# Cubie Evidence Pack — intc-v1.0

**Version:** intc-v1.0
**Scope:** Published-trace receipts behind every canonical number cited by `demo/dashboard.html` (66.85% TP4 blast radius, 87.53% DCGM blind spot, 33.15%/83.29% productivity split, 2.51× multiplier).
**Operation:** `cp -p` (timestamps preserved; sources untouched and verified by mtime).
**File-name prefix:** every file begins with `intc-v1.0-` for unambiguous version anchoring.

## Why these files exist here

The dashboard's hero comparison, the MONITORING BLIND SPOT panel, and the denial-reasons table all reference numbers whose ground truth lives in these JSON files. Embedding versioned copies inside the repo means the dashboard can cite a stable, hashed source even if the upstream analysis folder is later moved or superseded by a v1.1 / v2.0 pack.

## Update policy

- **APPEND ONLY.** Never edit these files in place. If the upstream analysis is updated, create `intc-v1.1/` alongside this directory and re-cite from the dashboard.
- The dashboard's footer carries the active version pointer; bumping versions is one constant in `dashboard.html` plus a new `intc-vX.Y/` directory.

## File inventory + sha256

| Filename | Description | sha256 |
|---|---|---|
| `intc-v1.0-tp4_nvlink_results.json` | Alibaba cluster-trace-v2026-GenAI TP4 NVLink cascade analysis — 157,411 GPU intervals, 141 LLM containers. Source of 66.85% blast radius, 45.54% NVLink-waiters, 51.76% VRAM-locked. | `7d1a90bfc6036e7d6eb8483ba1ddb4bf69cc69003a306016368d221286119bc9` |
| `intc-v1.0-hunter_results.json` | DCGM threshold sensitivity analysis. Source of 4.80% (tight DCGM), 8.34% (wide DCGM), 87.53% blind spot derivation. | `6fb720db2600441e6a7155d4a2ac6ee8cdc023de65fceef016edc7742d979685` |
| `intc-v1.0-stage_1_prometheus_metrics.json` | Reference vLLM Prometheus output from a 4,071-request load test; format anchor for `/prom` exposition. | `c6be0f7016a5bcabde05af446494c1526307ca10a6a9e2436f0c0f5e3f7e6723` |
| `intc-v1.0-alibaba_micro_stall_metrics.json` | Alibaba per-container micro-stall metrics. | `e31c1c1f8bb35ec8e12f038948a4a5d27ebc4880a308ab6ff3424b89616020ad` |
| `intc-v1.0-alibaba_v2026_squatter_metrics.json` | Alibaba v2026 GPU-squatter pattern metrics. | `2e4adf2dcd79233d603e1c68498434a469413e292b8843bca7bc5131844d43ad` |
| `intc-v1.0-azure_2023_unconstrained.json` | AzureLLMInferenceTrace 2023 unconstrained threshold sweep. | `123ef65809d69f40d90ce95c689ad9ea27493d85fb41dffe7a6d1d88781cdb26` |
| `intc-v1.0-azure_2024_unconstrained.json` | AzureLLMInferenceDataset2024 unconstrained threshold sweep. Source of per-dataset deny-rate calibration. | `4a2a451e4e104bc543cce67aae42f8446513ea7160c7c32a3a6a880c2bfada06` |
| `intc-v1.0-lmm_2025_unconstrained.json` | AzureLMMInferenceTrace multimodal 2025 sweep. | `13c841e8e87c746a6f6ffc3e3363bebd9b653e2c157e27e3e43ecaaf3f43a0b7` |
| `intc-v1.0-lmm_2025_sota_2026_findings.json` | LMM-2025 SOTA-2026 derived findings. | `d6b846a315e90ad4c81f5e7d9a08eb0bae4cea1120629a3f49a792936e42be49` |
| `intc-v1.0-maveriq_unconstrained.json` | MaverIQ fingerprint trace sweep. | `7f11991e4375c67ea8800584d6f47f0726f1b3ab0d4c942d1bbb15993835f90d` |

## Canonical numbers extracted (cited from this version by the dashboard)

| Number | Value | Source file → JSON path |
|---|---|---|
| TP4 full blast radius (capacity destroyed) | **66.8479%** | `intc-v1.0-tp4_nvlink_results.json` → `full_tp4_blast_radius.capacity_pct` |
| TP4 blast intervals (raw count) | **105,226** | `intc-v1.0-tp4_nvlink_results.json` → `full_tp4_blast_radius.intervals` |
| NVLink waiters (invisible to DCGM) | **45.5356%** / 71,678 intervals | `intc-v1.0-tp4_nvlink_results.json` → `nvlink_waiters.*` |
| GPU 0 deadlock (DCGM-visible portion only) | **5.8446%** / 9,200 intervals | `intc-v1.0-tp4_nvlink_results.json` → `gpu0_deadlock.*` |
| DCGM tight threshold catches | **4.7983%** | `intc-v1.0-hunter_results.json` → `alibaba_signals.dcgm_thrash_proxy.tight_dram90_sm5.pct` |
| DCGM wide threshold catches | **8.338%** | `intc-v1.0-hunter_results.json` → `alibaba_signals.dcgm_thrash_proxy.wide_dram80_sm20.pct` |
| Cubie catches (true exact-zero-SM total) | **66.8479%** | `intc-v1.0-hunter_results.json` → `alibaba_signals.dcgm_thrash_proxy.exact_zero_sm.pct` |
| **DCGM blind spot** = 1 − (8.338 / 66.8479) | **87.53%** | derived |
| Pattern persistence | **1,440 / 1,441 timestamps = 99.93%** | `intc-v1.0-tp4_nvlink_results.json` → `tp4_costal_pattern.*` |
| VRAM locked at any moment | **51.76%** | `intc-v1.0-tp4_nvlink_results.json` → `vram_locked_by_tp4.pct_total_vram_locked` |
| Total intervals analyzed | **157,411** | `intc-v1.0-tp4_nvlink_results.json` → `total_llm_intervals` |
| LLM containers in trace | **141** (134 with GPU 0 deadlock + 127 with waiters) | `intc-v1.0-tp4_nvlink_results.json` → `*.containers` |

### Per-dataset cascade-trigger rates (used to calibrate the demo's per-dataset cards)

| Dataset | Subset | Threshold | % | Source |
|---|---|---|---|---|
| Azure LLM 2024 | code (16.8M rows) | < 5 output tokens | 36.59% | `intc-v1.0-azure_2024_unconstrained.json` → `code.low_output_by_threshold.lt_5_tokens.pct` |
| Azure LLM 2024 | code | < 15 output tokens | 68.22% | `azure_2024_unconstrained.json` → `code.low_output_by_threshold.lt_15_tokens.pct` |
| Azure LLM 2024 | conv (27.3M rows) | < 5 output tokens | 10.97% | `azure_2024_unconstrained.json` → `conv.low_output_by_threshold.lt_5_tokens.pct` |
| Azure LLM 2024 | conv | < 15 output tokens | 25.33% | `azure_2024_unconstrained.json` → `conv.low_output_by_threshold.lt_15_tokens.pct` |

## Verification

```powershell
cd C:\Users\NickV\cubie-tf\demo\evidence\intc-v1.0
Get-ChildItem intc-v1.0-*.json | ForEach-Object {
    $h = (Get-FileHash -Algorithm SHA256 $_).Hash
    "$h  $($_.Name)"
}
```

Each hash should match its row in the inventory table above. If a hash drifts, **do not modify this file** — create `intc-v1.1-MANIFEST.md` documenting what changed.