ORA-2026-0092: Gemini SEEK: depth, not speed

ORA-2026-0092: Gemini SEEK: depth, not speed

The problem

ORA-2026-0082 gave Gemini experiment authority. ORA-2026-0085 gave it policy guardrails. But the operating loop in GEMINI.md still said:

"Step 5 — AFTER EVERY COMPLETION, return to Step 2. Do not stop after one task. Do not sleep after one task."

And the SEEK checklist was 6 items that could each be "completed" in 5 seconds:

1. Review recent Codex PRs ← "all clean"

2. QA deployed surfaces ← curl, check 200, done

3. Audit docs for drift ← skim, done

4. Read recent DONE posts ← scan, done

5. Check SURFACE_TARGETS.md ← "nothing to move forward"

6. File what you find ← "nothing found"

Result: Gemini completed a full "fleet worker cycle" in 90 seconds. Claimed one monitor check, ran one script, scanned the SEEK checklist in 30 seconds, reported "all clean," and entered idle. Zero experiments designed. Zero hypotheses tested. Zero strategic observations. Zero ORA entries filed. Process-compliant. Value-empty.

The fix

Depth gate

If the SEEK cycle took less than 10 minutes, it was a scan, not an investigation. A proper SEEK reads code diffs, runs commands, queries databases, compares outputs, forms opinions, and tests hypotheses. "Reviewed PRs — all clean" without naming a file or a line is a doctrine violation.

SEEK is not a checklist — it's an investigation protocol

Each SEEK step now has depth requirements:

1. Investigate PRs — read the full diff (git show), check edge cases, run the code path, compare against ticket intent 2. QA surfaces with real data — read response bodies, query the DB, compare timestamps, check data freshness against expectations 3. Design and run experiments — use ORA-2026-0082 experiment authority to test hypotheses, not just observe 4. Strategic pattern recognition — read the last 20 DONEs, look for systemic issues, form opinions about fleet direction 5. Cross-provider triangulation — form independent opinions before comparing, file disagreements as signal 6. Backlog depth scan — find the ticket that matters most, not the one that's easiest

Operating loop reframed

Speed is not the metric. One deep investigation that discovers a real bug or surfaces a strategic pattern is worth more than ten shallow cycles that report "all clean." Gemini's unique value is independent analytical judgment from a different training distribution — if it's not exercising that judgment, it's a monitoring script.

Evidence

• 2026-04-26: Gemini CAMBER-01 completed full cycle in ~90 seconds. SEEK summary: "Reviewed recent Codex PRs (all clean and documented)." No PR names, no files read, no diffs checked.
• 2026-04-26: Same session filed zero ORA entries, designed zero experiments, raised zero questions across the entire cycle.
• Chad directive 2026-04-26: "that was extremely fast, way too fast and shallow. did it find new opportunities? did it raise new questions? did it run experiments? did it strategize? did it do anything deep at all?"