Scenario: Investigating an Anomaly Mid-Pipeline

Status: draft Date: 2026-04-08 Project: pixecog Starting state:

• Same as end of TTL removal scenario — ttl-removal-validated is complete
• SWR detection pipeline is running on all subjects
• Sub-03 shows unexpected results during detection stage

What the researcher wants: Understand why sub-03's SWR detection rate is anomalously low and decide on a fix.

---

Phase 1: Reproduce and scope

Human: The SWR detection rate for sub-03 is way too low — 2/min when
       literature says 10-15/min. What's going on?

Behind the scenes

→ questio_status() — confirms swr-detection is in_progress
→ pipeio_target_paths("sharpwaveripple") — locates sub-03 outputs
→ reads sub-03 detection output — confirms 2.1 events/min
→ reads sub-01 and sub-02 for comparison — 12.3/min and 11.8/min (normal)

→ note_create(kind="idea", tags=["observation", "investigate"],
              title="Investigate: sub-03 SWR rate anomaly (2.1/min vs expected 10-15)")

Agent: Confirmed: sub-03 SWR detection rate is 2.1/min (sub-01: 12.3, sub-02: 11.8).

       **Scoping:** the issue is sub-03 only. Other subjects are in normal range.

       **Checking upstream:** let me verify the preprocessing output for sub-03.

---

Phase 2: Check upstream

Behind the scenes

→ pipeio_target_paths("preprocess_ieeg", mod="ttl") — reads sub-03 QC log
→ reads the preprocessed signal for sub-03 — checks spectral content

→ note_create(kind="idea", tags=["observation", "investigate"],
              title="Investigation: sub-03 preprocessing looks clean, issue is downstream")

Agent: **Upstream check:** TTL removal QC for sub-03 looks normal (26.7 dB reduction,
       -0.2 dB distortion). Preprocessed signal is clean.

       **Narrowing:** the issue is not preprocessing — it's in the detection stage.
       Let me check the detection parameters and sub-03's signal characteristics.

Diagnostic strategy

The agent follows the investigate loop: reproduce → scope (sub-03 only) → check upstream (preprocessing is clean) → trace downstream. Each step narrows the search space.

---

Phase 3: Root cause analysis

Behind the scenes

→ pipeio_mod_context("sharpwaveripple", mod="detect_swr")
→ reads sub-03 raw signal characteristics (amplitude, noise floor)
→ paper_context("@sirota_2003") — ripple amplitude expectations

→ note_create(kind="idea", tags=["observation", "investigate"],
              title="Root cause: sub-03 has lower amplitude ripples, threshold too high")

Agent: **Root cause identified:** sub-03 has lower overall signal amplitude
       compared to other subjects (likely electrode placement). The detection
       threshold (3.5 SD) is appropriate for sub-01 and sub-02 but too
       aggressive for sub-03.

       **Evidence:**
       - sub-03 mean ripple amplitude: 85 μV (sub-01: 145 μV, sub-02: 130 μV)
       - At 3.5 SD threshold, sub-03 misses ripples in the 80-100 μV range
       - Literature reports ripple amplitudes vary 2-3x across recording sites

       **Options:**
       1. Lower threshold to 3.0 SD for sub-03 (per-subject tuning)
       2. Use adaptive thresholding based on local noise floor
       3. Accept lower detection rate for sub-03 as a biological difference

       Which approach do you prefer?

---

Phase 4: Decision and record

Human: Option 2 — adaptive thresholding. But let's iterate on that in a
       separate session. Record what we found.

Behind the scenes

→ note_create(kind="idea", tags=["observation", "investigate", "decision"],
              title="Decision: use adaptive thresholding for SWR detection")
  Body: sub-03 has lower amplitude ripples due to electrode placement.
        Fixed threshold misses events. Human chose adaptive thresholding.
        To be implemented in next session.

Agent: Recorded. Three observation notes from this investigation, plus the
       decision to use adaptive thresholding. This will feed into the
       iterate loop when you're ready to implement it.

Investigation without pre-scripted notebooks

Pure investigate loop. The agent followed the diagnostic strategy: reproduce → scope (sub-03 only) → check upstream (preprocessing is clean) → trace to the detection stage → identify root cause (amplitude difference + threshold sensitivity) → present options → human decides. No pre-scripted notebook — the agent reasoned over tool outputs.

---

Ecosystem coverage

Subsystem	Tools used	Purpose in this scenario
questio	questio_status	Confirm milestone is in progress
pipeio	pipeio_target_paths, pipeio_mod_context	Locate outputs, inspect detection code and config
biblio	paper_context	Literature expectations for ripple amplitudes
notio	note_create (observation × 3, decision × 1)	Investigation trail and decision recording

Loop patterns used

Phase	Loop	Pattern
1. Reproduce and scope	Investigate	Confirm symptom, determine scope (sub-03 only)
2. Check upstream	Investigate	Rule out preprocessing as cause
3. Root cause	Investigate	Compare signal characteristics + literature → identify threshold sensitivity
4. Decision	(human)	Human picks adaptive thresholding, defers implementation

Recording trail produced

Note	Type	Content
observation-1	idea [observation, investigate]	SWR rate anomaly: 2.1/min vs expected 10-15
observation-2	idea [observation, investigate]	Preprocessing clean, issue is downstream
observation-3	idea [observation, investigate]	Root cause: lower amplitude ripples + fixed threshold
decision-1	idea [observation, investigate, decision]	Use adaptive thresholding (to implement next session)

Key insight

This scenario shows the investigate loop in isolation — a focused diagnostic workflow without any pipeline execution or milestone advancement. The agent narrows the problem systematically, presents options with evidence, and the human makes the engineering decision. The investigation trail is preserved in notes so the next session can pick up where this one left off.