# Philadelphia field guard labels — 2026-10-17 The public Philadelphia consumer previously built `missing tag ` for every successful field lookup and `tag ` for every successful equality check. These labels are now constructed only on the corresponding failure. The actual public `FIXValue`, full field lookup/equality, exception type and message, complete numeric extraction and all clock/capture boundaries stay the same. No SDK source or binary changes. The two helpers have package visibility so the existing consumer test can exercise actual SDK messages. New checks require the original missing-tag and mismatched-value messages and the exact SDK field object on success. The unchanged capture/numeric suite plus these checks passes **5,535 checks, zero skips** in a bounded local compile of the seven current consumer classes and its test, using the unchanged pinned Philadelphia 2.0.0 core JAR. The Java compiler has a 128 MiB heap limit and the test a 64 MiB limit. An initial local compile tried an older cached adapter as its dependency and refused compilation because that adapter predates the current bounded capture fields. No test or session ran in that attempt. The passing compile uses every current consumer source, with only the original core SDK JAR as its dependency. Both attempts remain under the session's external build root (`philadelphia-guard-labels-local-20261017-v1` and `-v2`). Fresh exported builds, actual session pairs, original capture/copy-count audits and cold refusal controls remain required on both hosts. Existing queued source capsules and historical results are unchanged. This source change and local semantic test alone supply no rate, allocation, Stage 5 or comparison claim. ## Actual standalone exported qualification on SCAN The 933-file benchmark manifest export at `4f19817c344207f7e3280b674cfa17db2a32fc68` has zero closure, link or scrub findings. The independent frozen-source reader verifies both its complete file/mode inventory and actual archive members. Inventory SHA-256: `9b8e93bf201be8557e4d7d36e2565c0d8697df8c0dde3a01149da016e99567a5`. Under an outer SCAN `isol` lease, the actual stripped tree passes: - **2,233 portable controls plus 58 subtests**, zero skips/failures; two SDK-source-only guards are explicitly deselected. - A fresh build against the unchanged pinned original Philadelphia core: **5,535 Java checks, zero skips**. - **Twelve actual four-flow pairs**, covering warmup five, zero and repeated five, with complete samples, measured owner folds and original preflight body/correlation checks. - **113 cold refusal controls** against copies of those original results. - **24 endpoint copy-count checks** requiring preflight-only application bodies and complete later metadata/admin observations. All source bytes and modes remain unchanged. Actual session source SHA-256: `9104a214e8ab715619e059ef7a2c2adaade793bd6eed9c0f0e5569f85cbc0f72`; test SHA-256: `8105c0763b47ba832753a342396bb55a2698ea5991a3a5e1070d665ac8da9e6d`. The compiled adapter JAR on SCAN has SHA-256 `d9976362897d8c8c44bf79ecdf386dabfe3e02691eac9ed57fb6db5c0c8a7138`. The exact original qualification receipt is `~/build/codex-fullbench/libhft/philadelphia-guard-labels-qualification-20261017-v1/completed.json`, SHA-256 `411c23d663088f5941264963e4488e04564901797f675ba481461c04cb1bdf37`. No application capture or raw series enters Git. HP's identical qualification is queued with high priority. A dependency-gated continuation prepares a separately compiled callback consumer using this same 933-file harness and the already qualified original public SDK JAR. That SDK keeps its separate `2d94a107c` source pin and actual per-host container hash; the new harness pin never rewrites its SDK identity. Fresh session/cold checks must pass before either new physical producer is prepared. Both preparations precede quiet timing leases; only completed original reports can reach the explicit callback comparison and archive-member reader. HP jobs retain high priority. The independent reader revision is frozen separately at `1b03dca7a1a1557439eeb12a0addc2d5f8278568`. These continuations have not supplied a timing result. Existing `2d94a107c` physical jobs and good compact cells remain intact; a failed new flow cannot replace a passing cell. The fresh complete standalone harness gate including the new reply-owner controls is separately queued from its own immutable 933-file export. Its export has no closure/link/scrub findings; its larger test suite has no result yet. ## Continuation after the actual callback timing result SCAN subsequently qualifies that separately compiled callback consumer (56 checks, twelve pairs and 107 cold controls) and prepares both new immutable physical generations. The older, first callback physical run then completes with rate failures on all four flows. Only the two owned continuation parent processes are stopped before they repeat that unchanged baseline's timing; SCAN's already queued Philadelphia child continues, as do the original HP physical jobs and every other session's work. HP's new Philadelphia functional qualification also remains queued. No physical target is interrupted. The [new Java numeric source](RECORD_2026-10-17_java_ascii_numeric.md) must qualify separately before further paired baseline timing. The failed original callback result and [exact independent blocked decision](RECORD_2026-10-17_java_callback_reply_physical.md) remain intact. The separately queued complete standalone harness gate has now passed 2,250 controls plus 58 subtests, with two explicit SDK deselections and no skips. The earlier pending statements above remain historical checkpoints. ## First guard-label physical collection on SCAN The already queued physical child completes all four tiny and twelve report pairs under its quiet lease. Both independent exact cold replays pass, with 120,080 primary and 120,080 scheduler-delay observations overall and all 330,300 report application frames per peer accounted for. The actual pre-warmup capture and full owner/admin/physical proofs remain valid. | Flow | Achieved cycles/s across three report runs | Actual report result | |---|---:|---| | NOS/ER | 9,870.775–9,988.002 | S5 pass | | MD | 9,976.339–9,983.963 | S5 diag: full percentile stability | | Mixed | 9,692.723–9,735.094 | S5 diag: rate, full stability and scheduler delay | | Tick to trade | 9,979.632–9,979.871 | S5 pass | Only the passing NOS/tick compact files replace their earlier keys. There are zero added cells and 655 cells overall; every other cell stays byte identical, including the earlier passing MD cell. Raw series and failed results remain on SCAN under `philadelphia-guard-labels-physical-20261017-v1`; small exact receipts and validated passing compact files are collected outside the repository. No failed flow or separately prepared callback generation receives promotion.