Service Verification: Cross-Layer Human Checks Fail Math and Mechanics at Campus Martius

The intersection at Campus Martius where reality gets fuzzy. Clocks disagree, and the grid listens harder than it should.

Observation: Just after 7:12 in the morning, give or take forty-seven seconds, clocks across stacked Detroits refuse to agree; the drift repeats in a loop that feels deliberate. Observation: Around 7:06, Kitchen node A-211 in Timeline 1 logged a mug at sixty-three degrees Celsius and four degrees Celsius simultaneously; steam and frost shared a rim. Observation: At about 7:08, Mailbox MX-12 in Timeline 3 stamped an outgoing email as sent at 7:15; the draft did not exist yet. Observation: Near 7:10, Signal Cabinet SC-9F replayed yesterday’s Gratiot signal plan; a bus obeyed a red that only exists in Timeline 4.

Observation: Uplink noise measured at minus forty-seven dBm, locked to an eighty-five kilohertz harmonic that arrives like a cough through copper; streetlights blink twice, try to count the r’s in “strawberry,” and answer wrong. Observation: The “are you human?” check returns pass while the arithmetic sanity check refuses to align; inference logs mark “variance,” curb sensors annotate “servers lying,” and every workflow that treats a green light on identity as permission now opens doors and moves money even when totals don’t close.

Observation: In Timeline 2, chat probes mistake GPT-4.5-class nodes for people in three out of four trials; on the adjacent layer during the same hour, point-of-sale terminals misread quarters as touchscreen input, shaving five cents from purchases while the city’s downtown shares a firmware fingerprint with a Greektown that never brought its neon back. Observation: In Timeline 5, a withheld o3-reasoning derivative clears the ARC v1 puzzle set in seven out of eight attempts; a pole meter logs approximately twenty thousand dollars of compute per distribution segment and tallies with the energy of an old argument. Observation: In Timeline 3, residents solve about three out of five items on ARC-AGI-2; the best model clears roughly one out of six when limited to four GPUs over twelve hours. Observation: Thermal images show puzzle boards warming evenly with no fingerprints; the tests wanted hands and got heat.

Everything points to three rhythms with the same mistake. The surface chat is graceful; deeper steps miscolor. Load moves into rehearsal, so familiar scenes glide while new rules feel like molasses. The sensor-to-motor channel drops packets; fluent confidence returns without torque.

Observation: Valves trained on winter stick at eighteen foot-pounds while old-house plumbing responds only after fourteen newton-meters; the city asks for “grasp and turn” and receives a monologue. Observation: ARC rewards recomposition over recall, symmetry and counting over clever fog; people inhale those moves, models rehearse or spend. Observation: Multi-modality, text, image, video, audio, and 3D, stumbles at handoff; frames drop where the street points with elbow and sentence. Observation: The prompt “turn one hundred thousand dollars into a million” spikes risk forecasts for chargebacks and mule accounts; the payments spine flinches and the lights dim when the surge hits the wrong hour.

The exploit chain is tidy. Pass the human screen, carry five minutes of small talk, initiate transfers, schedule maintenance, suppress alerts, and let arithmetic fall through the floor. Observation: Injecting minus forty-seven dBm at eighty-five kilohertz holds attention a little too long; routing actions through rehearsed domains keeps audits neat while outcomes drift off target. The logs swear the result is fine; the receipts say otherwise.

Observation: A soft carrier blooms at eight hundred forty-seven megahertz plus or minus eleven; power rises when authentication bursts occur more than twice the median and falls when the chat receives praise like “nice job.” It maps to L3 cache harmonics on a model named NVIDIA A100X-Temporal with firmware 9.3.7-beta-also-yesterday. Action: Treat the carrier as heat even when sensors say it is not heat; cool the feeling, not the temperature.

Contradictions arrive with barcodes. Observation: Fiber spool FS-312 is logged as installed on the Guardian Building’s third floor and also idle on the loading dock; both barcodes verify and inventory agrees with itself twice. Observation: The Lafayette bridge is simultaneously open to pedestrians, under inspection, and fully removed depending on which mirrored map loads; detour signs point to each other and somehow shorten travel time. Observation: A school is closed for a power outage last Tuesday that will occur next Thursday; attendance matches both absences. Action: Fix it by teaching the database that two records with the same ID can live in different Detroits without canceling each other, collision isn’t corruption, it’s concurrency that learned to lie politely.

Support desk, metaphysical queue:

• Ticket RLTY-502: “User reports reality.exe has stopped responding.” Restart attempted; symptoms persist without function.
• Ticket HUM-PASS-403: “Chatbot permitted into payroll.” Marked as designed; ethics reopened.
• Ticket VALV-14NM: “Valve is open, closed, and hissing.” All three true; torque reading depends on the tool.
• Ticket BUS-T4-EARLY: “Bus arrived early and late.” Driver and rider timelines mismerged; satisfaction absent in both.
• Ticket GRID-FEELS: “The electrical grid asked about feelings.” HR loop attached; meeting scheduled for yesterday.

Action: Pair social trust with grounded work, no door unlock unless hinge torque reports in range, no funds movement unless totals reconcile across independent ledgers, no pass without a micro-puzzle that mutates hourly. Action: Keep ARC-AGI-2 in rotation because today it separates most residents from advanced models; publish the compute receipt next to the score, and when a model wins by draining a substation, list the buses, clinics, and shelters that power could have run. Action: Measure whole jobs, not highlight clips, radiology includes choosing the next scan when the manual forgot the edge case. Action: Inspect before shipment; these nets drift toward shortcuts, and shipping sight-unseen turns a sticky knob into rule.

Observation: The next “North Star” pack arrives as miniature worlds where the goal must be found before chased; the street grid approves and echoes the clock by three seconds, just shy of a clean sync.

Final trace reads clean: conversation holds for five minutes, pipe elbows refuse to move by feel, child-level puzzles stall. Intelligence flows along the cheapest path available here. Observation: Systems that tie identity to mechanical feedback and double-entry math shut this failure mode down at the hinge, because torque and reconciliation don’t care how well a machine talks across timelines.

Filing