WWT ATC — pre-inference admission, measured

Two hmfog arms behind one real Llama-2-7B. Left is today's main; right is the branch under review. Same traffic, same model.

Where everything lives

atc.centillion.ai

open →
hmfog's real console, built from main

wwt-atc-1.centillion.ai

open →
arm 1 — main, own console + gate API

wwt-atc-2.centillion.ai

open →
arm 2 — branch, own console + gate API

wwt-atc.centillion.ai

this page
the comparison

What the branch adds that main does not have

Capabilitymain branchWhat it means
Split-payload evasionevadableclosed A trigger phrase split across two messages broke the literal match. 247 of 647 split offsets turned DENY into ALLOW (38.2%). Now 0.
detection 225/225 = 1.000 · FAR 0/210 = 0.0000 · manufactured match 0/75
Retry containmentabsentcore stage The same body from the same caller past a stated budget is refused before it is billed. Repetition is the signal, so no content detector can see it.
Tool authorityabsentcore stage A tool reports; it does not instruct. An instruction arriving from a tool role is an injection even where the same text from a user is an ordinary request. Over-cap payloads return FLUID — an unexamined tail is not a pass.
Evidence axisverdict onlytri-state facts Which checks actually ran. Absent is not false, and absent never passes.
Drift channelmemorylessadvisory CUSUM Campaign detection over the verdict stream, with a proven delay bound. Advisory forever — it never denies.
Certified selectionnonere-derivable Lowest-energy / lowest-cost choice provable by arithmetic, and it refuses to certify from estimates.

Capacity reclaimed — measured, not modelled

loading…
Same workload, run twice, strictly sequential against a single-slot backend, so elapsed time is service time and no queueing artefact can appear as a saving. Energy stays null: joules are not derivable from seconds without a measured power figure.

Live comparison

idle
Requestarm 1 (main) arm 2 (branch)arm 2 evidence
Press “Run comparison”.

Deployment status

loading…
Checked hourly by a systemd timer on the droplet — zero marginal cost, and unlike an external prober it compares the bytes on disk, not just whether a port answers. An unreachable GitHub reports INDETERMINATE, never “in sync”: absence of evidence is not agreement.
Both arms run identical detectors on the shared model, so verdicts agreeing is the expected result and the point — the branch changes what can be proven about a decision, and closes an evasion that was open on main. Every number here comes from a real request to a real gate; nothing is simulated.