" |
| N5 | Whisper large-v3 turbo: 2.4 s/clip on the NPU vs 4.2 s on the CPU. | MEASURED (scope) | NPU PDF |
| N6 | FastFlowLM NPU kernels: free for commercial use and redistribution in unmodified form under its binary license (the orchestration code is MIT); upstream issues were filed (FastFlowLM #880–882, #680; RyzenAI-SW #415). Verify the current terms before shipping. | FACT | FastFlowLM README (FastFlowLM NPU Binary License) — supersedes the earlier "non-commercial" reading; NPU PDF "Issues we filed upstream" |
| H1 | 53 tok/s for one user on Qwen3.6-35B-A3B (llama.cpp Q4_K_M); 16 chat streams at 15.8 tok/s each on Gemma 4 E4B (253 tok/s total, 0.86 s first token, official rocm/vllm). | MEASURED (scope) | EVIDENCE-SEED H1 |
| H2 | Quality equal to NVIDIA GB10 within noise on the shared hard-tier suites. | MEASURED (scope) | EVIDENCE-SEED H2 |
| H3 | The Halo box is a Ryzen AI Halo device, not a Dell SKU — Dell has no AMD Halo product today; the 128 GB Dell Pro Max is GB10. AMD list price $3,999. | FACT | EVIDENCE-SEED H; blueprint §07 |
| H5 | GB10 is about 1.3x Halo for one user and level at 16 users (our data). | MEASURED (scope) | blueprint §07 ("our data") |
| P8 | 16 interactive users per Halo box; 13.5 tok/s background generation on the NPU. | MEASURED (scope) | blueprint §05b |
| I1 | ~7,800 output tok/s per MI350P on a 35B FP8 model (128->128, concurrency >=1024); 282 W GPU draw. | MEASURED (scope) | EVIDENCE-SEED I1 |
| I2b | 3,150 output tok/s across 250 agents (~12.6 tok/s per agent); 282 W per GPU. | MEASURED (scope) | blueprint §05b capacity panel |
| I2c | ~2,000 concurrent coding agents per 8-card XE7745 node = 8 x the tested ~250 per MI350P. | MODEL | S11; V8 25, 26, 35; B2 |
| I2d | Sizing ladder: ~500 agents (R7725 + 2x MI350P, 288 GB HBM3E) · ~2,000 (R7725 + 1x XE7745, 8x MI350P) · ~4,000 (R7725 + 2x XE7745, 16x MI350P). | MODEL | S11; V8 26 |
| I3 | MI325X on DigitalOcean was a rented stand-in for the fine-tuning and demo runs. It is not MI350P performance and not on-premises proof. | FACT | EVIDENCE-SEED I3 |
| I4b | XE7745: 8x MI350P, 1,152 GB HBM3E total, 4U, air cooled, 4.8 kW GPU, 8 PCIe GPUs, 2x EPYC 9005. MI350P: 144 GB HBM3E, CDNA 4, 128 CU, 4 TB/s, PCIe Gen5 x16, 600 W. | VENDOR CLAIM | S8; V8 46, 53 |
| I4c | Dell's AI Platform with AMD names XE7745 and R7725 with MI350P and EPYC. The Dell catalog listing is in discussion, not live. | VENDOR CLAIM | S8; V8 22, 46; B2 |
| F1 | Run 1: tool-call strict exact match 4.5% -> 32.0%, but MMLU-Pro collapsed 60.5 -> 13.0 (IFEval 81.5 -> 63.5). The preregistered gate refused promotion. | MEASURED (scope) | EVIDENCE-SEED F1; inputs/fine-tuning-report.md |
| F2 | The same frozen router export (1,186 / 221 / 300 traces) trained on MI325X and RTX PRO 6000 came out statistically level (semantic 78 vs 74/100); both run-2 candidates stayed in canary. | MEASURED (scope) | EVIDENCE-SEED F2; report |
| F4b | One audited export contained 1,707 traces (1,186 train / 221 validation / 300 held-out); the uncaptured team was verified absent. | MEASURED (scope) | S13; V8 13 |
| F5 | Serving cost per harness success: $0.00142 (AMD MI325X) vs $0.00130 (NVIDIA RTX PRO 6000) on one replay harness, 300 held-out traces, serving only, excludes training. Training wall time 893 s (NVIDIA) vs 1,215 s (AMD) — the decks round the AMD run to "about 20 minutes". | MEASURED (scope) | report; seller-deck 13; update-v8 14 |
| F7 | Run-2 gates: NVIDIA exact 5/200 vs base 4/200 (+0.5 pp); AMD 7/200 vs base 9/200 (−1.0 pp); +10 pp required. | MEASURED (scope) | report |
| F9 | Teacher licensing: the provider's terms (§2.2(c)) prohibit using content to train AI models; real-teacher collection was stopped until a licensed teacher was chosen. | FACT | report |
| F10 | 0 of 2 candidates were promoted: the loop is proven, a better model is not yet. | MEASURED (scope) | S15; V8 13; B2 |
| F11 | LoRA training on governed, router-exported traces took about 20 minutes per run on Instinct MI325X (1,215 s; the NVIDIA run was 893 s). | MEASURED (scope) | S14; V8 14 |
| C2 | Illustrative workload (blueprint reference): 500 seats · 4 always-on agents per seat (2,000 agents) · 17 h/day · 30 days · 3% escalation to frontier · 12.6 tok/s sustained per agent · 46.3B agent output tokens/month · 12.1M jobs/month · ~3,704 output + ~74,088 input tokens per accepted job · node occupancy 68.7%. | ILLUSTRATION | COST-MODEL.md §3 |
| C16 | Price and cost inputs (from BLUEPRINT-REFERENCE.py): node capex $500,000 budgetary — not a quote; Halo box $3,999 x 32 boxes (1 per 16 users) -> total capex $627,968; DO MI325X $3.80/GPU-hr with no provisioning factor; GPT-6 Terra list $1/$5 per M tokens; agent input:output 20:1; 90% of agent input cached at 10% of list; electricity $0.12/kWh, PUE 1.4; node power 8 x 282 W measured + 1.5 kW host; ops $2,000/server/month + $20/Halo box; amortisation 36 months; 250 agents per MI350P, 12.6 output tok/s per agent. | MODEL | model §2 |
| C12 | Cost receipt, owned-priced at 68.7% busy: $0.0036 marginal / $0.0057 all-in per accepted job — ~9.9x below the frontier-priced job. Tool, retry and verification terms are identical across both columns. | MODEL | model §4.4 |
| C3 | All-frontier: $419,451/month for the modelled workload at GPT-6 Terra list ($1/$5 per M, 20:1 agent input, 90% cached at 10%). | MODEL | model §4.1 |
| C4 | Hybrid on-premises: $33,111/month all-in = node $16,350 + 32 Halo boxes $4,475 + cloud burst $0 + 3% escalation $12,287. 92.1% below all-frontier. | MODEL | model §4.1 |
| C5 | All-cloud (elastic): $34,644/month (elastic MI325X GPUs at $3.80/GPU-hr + hosted open models at $0.95/M + escalation). 4.4% above hybrid. | MODEL | model §4.1 |
| C6 | Break-even occupancy vs elastic cloud GPUs: 74.7% (~75%) — the node's all-in monthly cost equals the cloud cost of the same agent-hours. | MODEL | model §4.2 |
| C7 | Payback vs all-frontier: 1.6 months. | MODEL | model §4.2 |
| C8 | Payback vs all-cloud: 33.1 months. | MODEL | model §4.2 |
| C9 | On-prem cost per M output tokens at the modelled 68.7% occupancy: $0.055/M marginal (power + ops only) and $0.364/M all-in (including capex ÷ 36). | MODEL | model §4.3 |
| C10 | On-prem $/M output tokens: at 40% busy $0.094 marginal / $0.626 all-in; at 90% busy $0.042 / $0.278. | MODEL | model §4.3 |
| C14 | Escalation lever: 1% / 3% / 10% / 25% of output tokens to a frontier API -> hybrid monthly $24,920 / $33,111 / $61,780 / $123,212. The single biggest lever on the hybrid bill. | MODEL | model §4.5 |
| C15 | Cached-input lever: 0% / 50% / 90% / 99% cached input -> all-frontier monthly $1,168,980 / $752,575 / $419,451 / $344,498. Cached input dominates the frontier price. | MODEL | model §4.5; G7 |
| G7 | Cache reads are 96–99% of agentic coding token spend, so any savings claim must use cached-input prices. | EXTERNAL (cited) | blueprint §07 |
| A18 | Reproduction: final snapshot restored to a fresh directory with every file hash matched; offline rebuild with network denied reproduced all 136 CSV files byte for byte; the sealed prediction hash was re-scored. Teardown verified 0 running instances, volumes, security groups, key pairs or snapshots tagged for the pilot. | MEASURED (scope) | pilot #data, #cost teardown |
| A20 | Follow-up work identified by the pilot: more matched repeats per load, a constant-backlog test to separate cohort drain from capacity, a second core type, and a measured 3-core point. | PLAN | pilot #followup |
| R17 | The Apache-2.0 router has been in production internally since June 2026. | MEASURED (scope) | V8 40 |
| R18 | Mandatory disclaimer: testing performed by Metrum AI; performance and cost claims are Metrum AI's and have not been independently verified by AMD. | FACT | B2; V8 46 |
| P9 | Proposed next step for the enterprise tier: a Dell/AMD lab path to MI350P agent-workload validation, with no catalog or commercial commitment. | PLAN | PROMPT.md slide 12 |
| P4 | cost_per_accepted_job = sum of(attributable inference + tool/sandbox compute + retries and failed attempts + verification + operating overhead) ÷ verified completed jobs, with the latency requirement stated. | METHOD | PROMPT.md; blueprint §05 |
| A6 | $150.76 total ($59.19 compute + $91.57 model tokens) covers compute and model tokens across recovery, invalid and supplemental work, and excludes orchestrating-agent tokens. It is a study cost. Never divide it by SQLite successes. | MEASURED (scope) | pilot #cost |
| N1 | Qwen3.6-35B-A3B on the NPU scored 58/60 on a fixed 60-case suite (cloud gpt-4o-mini: 55/60). Decode 13.5 tok/s on the NPU vs 19.8 on the iGPU (llama.cpp Vulkan). | MEASURED (scope) | NPU PDF "Same models on two chips" |
| N2 | Umbrella for the three decision latencies — always shown as three separate figures (N2a alone on the NPU, N2b contended, N2c split), never merged into one "NPU latency". | MEASURED (scope) | NPU PDF |
| N2a | A d3:nano decision took 0.40 s alone on the NPU. | MEASURED (scope) | NPU PDF "Decisions while a chat answer streams" |
| N2b | The same decision took 39 s when a chat model ran on the same NPU (the NPU serves one request at a time). | MEASURED (scope) | NPU PDF |
| N2c | With chat on the iGPU and decisions on the NPU, the decision took 0.45 s, all 50 decisions were correct, and chat was 9% slower. | MEASURED (scope) | NPU PDF |
| N3 | d3:mini scored 24/24 on the labelled set. | MEASURED (scope) | NPU PDF |
| N4 | NPU 93% busy while running a 35B chat model, CPU 4%, iGPU 0% (generation run). | MEASURED (scope) | NPU PDF "Which chip does the work" |
| N5 | Whisper large-v3 turbo: 2.4 s/clip on the NPU vs 4.2 s on the CPU. | MEASURED (scope) | NPU PDF |
| N6 | FastFlowLM NPU kernels: free for commercial use and redistribution in unmodified form under its binary license (the orchestration code is MIT); upstream issues were filed (FastFlowLM #880–882, #680; RyzenAI-SW #415). Verify the current terms before shipping. | FACT | FastFlowLM README (FastFlowLM NPU Binary License) — supersedes the earlier "non-commercial" reading; NPU PDF "Issues we filed upstream" |
| H1 | 53 tok/s for one user on Qwen3.6-35B-A3B (llama.cpp Q4_K_M); 16 chat streams at 15.8 tok/s each on Gemma 4 E4B (253 tok/s total, 0.86 s first token, official rocm/vllm). | MEASURED (scope) | EVIDENCE-SEED H1 |
| H2 | Quality equal to NVIDIA GB10 within noise on the shared hard-tier suites. | MEASURED (scope) | EVIDENCE-SEED H2 |
| H3 | The Halo box is a Ryzen AI Halo device, not a Dell SKU — Dell has no AMD Halo product today; the 128 GB Dell Pro Max is GB10. AMD list price $3,999. | FACT | EVIDENCE-SEED H; blueprint §07 |
| H5 | GB10 is about 1.3x Halo for one user and level at 16 users (our data). | MEASURED (scope) | blueprint §07 ("our data") |
| I1 | ~7,800 output tok/s per MI350P on a 35B FP8 model (128->128, concurrency >=1024); 282 W GPU draw. | MEASURED (scope) | EVIDENCE-SEED I1 |
| I2 | ~250 concurrent coding agents per MI350P (Metrum tests: terminal-bench coding workloads + raw MI350P inference). | MEASURED (scope) | EVIDENCE-SEED I2 |
| I2b | 3,150 output tok/s across 250 agents (~12.6 tok/s per agent); 282 W per GPU. | MEASURED (scope) | blueprint §05b capacity panel |
| I3 | MI325X on DigitalOcean was a rented stand-in for the fine-tuning and demo runs. It is not MI350P performance and not on-premises proof. | FACT | EVIDENCE-SEED I3 |
| I4 | MI350P PCIe: 144 GB, air-cooled, up to 8 per server. Dell ships XE9785L (8x MI355X) and XE7745 / R7725 with MI350P. | VENDOR CLAIM | blueprint §07 |
| I5 | Shared KV cache on MI300X: 2.5–5.4x faster first token with LMCache/Mooncake on ROCm. | MEASURED (scope) | blueprint §06 |
| I6 | Dell lab health report (R7725, 2x MI350P x2): PSU input loss, PCIe errors, no OS login — a blocker for on-premises runs. | INT | blueprint §06 |
| R1 | Releases v4.0–v4.4 (21 Sep -> 27 Oct 2026), Apache-2.0; v4.4.0 ships Windows amd64/arm64 zips. | FACT | EVIDENCE-SEED R1 |
| R2 | vLLM Semantic Router d3 live run (11 Oct): 204/204 test prompts routed to the intended tier, 104 ms p50, fail-closed verified; shipped in v4.2. | MEASURED (scope) | EVIDENCE-SEED R2 |
| R3 | OpenJev 93% classification on a 60-prompt test set, 193 ms decision latency. | MEASURED (scope) | EVIDENCE-SEED R3 |
| R4 | 85% lower model spend vs the frontier API bill in one deployment (~25 developers). The demo's 92.5% is vs GPT-6 Astra list for one team. | MEASURED (scope) | EVIDENCE-SEED R4 |
| R5 | Per-agent key revocation is effective on the next request (~2 s across replicas). Key revocation stops model access; it is not whole-agent containment. | MEASURED (scope) | EVIDENCE-SEED R5 |
| R6 | Tokenomics meter: reasoning + cache tokens, list vs billed cost, savings split by routing and caching. | FACT | EVIDENCE-SEED R6 |
| R7 | d3-ultra: open 27B multimodal decision model from AMD + the vLLM Semantic Router team, trained and served on Instinct, Apache-2.0, launched 26 Oct 2026. | VENDOR CLAIM | EVIDENCE-SEED R7 |
| R11 | Router built from source ran 90/90 against Lemonade. | MEASURED (scope) | blueprint §02b status table |
| R12 | A sensitive verdict narrows the allow list to owned tiers, so the frontier group cannot be selected for that request whatever its difficulty. | FACT | blueprint §02b |
| R13 | Local requests $0 vs $0.00223 per 30 cloud requests. | MEASURED (scope) | blueprint §05 ledger table |
| R-G | Guardrail (not a claim): the Router is a gateway, not the serving runtime. No guaranteed quality, no aggregate atomic dollar budgets, no automatic privacy from a classifier, no whole-agent containment from key revocation. Verify release/runtime/model rights and distinguish vendor claims, measured results, models and illustrations. | FACT | PROMPT.md "EVIDENCE THAT MUST SURVIVE"; EVIDENCE-SEED |
| F1 | Run 1: tool-call strict exact match 4.5% -> 32.0%, but MMLU-Pro collapsed 60.5 -> 13.0 (IFEval 81.5 -> 63.5). The preregistered gate refused promotion. | MEASURED (scope) | EVIDENCE-SEED F1; inputs/fine-tuning-report.md |
| F2 | The same frozen router export (1,186 / 221 / 300 traces) trained on MI325X and RTX PRO 6000 came out statistically level (semantic 78 vs 74/100); both run-2 candidates stayed in canary. | MEASURED (scope) | EVIDENCE-SEED F2; report |
| F3 | Run 1 trained on 1x NVIDIA RTX PRO 6000 Blackwell (Shadeform); GPU cost $3.19. | MEASURED (scope) | report §"run 1" |
| F4 | Governed capture produced an audited export of 1,186 / 221 / 300 train / validation / held-out traces with the uncaptured team verified absent. | MEASURED (scope) | report; blueprint §04d |
| F5 | Serving cost per harness success: $0.00142 (AMD MI325X) vs $0.00130 (NVIDIA RTX PRO 6000) on one replay harness, 300 held-out traces, serving only, excludes training. Training wall time 893 s (NVIDIA) vs 1,215 s (AMD) — the decks round the AMD run to "about 20 minutes". | MEASURED (scope) | report; seller-deck 13; update-v8 14 |
| F6 | Total measured fine-tuning spend ~$51 (GPUs ~$17, a short Fireworks deployment ~$32, teacher calls a few dollars). | MEASURED (scope) | report |
| F7 | Run-2 gates: NVIDIA exact 5/200 vs base 4/200 (+0.5 pp); AMD 7/200 vs base 9/200 (−1.0 pp); +10 pp required. | MEASURED (scope) | report |
| F8 | Router v4.3.0 validated and tagged 2026-10-24. | FACT | report |
| F9 | Teacher licensing: the provider's terms (§2.2(c)) prohibit using content to train AI models; real-teacher collection was stopped until a licensed teacher was chosen. | FACT | report |
| A1 | Scope: SQLite coding agents on pinned AMD EPYC 9R14 cores on AWS, Qwen3 Coder Next on Amazon Bedrock, 25–26 Oct 2026. It is not NPU, Halo or MI350P validation and does not establish Router uplift. | MEASURED (scope) | pilot #cpuonly, #model |
| A2 | Primary sizing question: CPU-only — the number of agents that keeps >80% of the allocated CPU busy, regardless of completion or verifier result. Latency-safe load and verified success are separate measures. | MEASURED (scope) | pilot #cpuonly |
| A3 | The 3 / 7 / 14 on-time loads were single waves, not validated sustained capacity. All six sustained studies were invalid (instrumentation faults); the held-out four-core prediction failed. | MEASURED (scope) | pilot #cpuonly, #holdout |
| A4 | Failed vs passed SQLite attempts: median 48 vs 47 core-seconds (p10–p90 44–103 vs 42–112). Failures cost real CPU. | MEASURED (scope) | pilot #outcomes |
| A5 | Zero 429s does not mean no model-response waiting. 0 of ~4,000 calls throttled; max relay admission wait 0.0002 s; agents still spent a large share of each run waiting on the model. | MEASURED (scope) | pilot #model, #results (V4) |
| A6 | $150.76 total ($59.19 compute + $91.57 model tokens) covers compute and model tokens across recovery, invalid and supplemental work, and excludes orchestrating-agent tokens. It is a study cost. Never divide it by SQLite successes. | MEASURED (scope) | pilot #cost |
| A7 | T1 = 184.7 s (median of the first three single-agent trials); SLO = 2 x T1 = 369.3 s (6.2 min). | MEASURED (scope) | pilot glossary |
| A8 | Valid whole waves above 80% busy: 5 and 8 agents on 1 core (81.9%, 83.9%); 12 agents on 2 cores (82.8%). No valid 4-core whole wave exceeded 80% (highest load tested: 24 agents). | MEASURED (scope) | pilot #cpuonly |
| A9 | Provisional per-agent estimate: 3 / 6 / 12 agents at >80% busy on 1 / 2 / 4 cores, ranges 2–5 / 4–9 / 7–17. | MODEL | pilot #cpuonly |
| A10 | Per-agent CPU demand: SQLite 0.276 cores; CompCert 0.543 cores (stated in the source decks as "about 0.28 of a core per coding agent on EPYC"). | MEASURED (scope) | pilot #formulas; seller-deck 4; update-v8 27; battlecard 1 |
| A11 | Latency-safe load (secondary): estimate 2 / 5 / 9 agents on 1 / 2 / 4 cores; largest valid single waves that met the target held 3 / 7 / 14 agents at 67% / 72% / 57% busy. | MEASURED (scope) | pilot #sc-latency |
| A12 | Registered verdicts: V1 FAIL (no load met >=90% busy with >=80% on time); V2, V3, V5 INCONCLUSIVE; V4 PASS (no throttling); V6 FAIL; V6′ FAIL. | MEASURED (scope) | pilot #results |
| A13 | Three repeats of 12 agents on 4 cores: 31% / 50% / 60% busy, 83% / 83% / 100% on time; one run used 54% more CPU work than another. | MEASURED (scope) | pilot #results |
| A14 | Verified success: 4 of 5 low-load trials passed the verifier; at >=12 agents the pass rate ranged 33–69%; the heaviest loads passed about a third. | MEASURED (scope) | pilot #outcomes |
| A15 | Sealed 4-core prediction (SHA-256 b95bbabb...eeeb81, recorded before any 4-core run) expected 62–67% on time at 4–14 agents; measured 83–100%. Scored as registered: V6 FAIL. | MEASURED (scope) | pilot #holdout |
| A16 | Model speed: 1 request -> 0.79 s p50 time to first token, 4.3 ms per output token, 167 tok/s per request; at concurrency 8 -> 0.79 s / 5.0 ms / 137 tok/s; 20/0 accepted/throttled in the probe. Applied quota 10,000 requests/min, 100M tokens/min. | MEASURED (scope) | pilot #model |
| A17 | Supplemental tasks (never pooled with SQLite): CompCert 573–1,984 core-seconds per attempt, 24–79% busy, 2 of 7 passed; pMARS 2–9% busy; POV-Ray 0–3% busy, 0 passed. | MEASURED (scope) | pilot #supplemental |
| A18 | Reproduction: final snapshot restored to a fresh directory with every file hash matched; offline rebuild with network denied reproduced all 136 CSV files byte for byte; the sealed prediction hash was re-scored. Teardown verified 0 running instances, volumes, security groups, key pairs or snapshots tagged for the pilot. | MEASURED (scope) | pilot #data, #cost teardown |
| A19 | Qwen3 Coder Next was chosen from six Bedrock coding models; it solved 4 of 5 low-load trials (95% interval 38–96%); gpt-oss-120b solved 2 of 5 and was dropped. | MEASURED (scope) | pilot #model, journal |
| A20 | Follow-up work identified by the pilot: more matched repeats per load, a constant-backlog test to separate cohort drain from capacity, a second core type, and a measured 3-core point. | PLAN | pilot #followup |
| P1 | Working theme: "The token bill is not the agent bill." Earlier blueprint framing: "Right token, right silicon" — retained only as a closing line. | PLAN | PROMPT.md; blueprint §01 |
| P2 | Illustrative workload: one 8x MI350P server ~69% occupied; owned beats elastic cloud above ~75% occupancy; payback ~1.6 months vs all-frontier (budgetary capex). | MODEL | EVIDENCE-SEED P2; blueprint §05b |
| P3 | the agent team internal: 11 leads + 89 named subagents = up to 100 agents (concept/estimate); a coordinator agent coordinates. Real lessons: NPU contention (39 s vs 0.45 s); a fine-tune candidate rejected by its gate; claims discipline; the first preview film of this plan was made in ~50 minutes by one agent. | INT | EVIDENCE-SEED P3; blueprint §08c |
| P4 | cost_per_accepted_job = sum of(attributable inference + tool/sandbox compute + retries and failed attempts + verification + operating overhead) ÷ verified completed jobs, with the latency requirement stated. | METHOD | PROMPT.md; blueprint §05 |
| P5 | NPU tier: $0 marginal, plus an optional device-share amortization; Dell's 5,000-seat model implies ~$50 per seat per year saved. | ILLUSTRATION | blueprint §05 |
| P6 | Strix Halo: $0.77–1.05 per M output tokens at 16 users and 25% busy. At 4 ops hours a month the box never beats a hosted API. | MODEL | blueprint §05 |
| P7 | XE7745 payback in about 10 weeks vs Opus at 90% cache hit. | MODEL | blueprint §05 |
| P8 | 16 interactive users per Halo box; 13.5 tok/s background generation on the NPU. | MEASURED (scope) | blueprint §05b |
| P9 | Proposed next step for the enterprise tier: a Dell/AMD lab path to MI350P agent-workload validation, with no catalog or commercial commitment. | PLAN | PROMPT.md slide 12 |
| G1 | AMD claims Helios delivers up to 30% more tokens per dollar than Vera Rubin NVL72. | VENDOR CLAIM | blueprint §07 |
| G2 | AMD reports that routing its own internal traffic to MI350P cut its token bill 43% (self-reported). | VENDOR CLAIM | blueprint §07b |
| G3 | Dell's message: "tokenomics up to 90% below public cloud". | VENDOR CLAIM | blueprint §07 |
| G4 | Dell's deskside agentic AI offer (NVIDIA-only) claims up to 87% lower cloud token cost. | VENDOR CLAIM | blueprint §07b |
| G5 | AMD claims up to 4.2x tokens per dollar for MI350P; MLPerf v6.1 debut. | VENDOR CLAIM | blueprint §07 |
| G6 | ROCm delivered +38% on the same MI355X between MLPerf rounds. | VENDOR CLAIM | blueprint §07 |
| G7 | Cache reads are 96–99% of agentic coding token spend, so any savings claim must use cached-input prices. | EXTERNAL (cited) | blueprint §07 |
| G8 | Goldman Sachs projects 24x token growth by 2030 from agents. | EXTERNAL (cited) | blueprint §07 |
| G9 | Forrester: open-weight models carry about half of tokens but ~5% of spend. | EXTERNAL (cited) | blueprint §07 |
| G10 | Dell's own survey: 59% cite unclear ROI on AI. | VENDOR CLAIM | blueprint §07 |
| G11 | Gartner: 40% of new AI data centers will be power-limited by 2027. | EXTERNAL (cited) | blueprint §07b |
| G12 | Dell AI Factory: 6,500+ customers. | VENDOR CLAIM | blueprint §07 |
| G13 | Ryzen AI PRO 400 NPUs reach up to 60 TOPS; Qualcomm X2 Elite claims 80 TOPS; Intel Panther Lake 50 TOPS. | VENDOR CLAIM | blueprint §07 |
| G14 | Schedule items: AMD/Dell follow-up 30 Oct; AMD Q3 earnings 3 Nov; CES 2027 6–9 Jan (AMD keynote unconfirmed); DTW 2027 24–27 May. | PLAN | blueprint §07b |
| G15 | AMD is planned to lead the Router v5.0 launch; Dell AI Factory validation dated 11 Dec. | PLAN | blueprint §06 |
| G16 | Governance context: OWASP Agentic Top 10; NIST agent-identity work; EU AI Act high-risk obligations point to Dec 2027; ~1 in 5 organisations have mature agent governance (Deloitte, via Dell). | EXTERNAL (cited) | blueprint §07b |
| G17 | Helios (MI455X, 72 GPUs per rack) is in production at OpenAI, Meta, Microsoft and Oracle; Helion is an OCP agentic rack with a built-in model router, partner configurations from November. | VENDOR CLAIM | blueprint §07 |
| I2c | ~2,000 concurrent coding agents per 8-card XE7745 node = 8 x the tested ~250 per MI350P. | MODEL | S11; V8 25, 26, 35; B2 |
| I2d | Sizing ladder: ~500 agents (R7725 + 2x MI350P, 288 GB HBM3E) · ~2,000 (R7725 + 1x XE7745, 8x MI350P) · ~4,000 (R7725 + 2x XE7745, 16x MI350P). | MODEL | S11; V8 26 |
| I4b | XE7745: 8x MI350P, 1,152 GB HBM3E total, 4U, air cooled, 4.8 kW GPU, 8 PCIe GPUs, 2x EPYC 9005. MI350P: 144 GB HBM3E, CDNA 4, 128 CU, 4 TB/s, PCIe Gen5 x16, 600 W. | VENDOR CLAIM | S8; V8 46, 53 |
| I4c | Dell's AI Platform with AMD names XE7745 and R7725 with MI350P and EPYC. The Dell catalog listing is in discussion, not live. | VENDOR CLAIM | S8; V8 22, 46; B2 |
| R14 | The router pod is CPU-only and requests no GPU. | FACT | S5; V8 38 |
| P7b | MODELED: one XE7745 (8x MI350P) pays back in ~9 weeks against a frontier model list at 24x7 — ~$3.0M/yr of API spend vs a ~$500K budgetary node. Basis: 250 concurrent agents, ~103 generating at any instant at 30 output tok/s each, 3,150 output tok/s sustained, 90% cache hit, $5/$25 per M tokens. Without caching ~$11M/yr. | MODEL | S7, S12; V8 28, 50; B2 |
| P10 | MODELED: at 5,000 developers and public $250–1,500 per developer-month spend, the measured 67–85% cuts imply a $10–77M yearly reduction. | MODEL | V8 2, 31 |
| P11 | MODELED: at 5,000 developers and $250/$500 per developer-month, API spend is $15M/$30M a year and the router cuts $10–12.8M / $20–25.5M. | MODEL | V8 31 |
| P12 | The economic proof must separate three things and never count the same saving twice: route more efficiently · substitute local serving (include the complete local platform, operations and the API spend that remains) · utilization. Owned tokens are not free. | METHOD | V8 50 |
| R15 | 100% of routed tokens land on a usage row; three ceilings per key (daily, monthly, lifetime); budgets reserve input plus the output cap before dispatch, then reconcile. | MEASURED (scope) | V8 18 |
| R16 | Salted PBKDF2-HMAC-SHA3-512 key verifiers (~4.7 us per cached request); upstream identity hidden from callers and error messages; credential values redacted from config, startup and CLI errors. | MEASURED (scope) | V8 20 |
| R17 | The Apache-2.0 router has been in production internally since June 2026. | MEASURED (scope) | V8 40 |
| R18 | Mandatory disclaimer: testing performed by Metrum AI; performance and cost claims are Metrum AI's and have not been independently verified by AMD. | FACT | B2; V8 46 |
| R19 | The quality floor is a prediction, not a guarantee. | FACT | S3; V8 51 |
| R20 | Boundary statements: a predicted floor is not a guarantee · regex PII filtering is not a guarantee · capture is off by default and export is admin-authorized and audited · the customer clears consent, retention, base-model licence and provider training terms. | FACT | S15; V8 51 |
| R21 | Frontier and open model names are as published in September–October 2026 (a frontier model, GPT-6 Astra, K2 Horizon 7B, Muse Glimmer 30B, DeepSeek V4.1 Flash, Gemma 4 E4B, Qwen3.6-35B-A3B, Qwen3.8 27B). | FACT | S9, S10, S12; V8 24, 49 |
| R22 | Community edition is Apache-2.0 with no licence key; the Enterprise edition adds production validation, named support and signed releases. | FACT | S20; V8 37 |
| S-M1 | Starter portfolio on one XE7745, published checkpoint footprints, not measured VRAM, and not a validated partition: K2 Horizon 7B ~18 GB (TP=1, ~126 GB gross remainder) · Muse Glimmer 30B ~60 GB · Qwen3.8 27B ~56 GB x2 (TP=1 replicas) · DeepSeek V4.1 Flash ~510 GB FP8/FP4 (TP=4, ~66 GB gross remainder — the tightest budget on the node). | VENDOR CLAIM | S9; V8 23, 49 |
| R2b | d3 boundary: the live run was d3 26B in enforce mode, fail-closed verified when d3 was unreachable. It does not establish typical savings (1.4% on the seed set, 25% on one 8-turn take) or long-context latency. | MEASURED (scope) | V8 47; S13 |
| R3b | OpenJev boundary: live run 25 Sep, 60 prompts x 2 on one RTX PRO 6000, 0 of 60 repeat flips; spend projected from a 15-call price sample. It does not establish production accuracy, Instinct performance, or commercial rights. | MEASURED (scope) | V8 47; S13 |
| R4b | Savings-report boundary: 1,379 requests over 30 days, one developer, three list-price baselines, $58.06 vs $772.18 -> 81%–92.5% lower cost; the baseline is list price with no cached-input discount, so at Astra's cached-input rate the saving is about 67%. The team-wide figure is 85%. | MEASURED (scope) | S16; V8 47; B2 |
| R23 | Neither decision model (d3, OpenJev) has been measured on Instinct yet; both ran on one NVIDIA RTX PRO 6000. | MEASURED (scope) | S13, S14; V8 47 |
| F4b | One audited export contained 1,707 traces (1,186 train / 221 validation / 300 held-out); the uncaptured team was verified absent. | MEASURED (scope) | S13; V8 13 |
| F10 | 0 of 2 candidates were promoted: the loop is proven, a better model is not yet. | MEASURED (scope) | S15; V8 13; B2 |
| F11 | LoRA training on governed, router-exported traces took about 20 minutes per run on Instinct MI325X (1,215 s; the NVIDIA run was 893 s). | MEASURED (scope) | S14; V8 14 |
| G18 | vLLM Semantic Router x AMD is an AMD-backed open project, not an AMD product (collaboration 16 Dec 2025; six-model playground Jan 2026; arXiv 2603.12646, Mar 2026: 50 ms, under 800 MB on MI300X). | EXTERNAL (cited) | V8 42, 44 |
| G19 | Read from public NVIDIA repositories on 17–18 Sep 2026: NeMo Switchyard pre-1.0 (v0.3.0, 22 Sep), standalone proxy demo/evaluation only, credentials and budgets in a host gateway; LLM Router Blueprint deprecated; Dynamo KV-aware router at the worker layer. | EXTERNAL (cited) | S19; V8 33, 34, 35 |
| G20 | Hosted routers offer many models with no hardware, but prompts and the routing decision leave the walls; gateways do keys, budgets and load balancing well. OpenRouter can sit behind Metrum as an approved provider. | EXTERNAL (cited) | S19; V8 34 |
| G21 | The Linux Foundation's Tokenomics Foundation is standardising how AI cost and value are counted; the router's meter maps to its roadmap — a mapping, not a membership claim. | EXTERNAL (cited) | V8 30 |
| C1 | cost per accepted job = ( inference + tools and sandbox + retries and failed attempts + verification + operating overhead ) ÷ accepted jobs, with the latency requirement stated. | METHOD | PROMPT.md; COST-MODEL.md §1 |
| C2 | Illustrative workload (blueprint reference): 500 seats · 4 always-on agents per seat (2,000 agents) · 17 h/day · 30 days · 3% escalation to frontier · 12.6 tok/s sustained per agent · 46.3B agent output tokens/month · 12.1M jobs/month · ~3,704 output + ~74,088 input tokens per accepted job · node occupancy 68.7%. | ILLUSTRATION | COST-MODEL.md §3 |
| C3 | All-frontier: $419,451/month for the modelled workload at GPT-6 Terra list ($1/$5 per M, 20:1 agent input, 90% cached at 10%). | MODEL | model §4.1 |
| C4 | Hybrid on-premises: $33,111/month all-in = node $16,350 + 32 Halo boxes $4,475 + cloud burst $0 + 3% escalation $12,287. 92.1% below all-frontier. | MODEL | model §4.1 |
| C5 | All-cloud (elastic): $34,644/month (elastic MI325X GPUs at $3.80/GPU-hr + hosted open models at $0.95/M + escalation). 4.4% above hybrid. | MODEL | model §4.1 |
| C6 | Break-even occupancy vs elastic cloud GPUs: 74.7% (~75%) — the node's all-in monthly cost equals the cloud cost of the same agent-hours. | MODEL | model §4.2 |
| C7 | Payback vs all-frontier: 1.6 months. | MODEL | model §4.2 |
| C8 | Payback vs all-cloud: 33.1 months. | MODEL | model §4.2 |
| C9 | On-prem cost per M output tokens at the modelled 68.7% occupancy: $0.055/M marginal (power + ops only) and $0.364/M all-in (including capex ÷ 36). | MODEL | model §4.3 |
| C10 | On-prem $/M output tokens: at 40% busy $0.094 marginal / $0.626 all-in; at 90% busy $0.042 / $0.278. | MODEL | model §4.3 |
| C11 | Cost receipt, frontier-priced: one accepted job = $0.0359 (inference $0.0326 + tools $0.0024 + retries $0.0008). | MODEL | model §4.4 |
| C12 | Cost receipt, owned-priced at 68.7% busy: $0.0036 marginal / $0.0057 all-in per accepted job — ~9.9x below the frontier-priced job. Tool, retry and verification terms are identical across both columns. | MODEL | model §4.4 |
| C13 | Retry lever: a failure rate of 20% / 33% / 50% / 67% multiplies the cost per accepted job by 1.25x / 1.49x / 2.00x / 3.03x. | MODEL | model §4.5; A14 |
| C14 | Escalation lever: 1% / 3% / 10% / 25% of output tokens to a frontier API -> hybrid monthly $24,920 / $33,111 / $61,780 / $123,212. The single biggest lever on the hybrid bill. | MODEL | model §4.5 |
| C15 | Cached-input lever: 0% / 50% / 90% / 99% cached input -> all-frontier monthly $1,168,980 / $752,575 / $419,451 / $344,498. Cached input dominates the frontier price. | MODEL | model §4.5; G7 |
| C16 | Price and cost inputs (from BLUEPRINT-REFERENCE.py): node capex $500,000 budgetary — not a quote; Halo box $3,999 x 32 boxes (1 per 16 users) -> total capex $627,968; DO MI325X $3.80/GPU-hr with no provisioning factor; GPT-6 Terra list $1/$5 per M tokens; agent input:output 20:1; 90% of agent input cached at 10% of list; electricity $0.12/kWh, PUE 1.4; node power 8 x 282 W measured + 1.5 kW host; ops $2,000/server/month + $20/Halo box; amortisation 36 months; 250 agents per MI350P, 12.6 output tok/s per agent. | MODEL | model §2 |
| C17 | 92.5% lower cost vs GPT-6 Astra list for one team over 30 days in the demo (1,379 requests, $58.06 vs $772.18); at Astra's cached-input rate the saving is about 67%. | MEASURED (scope) | R4; S16; V8 47 |
| C18 | 85% lower model spend vs the frontier API bill, measured on Metrum's own team (~25 AI developers, one deployment). | MEASURED (scope) | R4 |
Our own project budget and production spend — the campaign budget, the agent, model and voice-generation spend for producing this work, and any build-bill figure (claim IDs F3 and F6). The deck shows what the customer's workload costs, never what the chase cost us.