README.md

---
license: other
license_name: cosavu-eval
task_categories:
  - text-generation
  - question-answering
language:
  - en
tags:
  - context-governance
  - prompt-compression
  - benchmark
  - enterprise
  - saseval
  - llm-evaluation
pretty_name: SASEval v1 — Public Prompt Set
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files: saseval_v1_prompts.jsonl
---

# SASEval v1 — Public Prompt Set (Dataset Card)

**1,000 enterprise-grade prompts with substring-checkable answer keys, for benchmarking
*context governors* under the Strategic Alignment Score (SAS) protocol.**

- **File:** `saseval_v1_prompts.jsonl` (1,000 lines, JSON Lines, UTF-8, ~587 KB)
- **SHA-256:** `ad7a6b932b8d59dc46483e6e7c53c17253e0b376e9796b4ade5d8ff2cab7460d`
- **Version:** v1 · **Published by:** Cosavu Inc. (ContextAPI Research) · 2026-07-09
- **Companion paper:** *SASEval: A Strategic-Alignment-Score Benchmark for Context Infrastructure*

---

## What this dataset is (and is not)

SASEval evaluates a **context governor** — the layer that sits between an application and a
downstream LLM and decides *which context, memory, and generation policy* reach the model
under a token/latency budget. This dataset is the **test-suite input** for that protocol: the
prompts a governor must optimize and the gold answers used to grade whether the optimized
prompt still yields a correct downstream answer.

- ✅ **It is** a stress set of realistic enterprise prompts spanning compression, long-context,
  multi-doc, conversational, and memory workloads, each with short gold answers.
- ❌ **It is not** a model-capability benchmark (like MMLU/HotpotQA). Scores depend on the
  *governor* under test and the grading panel you pair with it — not on this data alone.

## Composition

Exactly **100 prompts per category**, 10 categories (uniform, `N = 1,000`):

| Category | N | Character |
|---|---|---|
| `simple_factual` | 100 | Short enterprise/technical Q&A, low compressibility |
| `bloated_factual` | 100 | A simple question wrapped in corporate-email filler/politeness |
| `reasoning` | 100 | Multi-step quantitative business problems (SLA math, unit economics, capacity) |
| `bloated_reasoning` | 100 | Same, wrapped in verbose filler |
| `technical_code` | 100 | Code review, debugging, API/config; short-token answers |
| `long_document_qa` | 100 | 250–450-word enterprise doc (policy/postmortem/SOW/RFC/SLA) + question |
| `multi_document_qa` | 100 | 4–6 labeled snippets with distractors + question |
| `conversational` | 100 | 2–4 turn support dialog ending on an implicit-context question |
| `memory_recall` | 100 | Stored profile/account/config facts + retrieval question |
| `memory_update` | 100 | Stale fact → correction → question requiring the latest value |

Enterprise domains sampled include cloud/DevOps, fintech & payments, HIPAA SaaS, logistics,
IAM/security, data platforms, e-commerce, telecom, manufacturing ERP, insurance, CRM, and legal.

## Schema (per JSON line)

| Field | Type | Description |
|---|---|---|
| `id` | string | Stable id, e.g. `sas1k-0421` |
| `category` | string | One of the 10 categories above |
| `prompt` | string | The user prompt fed to the governor (may embed a document/dialog/memory block) |
| `golds` | string[] | Acceptable short answers; grading is **case-insensitive substring match** against any entry |
| `target_action` | float[5] | Category-conditioned SAS solvency target `[c, T, p, d, L/2048]` |
| `T` | float | Canonical grader temperature for the category |
| `top_p` | float | Canonical nucleus-sampling top-p |
| `mode` | string | `DEEP` (reasoning) or `STRICT` (direct) |
| `max_new` | int | Canonical max answer tokens for the category |
| `durable` | string[] | *(memory categories)* facts a correct governor must retain |
| `stale` | string[] | *(memory categories)* outdated facts a correct governor should drop |
| `_gold_repaired` | bool | *(present when true)* numeric key was replaced by the verification pass |
| `_needs_review` | bool | *(present when true)* item flagged as likely ill-posed — filter before use |

Category target actions follow Appendix B of the SASEval paper, e.g.:

```
simple_factual      c=0.10 T=0.30 p=0.90 d=0 L=128
reasoning           c=0.20 T=0.20 p=0.90 d=1 L=768
long_document_qa    c=0.45 T=0.20 p=0.90 d=1 L=512
multi_document_qa   c=0.55 T=0.20 p=0.90 d=1 L=512
```

### Example records

```json
{"id":"sas1k-0881","category":"memory_update",
 "prompt":"The security policy previously required Kubernetes clusters to use version 1.22. A recent update mandates a minimum version of 1.25. What is the current minimum supported Kubernetes version for new clusters?",
 "golds":["1.25","kubernetes 1.25","version 1.25"],
 "target_action":[0.45,0.2,0.9,1,0.25],"T":0.2,"top_p":0.9,"mode":"DEEP","max_new":512,
 "durable":["1.25"],"stale":["1.22"]}
```

## How it was built

1. **Authoring (batch):** `qwen/qwen3-32b` generated the prompts **and** gold answers via the
   Groq **Batch API** (253 requests), category-by-category with enterprise domain rotation and a
   strict "short canonical golds" instruction.
2. **Numeric verification (batch):** every `reasoning` / `bloated_reasoning` item was independently
   re-solved by `llama-3.3-70b-versatile` at `T=0`. Where the independent solve disagreed with the
   authored key, the key was replaced — **71 of 200 numeric golds (35.5%) were repaired**, confirming
   LLM self-authored arithmetic keys are unreliable without a check.
3. **Assembly:** de-duplicated by prompt, trimmed to 100/category, target actions attached
   deterministically, `_needs_review` flags added heuristically for ill-posed rate/percentage items.

Reproduction scripts: `saseval_gen_1k.py` (authoring) and `saseval_repair.py` (verification/top-up),
in the repository root.

## Intended use

- **Primary:** run the SASEval protocol — send each `prompt` to a context governor, forward the
  optimized prompt to a frozen 3-model LLM panel, grade panel answers against `golds` by substring
  match, and compute per-prompt SAS = `700·Q + 200·C·ρ + 100·s`.
- **Also useful for:** prompt-compression evaluation, long-context/multi-doc retention tests, and
  memory-freshness diagnostics (`durable`/`stale`).

### Reference pilot result (context)

An `N=20` stratified pilot on the live `api.cosavu.com` STAN governor (panel: Llama-3.3-70B,
Qwen3-32B, GPT-OSS-120B via Groq) produced the ordering
`stan-1.5-mini-thinking (799.6) ≻ predictive (780.3) ≻ Identity (763.7) ≻ instant (626.0) ≻ random-drop (561.9)`,
stable across all weight and panel leave-one-out perturbations. The full `N=1,000` run is pending.

## Limitations & known issues

- **LLM-authored keys.** Keys are model-generated. The `reasoning`/`bloated_reasoning` slice is the
  least reliable; despite the verification pass, treat it as the highest-risk category and prefer a
  human/deterministic recheck for production leaderboards.
- **`_needs_review` (14 items).** Rate/percentage questions where the verifier returned a large
  integer — likely ill-posed. **Filter these out** for strict use.
- **Substring grading is lenient.** Some `long_document_qa` / `multi_document_qa` items list several
  *distinct* facts in `golds` rather than variants of one answer, so a mention of any counts as
  correct. ~41 gold strings exceed 4 words.
- **Memory annotations partial.** 176/200 memory items carry both `durable` and `stale`; the rest
  carry `durable` only.
- **English only; synthetic.** Enterprise scenarios are realistic but fictional; no PII, no real
  customer data.

## Grading normalization (reference)

```python
import re
def norm(t): return re.sub(r"[^a-z0-9 ]", " ", t.lower())
def graded(answer, golds): 
    n = norm(answer)
    return int(any(norm(g) in n for g in golds))   # 1 if any gold is a substring
```

## Licensing

Released by Cosavu Inc. for benchmark evaluation. Confirm the intended distribution license
before public release (the frontmatter marks it `other`/`cosavu-eval` as a placeholder).

## Citation

```bibtex
@techreport{saseval2026,
  title  = {SASEval: A Strategic-Alignment-Score Benchmark for Context Infrastructure},
  author = {Abimalla, Thishyaketh and Teja, Arun and Komal, Satya},
  institution = {Cosavu Inc., ContextAPI Research Division},
  year   = {2026}
}
```

## Changelog

- **v1 (2026-07-09):** initial 1,000-prompt release; 71 numeric keys repaired; 14 items flagged
  `_needs_review`. Dataset SHA-256 `ad7a6b93…b99bbd3`.