Experiment · 2026-10-13

ORACLE-ST

Retrieval ceiling: does injecting the known fact chunk into the 4B cut 7 context (ORACLE-ADD extra slot; ORACLE-ONLY lead+fact chunk) on the 180 public self-tests raise the score vs the S runs (item's pack, today's build)?

Chapter 9 · Knowledge packsopen

From the record

What was tested: ORACLE-ST
Result: inject fact chunk (ADD) +0.095 (+0.070…+0.122); fact chunk only (ONLY) +0.115 (+0.088…+0.142)
Verdict: open: parked, running or awaiting a result
Detail: running (stage 0 PASS; stage 1 ORACLE-ADD generation 138/180 at 21:57, ORACLE-ONLY not started, 0 judge calls; no ORACLE-ST.md yet)
Set: 180 self-tests
Judge: Claude Opus, Claude Haiku (L1)
n: 180 (planned)
Seeds: 1 (s42, deterministic generation)
Pre-registered: yes, sha256-frozen

Linked facts · 1

Ceiling test: handing the model the exact fact-bearing chunk shows the headroom is retrieval, not reading.Ceiling test: handing the model the exact fact-bearing chunk shows the headroom is retrieval, not reading.inject fact chunk (ADD) +0.095 (+0.070…+0.122); fact chunk only (ONLY) +0.115 (+0.088…+0.142)Set180 pack self-testsStatusdevJudgeClaude Opus (judge.v1, both orders)n180Seeds1Date2026-10-13CaveatOracle (cheating) cell: an upper bound, not a deployable result.Sourceresearch record arms/oracle-st/ORACLE-ST.md:3research record reports/REPORT-PACKS-2026-10-13.md:19research record arms/oracle-st/ORACLE-ST.md:3,8Open the data file