signaldepth/e15-context-budget

SignalDepth E15 Context Budget

This is a small prompt-sensitivity benchmark slice for separating two explanations that often get conflated:

1. the prompt is too short
2. the task contract is underspecified

The narrow result: on this deterministic Python code-task suite, making sparse prompts longer did not help. Making the task contract explicit did.

Key Result

Condition	Average pass rate	Read
short_sparse	0.25	short and underspecified
long_sparse	0.25	longer, same missing contract
short_dense	1.00	short but contract-explicit
long_dense	1.00	longer and contract-explicit

Length marginal: short = 0.625, long = 0.625.

Density marginal: sparse = 0.25, dense = 1.00.

Scope

• deterministic Python code tasks
• 4 local-model run archives
• 4 tasks
• 4 prompt conditions
• 192 graded calls
• k=3
• temperature 0.0

This is not a general long-context benchmark, not a model leaderboard, and not evidence about non-code task families.

Conditions

• short_sparse: short prompt, task contract underspecified
• long_sparse: longer prompt, same missing task contract
• short_dense: short prompt with explicit I/O and edge-condition constraints
• long_dense: longer prompt with explicit task contract

Dense here means contract-explicit, not merely wordy.

Files

• data/e15_summary.json: extracted E15 aggregate from SignalDepth public findings
• data/e15_conditions.csv: condition-level pass rates
• data/e15_models.csv: model-by-condition pass rates
• data/e15_tasks.csv: task-by-condition pass rates
• data/e15_failure_modes.csv: sparse-prompt failure counts
• schemas/experiment.schema.json: public experiment schema
• docs/context-budget.md: method/result note
• docs/methodology.md: benchmark methodology
• assets/context-budget.svg: chart

Dataset Viewer configs are conditions (default), models, tasks, and failure_modes. The JSON summary is included as the canonical aggregate object and is not mapped to a viewer table.

Reproduce Comparable Runs

The public harness is available here:

https://github.com/signaldepth/prompt-sensitivity-bench

Example:

cd harness
uv run python validate.py e15 --model-name qwen3.5:4b --k 3

Fresh runs should be read as replications on the same prompt family, not byte-for-byte replays of the archived local runs.

Raw Data Boundary

This dataset publishes the derived public E15 summary and reproducibility scaffold. Full raw per-trial archives remain local/private unless curated for a later release.

That boundary is intentional: the artifact is meant to make the condition design, aggregate result, and public harness easy to inspect without implying a full raw-output dump.

Citation

If you use this result, cite the Hugging Face dataset and the GitHub repository:

• Dataset: signaldepth/e15-context-budget
• Code: https://github.com/signaldepth/prompt-sensitivity-bench