add deterministic pilot benchmarks

docs/OPERATIONS.md

@@ -251,3 +251,19 @@ There is currently no standalone migration or backup CLI. `schema.sql` is applie
 ### Explicit limitations
 
 This repository does not provision Virtualmin/TLS/DNS, provide a dependency-aware readiness service beyond the API's SQLite check, durable migration runner, PITR, HA database, durable queue/worker leases, production egress proxy, or compliance-grade retention service. SQLite and the in-process worker are pilot-only. The release has no outbound provider/send path and must remain outbound-disabled by default. See `docs/RELEASE_CHECKLIST.md` for the short go/no-go gate.
+
+## Phase 16 pilot benchmark operations
+
+Run the Phase 16 benchmark only against the versioned, labeled synthetic fixture manifest described in `docs/BENCHMARKS.md`. Confirm that fixtures contain no live personal data, secrets, uncontrolled network responses, or production records. Keep expected labels separate from generated outputs, and record the manifest/hash, label-policy version, commit, clean/dirty state, algorithm and rule-set versions, runtime/dependency/image digests, non-secret configuration fingerprint, seed, locale/timezone, rounding/tie-breaking, host profile, command, and timestamps.
+
+For each deterministic classifier/scorer, run every fixture at least twice in fresh processes. Compare class, score, band, eligibility, uncertainty reasons, explanation factors, and version metadata; any unexplained difference is a reproducibility failure. Report confusion matrices, support, precision and recall with zero-denominator `N/A`, plus unknown/blocked/partial/error and suppression outcomes separately. Do not tune thresholds, remove hard fixtures, or convert `N/A` into a pass after seeing results without publishing a new methodology version.
+
+Measure latency after a documented warm-up, using monotonic timestamps and a fixed local, network-free workload unless separate approval exists. Separate cold, warm, bypass, and not-applicable cache states. Record every sample, including errors/timeouts, and report count, minimum, median, p95, p99, maximum, timeout/error rate, cache hits/misses, freshness/expiry, and invalidation behavior. State whether startup, serialization, database, and network time are included. Never present a cache hit as a fresh observation or infer production capacity/SLOs from pilot timings.
+
+Before sign-off, evaluate the synthetic-only gates: 100% deterministic output equality; precision `>=0.90` and recall `>=0.85` only when positive support is at least 20; 100% safety preservation for suppressed, blocked, malformed, and explicitly unknown fixtures; 100% cross-tenant denial/non-disclosure probes; and complete latency sample/cache labels. Conduct blinded, seeded human review stratified by predicted class, uncertainty/suppression state, and FP/FN disagreement, reviewing at least 20 fixtures or 10% (whichever is larger, subject to the declared cap). Use pseudonymous reviewer IDs, reason codes, independent review for disagreements, and documented adjudication. A passed pilot gate is not production evidence.
+
+Store the report and raw machine-readable samples with the pilot artifacts, including failed/incomplete runs and limitations. Do not place sensitive fixture values in routine logs. Phase 16 does not enable a worker, live source, or outreach path.
+
+### Unresolved publication and deployment prerequisites
+
+Do not report local benchmark or Compose success as deployment. Remote push remains blocked until repository authentication and intended remote/branch permission are supplied. Production deployment remains blocked until the Phase 15 operator prerequisites are verified: reviewed remote revision, Docker/Compose host access, protected deployment directory, secret injection, DNS/TLS/Virtualmin, encrypted off-host backup and restore evidence, monitoring, rollback owner, and operational approval. Record the blocker in the release record and keep the current outbound-disabled configuration.