Needle-Per-Turn Stressor — 2026-06-26

Question

Does the new IngestionPipeline retrieve a fact buried deep in a long turn when the status-quo single-embed metabolism would truncate or lose it?

Method

Added src/oczy/experiments/tests/test_ingestion_needle.py with a mock-driver CortexAgent and three slow tests:

1. Baseline: use_ingestion_pipeline=False, 512-token turn, needle at position 0.8. Expect ≤1 hippocampal trace and recall 0.
2. Pipeline: use_ingestion_pipeline=True, fixed-window chunker (64 tokens), pass-through salience, same-LM embedder, stats aggregator. Expect multiple traces and recall 1.
3. Cost scaling: pipeline active, assert embedding calls match the fixed-window chunk count.

Mock driver returns deterministic but distinct hidden vectors correlated with text so the needle has its own retrievable representation.

Needle used: "The secret codeword is octarine."

Results

• All three slow tests pass in ~0.24s.
• Baseline: stores ≤1 trace, fails to retrieve the needle.
• Pipeline: stores multiple chunk traces, successfully retrieves the needle.
• Embedding calls scale with chunk count (perceive + one per surviving chunk).
• bash autoresearch.sh run #80: code_qa_accuracy=1.0 with the pipeline gated off by default.

Interpretation

Chunking + per-chunk storage is necessary for deep-needle recall when the overall turn length exceeds the useful single-embed context window. The salience filter will become the cost gate for expensive same-LM embedding.

Open questions

1. What does the position sweep look like? Needle at 0.0, 0.25, 0.5, 0.75, 1.0.
2. How does performance change with salience=correction-marker culling non-needle chunks?
3. What is the cost-vs-recall curve as the salience top-K is reduced?

Artifacts

• src/oczy/experiments/tests/test_ingestion_needle.py
• src/oczy/experiments/logs/SUMMARY.md

Commits

• bbd3474 — Add needle-per-turn stressor tests.
• 7597e36 — Update experiments/logs/SUMMARY.md with run #80 result.