Asking the fleet what it is doing…
monad-knowledge Wi-Fi sensing lab · FIIT STU
Campaign

Placement oracle, re-earned — the 57-floor coverage-vs-counting claim with attached runs and rendered figures

c-placement-oracle-powered

Archive snapshot, as of 5 h ago — the run corpus is rebuilt once a day, so this page is not a live reading. The fleet panel is the live one; it refreshes every 30 s.

Sessions

No sessions yet — this campaign has a brief but no execution.

Brief

Why this campaign exists

The 2026-07-15 integrity audit found that the prior "POWERED GO" placement-oracle session (01KXE2P8YXF93EY0Q898HDFER9) sealed a confident 57-floor + ray-traced verdict in 44 seconds with 0 attached runs, 0 figures, and 0 CPU/LLM budget spent — a write-up presented as a run. The one fetchable figure was byte-identical to the earlier honest n=6 pilot (01KXDYYK…), and the claimed 57-floor + stage-3 figures 404'd. The pilot (n=6) is sound; the powered claim was never actually computed.

This campaign re-earns the claim under the now-mechanical seal gate: sim_campaign_session_close refuses to seal resolved criteria with 0 runs/0 figures, and the runner now registers every rendered PNG. Same design as c-placement-oracle (56 real floors, 5 personas, coverage↔counting submodular reduction, 3 ray-traced verify floors) — but every headline number must trace to an attached run and a figure in S3.

What the supervisor should do

Run the full 56-floor Stage 0-2 corpus with runs attached to THIS session; compute the divergence CI over the attached corpus (not the pilot); render + upload every figure. Then the 3 verify floors for Stage 3. Honest framing throughout: this is a generative placement ranker — a hypothesis generator until a real AX210/Pi5 floor confirms the ranking (IP-106 / IP-112). No detection-rate claim.

Adversarial-review notes (2026-07-15): pilot_sigma: 0.20 is borrowed from a different campaign (c-coverage-meets-crowds, blind-spot-fraction variance) — the true placement-divergence σ is unknown, so re-estimate σ from the first real floor batch and confirm n=56 still powers the divergence CI. The counting objective in Stages 1–2 is a geometric proxy, and Stage-3 ray-traced I(C;Φ) is sparse (3 floors, discriminates only at high occupancy); a clean powered re-run therefore supports "coverage-optimal ≠ counting-proxy-optimal placement", NOT real counting performance. The "ranker / hypothesis-generator until a real AX210/Pi5 floor confirms the ranking" framing (above) must survive into the synthesis or the same over-claim returns in honest clothing.

Prior sessions: c-placement-oracle session 01KXDYYK… (honest n=6 pilot — keep as the pilot anchor), 01KXE2P8… (the phantom — do not cite its numbers).