# HP callback Java/Philadelphia physical collection — 2026-10-14

Both fixed callback generations are installed without another sudo registration.
Java generation `234c42c9f3e260e65ba022b3a9347c45a750ebf0b18f274ee82953ef48d71d05`
and Philadelphia generation
`277bac24e115972060e6d74bb70a17d6c6f71223a239ce2579abcd30376c5b37`
use the qualified split-placement products. Their complete clean source controller is
`c176e7512fd85f10743686dcb52a11e297a21d72`, with 867 selected files and source
inventory SHA-256 `0f151caa4b8a9a4cfb7ecde252a646a4d31d7d6c3c67fd958d7cc4190661446d`.
SDKs and the two fixed root helpers remain unchanged; payloads execute as the normal user.

HP completes 24 functional and 32 actual physical pairs under the required high-priority
leases. Each engine has four tiny and twelve full physical pairs. All body, full owner,
prepared/callback binding, original administration, per-thread kernel read-back,
namespace/endpoint/runtime, process-exit and physical-counter checks pass. There are
240,640 primary samples, 240,640 delays and 661,900 application frames per peer,
including functional qualification. The full physical collection uses the unchanged
1,000 warmup / 10,000 samples / three runs / 10,000 per second contract.

Java remains **S5 diag** on all four flows: every row misses achieved rate, and NOS also
misses percentile stability. Philadelphia holds **S5 pass** for MD and mixed; NOS and
tick remain **S5 diag** for achieved rate. No row misses the p99 delay gate. Four
original SDK-free physical cold receipts pass, independently recomputing every metric,
full owner/body observation and unchanged gate.

The new comparison reader checks both original generations, matching full profiles,
physical topology, total endpoint CPU budgets, actual callback tick spans, source/config
bindings and complete producer completion. Its 18 controls refuse partial/borrowed
profiles, failed grades and incompatible clocks, and prove that worst is the maximum
of all run extrema. Actual HP replay is **blocked on every flow**, because Java is
not admitted. All metric/ratio objects are empty. This result explicitly requires one
rig and cannot be advertised as the two-rig campaign comparison. SCAN remains queued.

The [immutable seal](evidence/java-callback-hp-physical-20261014/retention-seal.json)
binds 2,199 original inputs / 371,982,128 bytes; SHA-256
`1721434176ec12bfbcf73f5e7c3f112fc5b6d09f8b1fdf11ff74ba124cb8323c`.
It retains the exact clean source, all HP originals, physical counters/polling, launcher
and CPU proofs, cold receipts and actual blocked one-rig comparison with its readers.
No source, sample, quantile, extremum or admission threshold is changed.