CJ-05 data dictionary

Identity and scope
- group: matrix / all_scope_control / budget_control / speed_control.
- case: engineering-fixture identifier; exact dimensions, time base, preparation
  commands and source SHA-256 are in CORPUS.json.
- profile: economy_5s, economy_2s, resilient, image; not_applicable for unmarked input.
- scope: current (new-issuer scheme) or explicit all control (legacy included).
- temperature: process_model_cold = first job in fresh SDK/container;
  resident_warm = subsequent job. Dependencies provisioned, OS caches not evicted.
- id: exact raw record and per-operation timeline filename.

Time and CPU
- wall_s: host request-to-response time including small IPC/JSON overhead.
- sdk_call_s: measured SDK method only; first-job initialization separately sdk_init_s.
- worker_operation_s: complete worker-side job including initialization/result prep.
- total_cgroup_cpu_s: canonical cpu.stat usage delta for every process/thread in
  this container, including live Node, native libraries, reaped children.
- python_process_cpu_s: Python process SELF, including native PyAV/C2PA/Torch work;
  not exclusively interpreted Python bytecode.
- live_node_cpu_s: diagnostic process CPU delta using PID + starttime identity.
- reaped_child_cpu_s: diagnostic exited-child CPU. Never add shutdown lifetime
  reaped Node counters to per-job totals a second time.
- cpu_accounting_difference_s: cgroup total minus SELF + reaped + live child deltas.
  Minor clock-tick/serialization differences are expected; raw counters retained.
- average_cpu_cores: total_cgroup_cpu_s / wall_s, not a percentage of the whole VM.
- cold_inclusive_*: spawn-to-first-response accounting including container startup
  and controller setup intervals; not install/download/build latency.
- receipt_encoder_inference_*: SDK receipt's narrow video-embedding counters;
  not a complete CPU or wall breakdown of decode/encode/gate/signing work.

Memory (MiB = 2^20 bytes)
- memory_peak_mib: canonical cgroup kernel memory.peak reset/read on the SAME FD
  for this request. Charged memory includes anonymous/native pages and charged
  page cache; it is not RSS, the 4096 MiB cap or global physical RAM attributable
  to every shared page. Shared pages charged elsewhere need not appear here.
- memory_start/end_mib: cgroup memory.current at request boundaries.
- sampled_anon/file_charge_peak_mib: independent sampled maxima from memory.stat;
  maxima need not occur at the same instant and must not simply be added.
- sampled_cgroup_peak_mib: maximum memory.current observation, normally 100 ms.
- sampled_rss_sum_peak_mib: sampled sum of process RSS; shared mappings can count
  multiple times. Role RSS maxima are also non-coincident diagnostics.
- sampling_enabled=false: no periodic sampling; endpoint snapshots and kernel
  peak/CPU deltas remain. Sampled fields then mean endpoints only, not 100 ms data.

Host and I/O
- observer_cpu_s: host observer's own process CPU, outside measured container.
  It competes for host resources and is checked by a no-sampler control.
- host_steal_percent, host_busy_percent: aggregate host tick deltas for the request.
- throttled_s/periods: kernel cgroup CPU quota counters, not wall delay estimates.
- io_read/write_bytes: summed cgroup io.stat device deltas. Raw per-device values
  remain in timelines. They are not user-level payload bytes or guaranteed unique
  physical I/O across stacked devices; page-cache hits may not cause block I/O.

Verdicts
- accepted_export: SDK accepted the full export/read-back gate.
- correct_id: returned ID equals the expected export's ID (assessed AFTER verify).
- completed_negative: watermark.status=not_found on a clean fixture.
- incomplete: watermark.status=not_checked; never evidence of absence.
- wrong_id / unexpected_id / operational_error: retained adverse outcomes.
- API completion (n_ok) is distinct from semantic outcome.
- Metadata-loss positive proves ID/record recovery, not the authenticity of a
  modified copy. Development signature validity is distinct from issuer trust.

Statistics
- Warm median/min/max over three sequential observations of one source per cell.
  IDs/files differ for exports and positives; negative repeats use the same clean
  source. They are timing repeats, not independent content-population samples.
- Cold n=1 per cell. No p95, SLA, population accuracy/FPR or legal cadence claim.
- Mixed/idle controls and earlier client pilot are separate, not pooled into
  primary medians. For runtime CLI vs PyAV library versions see RUNTIME.json.
- Exception: the separate 60-second budget control has one cold and one warm
  observation, not three warm repeats; its summary n is explicit. It is a limit
  probe and does not estimate warm timing spread.
