KUBERNETES / MEMORY PRESSURE / COST CONTROL

Spot rising memory pressure. Stop paying for blind headroom.

A bounded, read-only advisory guard for memory pressure decisions. No agent-side allocations on the hot path and no application rewrite in the evaluation phase.

Synthetic reference results and a native Mac2 scope-limited benchmark only. The customer selects the industrial baseline and acceptance gate. No production SLA, crash immunity or independently audited saving is claimed.

Try synthetic Proof Lab Fund a bounded assessment

01 / THE COSTLY BOTTLENECK

Audit first. Change only what survives a controlled comparison.

The problem

OOMKilled incidents and over-provisioned memory consume engineering time and budget. A narrow, measurable admission and capacity signal can improve decisions without invasive heap instrumentation.

Our boundary

Public synthetic reference cases show stable envelopes, pressure warnings, budget breaches and invalid limits. This preview does not block malloc, prevent OOM or measure production savings.

Measured acceptance

Customer-selected p50/p99 tail, exact outputs, adverse cases, CPU/memory baseline, rollback and a written stop/go decision.

02 / INTERACTIVE PUBLIC-REFERENCE PREVIEW

Run, verify, tamper, export.

This browser-side receipt is generated from disclosed synthetic scenarios. The private C++20 execution path never enters the browser.

Synthetic scenario

Choose a fixture and create a receipt.

Public scenarios—

Receipt SHA-256Not generated

Evidence envelope

No receipt created.

SHA-256 verifies the integrity of the locally constructed receipt. It does not independently validate the proprietary engine or prove external performance.

Download public technical brief (PDF)

03 / SCOPED NATIVE EVIDENCE

Benchmarks with boundaries, not slogans.

Private engine

C++20

Mac2 native candidate. The owner retains source and licensing rights.

Median local measurement

54.99 ns

ns per advisory decision. Single host, R1; five repetitions. Not production latency.

Differential / adverse checks

PASS scoped

Kernel eBPF or runtime enforcement needs separate privileged Linux qualification; advisory engine only.

Inspectable synthetic cases

#	Scenario	Expected local output	Evidence category

Industrial evidence · measured scope

30/30

Jobs completed with reservation

0 OOM

Per guarded policy (historical)

120/120

Independent synthetic references

Measurement boundary: Linux cgroup v2: 18 observational trials (8 kernel-confirmed OOM), median advisory lead ~446 ms. Separate admission test: no guard 9/30 (21 OOM), passive PSI 9/30 (21 OOM), reservation-only 30/30 (0 OOM), reservation plus native advisory 30/30 (0 OOM). No additional OOM prevention benefit of native advisory versus reservation alone demonstrated. Independent Linux 6.18.57 reference suite: 120/120 tests (steady, burst, fault, recovery; 30 each), not a proprietary-engine benchmark.

Evidence quality and eight industrial gates (G0–G7)

Scope-qualified, not industrially certified. Incremental benefit must be judged against the best correct optimized baseline using false blocks, overhead, failure cases, p99 and cost per completed work unit.

G0 · EvidenceHash-linked independent reference logs retained privately; historical evidence is bounded.

G1 · CorrectnessHistorical scoped tests plus 120/120 independent reference fixtures; native engine has a separate test oracle.

G2 · Incremental valueNot demonstrated against an optimized, same-workload production competitor.

G3 · Tail and scaleExisting local timings are scoped; no wire-to-wire or production p99/p99.9 SLA.

G4 · Fault / recoverySynthetic adverse and recovery fixtures only; native chaos testing remains separate.

G5 · Competitive baselineReservation-only, active systemd-oomd, memory.high and VPA where applicable; active optimized oomd or VPA comparison not yet performed.

G6 · Security / IPNo private code, thresholds, credentials, customer traces or raw private logs in the public view.

G7 · Economic acceptanceNo verified customer savings. An isolated substitute is not a client workload.

Historical per-run public summaries: OOM admission outcomes · warning comparison.

Original raw traces remain private; published hashes alone cannot independently validate unseen logs.

PUBLIC EVIDENCE · PRODUCT BOUNDARIES · PRE-CONTRACTUAL TERMS

Check the evidence and price before making contact

System boundaries and OOM results

A read-only cgroup v2 memory-pressure advisory. Actual OOM prevention in the controlled experiment required a separate pre-admission memory reservation gate. Native advisory by itself is not an OOM actuator.

18 observed runs: 5 stable, 5 recoverable spikes and 8 kernel-confirmed OOM; first warning median lead approximately 446 ms.

Separate admission experiment: No guard 9/30 jobs complete with 21 OOM kills; passive PSI 9/30 and 21 OOM; reservation-only 30/30 and 0 OOM; reservation + native advisory 30/30 and 0 OOM. No incremental native prevention advantage has been demonstrated over reservation-only.

Active systemd-oomd, memory.high and VPA comparisons remain untested.

Admission per-run summary · 18 observational summaries

Downloadable synthetic reference evidence

720/720 checks across six independent reference models, four scenarios and 30 repeats per scenario. Not proprietary-engine, customer or production results. Reference timing is not native-engine latency.

Raw 720-row JSONL Full numerical summary Python checksum and chain verifier Exact synthetic reference harness (Python)Protocol and limitations

SHA-256 of original raw reference file:

19b39154b7452809260bb3fb248971f2b871b6a38be9fc3c669ba4a7872f308c

These downloadable bytes reproduce the 720-row reference log. The raw kernel traces behind the separate OOM experiment remain private; their already published sanitized summaries are linked above.

€3,900 · Exact diagnostic perimeter

Fixed fee: €3,900. One approved de-identified trace capped at 2 GiB or 1,000,000 records, one isolated non-production environment, up to four scenarios and 30 paired repetitions per configuration where supported. Comparators: reservation-only, passive PSI, reservation + advisory and one client-approved optimized incumbent; unsupported active oomd / memory.high / VPA policies are marked NOT TESTED.

Included: one 45-minute kickoff, one 45-minute results call, one clarification email thread within five business days; 8–12-page dossier, aggregate results, SHA-256 manifest and written GO / NO-GO / INCONCLUSIVE.

Payment: 50% on signed acceptance, 50% on delivery; taxes/withholding per quote. Target 10 business days after complete approved input. Additional meetings, environments, lab costs, reruns or features require separately priced written approval.

Excluded: source or application changes, production access/deployment, Kubernetes changes, architecture consultancy, on-call, new custom engine features, guarantees of savings or production SLA. Client retains its data; provider retains proprietary IP. A documented NO-GO is a completed deliverable. Public terms become binding only through a mutually signed quote, subject to applicable legal and data-processing requirements.

Printable public one-page scope Download signed-scope preview (PDF) · Contact about this scoped evaluation

04 / FUNDED BUYER VALIDATION

Approve the scope. Measure the savings. Then discuss the license.

Fixed-scope diagnostic

€3,900

One workload, agreed memory pressure traces, paired baseline and written failure/rollback criteria. Production changes remain opt-in.

Optional Proof of Value

€9,000–18,000

Customer data stays under a restricted agreement. License / gain-share pricing is agreed separately after measured value and acceptance.

Indicative commercial quotes for negotiation, not claims of objectively appraised market value or guaranteed savings.

05 / PAID TECHNICAL EVALUATION

Request a funded confidential technical evaluation.

Draft output

Fallback draft will appear here if email delivery fails.

INCREMENTAL CONTROL · MEASURABLE VALUE

An early memory-pressure signal on cgroup v2, without application rewrites.

Control boundary

Observe Linux cgroup memory accounting and issue a bounded capacity advisory. This is a read-only control signal that may complement PSI, oomd and VPA; it does not suppress arbitrary allocations.

Why it adds value

The economic value is safe headroom reduction or fewer avoidable incidents after counting interventions and unnecessary pressure warnings—not a guaranteed percentage of RAM savings.

What is proven

Open the interactive synthetic Proof Lab, reproduce the inputs, inspect SHA-256 receipts and test tampering. A controlled Linux admission test now also demonstrates completed-work outcomes without exposing the private engine. Customer production qualification remains separate.

Measurement method, baselines and limitations

18 aligned Linux cgroup-v2 trials: five stable/no alerts, five recoverable bursts/warnings, eight kernel-confirmed OOM runs/warnings. First advisory lead: median 446 ms (range 444–532). All eight forced-OOM trials reached HOLD, with a median HOLD lead of 446 ms (range 179–532); zero HOLD in five stable and five recoverable-burst trials. Each recoverable burst did cause a transient WATCH, so false-warning tradeoffs remain. A PSI cumulative increase appeared before exit in two of eight runs, and PSI some.avg10 never reached 1%. This compares signals on the same sample timeline; active oomd and Kubernetes VPA were not run. A separate controlled boundary-admission experiment did actuate a queue; the outcome below is independent of the advisory-only timing test.

Actuated prevention, bounded scope: in three comparable Linux cgroup-v2 batches of ten unchanged workers, no admission guard completed 9/30 jobs with 21 OOM kills; passive PSI-only also completed 9/30 with 21 OOM kills. A fixed memory-reservation gate completed 30/30 with zero OOM; Memory Shield plus that reservation also completed 30/30 with zero OOM. The admission gate trades batch latency (about 4.9 s versus 1.3 s for incomplete ungated batches) for completion. No incremental OOM-prevention advantage beyond a simple reservation gate has yet been shown. This controlled result is not a production guarantee, and neither active systemd-oomd nor VPA was evaluated. Method and per-run results.

Per-run comparison data (privacy-safe).

Loris Longarini · Independent Technology Researcher · Deterministic Systems Architect · LinkedIn and technical background. Private IP retained; funded diagnostic, financed Proof of Value and licence are separate contractual stages.