ServerBudget

What Counts as Evidence When a Delivered Cheap VPS Does Not Match Its Plan Page

By Marcus Reid · Published Nov 11, 2026 · Last updated Nov 11, 2026

Short answer: Keep four things, and keep them paired. First, the untouched readouts of the machine's own interfaces for memory, block-device type, processor identity and the link setting [S1][S5][S6][S7]. Second, the plan wording those readouts are held against, saved before anyone can revise it [S12][S13]. Third, a side-by-side note naming the field behind each value and the official definition that gives the field its meaning [S2][S3][S7]. Fourth, a dated, two-sided trace: a ticket opened through the channel the contract names, with the reply kept beside it [S14]. A record becomes checkable when somebody else can repeat the same reads, obtain the same plan text and see the order in which the two sides were compared. A lone screenshot of one command is not that record [S14]. This page reports no self-run result, names no tool for a reader to run, and prints no outcome figure of any kind.

What the machine will tell you, and what each field means

Begin with memory, because it is the field most often read as if it had one meaning. The kernel's own table defines MemTotal as total usable RAM, physical memory minus reserved bits and the kernel binary code; MemAvailable is described there as an estimate for starting new applications without swapping [S1][S2]. The free utility prints an available column built from those counters [S9]. None of those three is a statement about an allocation ceiling. The ceiling, where one exists, is a different object: the unified cgroup interface exposes memory.max as a read-write single-value file described as the hard limit on memory usage [S3], and the cgroup mechanism itself is a way of binding processes to limits through a pseudo-filesystem [S4].

Put those definitions next to each other and a narrow mistake becomes visible. Comparing a plan's advertised memory against MemTotal compares a commercial allocation against a kernel statement about physical memory, while comparing it against the cgroup limit compares it against an enforcement boundary [S1][S2][S3]. Those are different questions with different owners. I am classifying this as a reading of documents rather than a discovery about any delivered box. It rests on a short chain: two separate manuals define two separate quantities, both are visible from inside a delivered system, and a plan page names neither [S1][S3][S12]. So the first line of any record has to say which quantity was read, not just what the readout said.

Block-device type has a similar split. The block-device utility lists ROTA among the columns of its topology output and documents filtering devices by their TRAN value [S6]. The kernel's net and block interfaces are exported through sysfs, a pseudo-filesystem whose purpose is to expose kernel data structures [S11]. What follows from the column names is limited: a rotational flag answers whether a device reports itself as spinning, while a transport column answers how the device is attached [S6][S11]. Neither one, by itself, resolves the finer wording a catalogue may use for solid-state media [S12][S16]. I am labelling that boundary as an inference: it comes from the column names and the documented filter behaviour, not from inspecting what any provider installs.

Processor identity is the third field, and the official documentation is unusually candid about it. The CPU utility gathers its information from sysfs, the cpuinfo file and architecture-specific libraries [S5]. Its own known-issues section says that under one hypervisor the kernel sometimes reports wrong data, and that on virtual hardware the reported core counts can be wrong [S5]. This matters for evidence rather than for shopping. A model string is a value the guest was given to display, and the same manual that documents the read path also documents a class of cases where that path misreports [S5][S11]. Writing down the string is therefore necessary and insufficient: the record also needs to say where the string came from and that the source itself carries a caveat.

The fourth field is the link setting, and here the wording matters most. The kernel's interface attribute is defined as the interface's current speed value, expressed as a link rate, with a note that it is valid only for interfaces implementing a particular driver query method [S7]. The corresponding utility treats speed and duplex as device settings that can be displayed and set [S8]. A link setting is a negotiated property of the interface, not a record of how much traffic moved. On a virtual interface the attribute may simply be absent, because the official note limits its validity to interfaces that implement that method [S7]. Absence of a field is not evidence of a broken promise. I am treating that as an inference, with the chain written out: the documentation attaches a precondition to the attribute, a guest interface need not satisfy it, and no readout exists when the attribute is missing.

One more property of these reads is worth writing into the record before anything else. Sysfs is an export of current kernel data structures, and the project's rules for that filesystem govern how attributes are exposed [S10][S11]. A read is a snapshot of a state, not a durable object. Repeat it later and the state may differ. So reproducibility here cannot mean that every later read returns the same characters; it has to mean that the path, the moment and the comparison are documented well enough for a second person to retrace them.

How to keep it so somebody else can repeat it

Save the readout as it came out. Do not tidy it into a summary, because a summary silently discards the field names that carry the definitions [S1][S2][S7]. Alongside it, write which interface produced each field and which official document defines that field. This is the part that turns a number into a claim that can be audited: without the definition, a reader cannot tell whether a memory line came from a physical-memory counter or from an enforcement limit [S1][S3].

Save the other side at the same time. A plan page is editable, and the wording that matters is the wording that was shown when the order was placed [S12][S13]. The comparison sites in this set list parts and capacities in their editorial tables, and none of them describes a way to check a delivered machine [S12][S13][S16][S17]. That absence is the reason the plan text has to be preserved by the buyer rather than recovered later from a roundup.

The contract then sets a higher bar than a personal note can clear. Its terms rule out compensation for misinformation when the only basis is whatever the provider's website, system and lookup tools returned [S14]. Read literally, a private file of readouts is not a claim the agreement recognises; it is only the raw material for one. The same document names the HelpDesk ticket as the sole channel for technical support, requires the appropriate category, and declines any duty to answer a request filed under the wrong one [S14]. It also asks the customer to supply the information and access reasonably required for support [S14]. A two-sided trace, kept with its reply, is what bridges the gap between the two.

Two further clauses shape the filing. Full account access, account management and refunds all depend on the customer first completing a verification procedure [S14]. Notices under the agreement are deemed given when delivered according to the account or domain contact information on file [S15]. Both are procedural, and both belong in the record before any disagreement starts, because they decide where a written exchange can happen and whether it will be treated as delivered [S14][S15].

Keep the evidence outside the machine as well. The acceptable-use policy disclaims any guarantee as to the integrity of the backup copies it describes, and the terms say that everything stored on the machine is erased when the agreement ends [S14][S15]. A record that exists only on the disk being questioned has a single point of failure. I am marking that as an inference, and the chain is short: the documents decline to warrant the copies, they say the agreement's end removes stored content, and local-only retention therefore is not independent retention [S14][S15]. No claim is made here about how often any copy is made.

What the record cannot do is settle the disagreement by itself. It is written by one party about a machine controlled by the other. Calling it checkable means the reads can be repeated, the plan text can be produced and the sequence is visible; it does not mean the file is neutral testimony. That distinction is mine, and it follows from the contract's own exclusion of claims resting only on system results [S14].

The order to use when the two sides do not line up

Start by separating a definitional mismatch from a shortfall. Four of the fields above have documented reasons to differ without anything being wrong: physical-memory counters and enforcement limits are different objects [S1][S3]; a rotational flag does not resolve the catalogue's media wording [S6][S12]; a link attribute carries a stated validity precondition [S7]; and the CPU utility's own known-issues section warns about virtual hardware [S5]. Writing "the plan is wrong" before checking those four is a category error, not a complaint.

Then freeze both sides in the form described above: untouched readouts, the saved plan wording, a field-by-field note with definitions, and the timing [S1][S2][S12][S13]. Do this before contacting anyone. Sysfs exports current state, so a later read is a different observation rather than a repeat of the first one [S10][S11].

Only then open the written exchange, through the channel and category the terms name, with the field note attached rather than a conclusion [S14]. Supply the information and access the terms ask for [S14]. Keep the ticket identifier and the reply. If the account has not passed the verification procedure, the same terms say that account management and refunds are not available yet, so that step comes earlier than the substance of the complaint [S14]. Notices are deemed delivered to the contact details on file, which is worth checking before assuming silence [S15].

Read the promise itself once more before escalating. The plan page describes dedicated resources and unmetered bandwidth, and separately states that the port setting is adjusted once a published traffic threshold is crossed [S12]. The acceptable-use policy carries its own version of that adjustment and disclaims warranties, while the terms state that the services are provided as is and do not warrant that they will meet a customer's requirements [S14][S15]. A complaint framed as a fixed specification may therefore be answered with a policy clause rather than a repair. That is a reading of three documents together, not a statement about how any provider behaves [S12][S14][S15].

Hold the boundary. This sequence is about what can be written down and in what order; it is not legal advice, and it does not predict an outcome. Nothing is measured in it, nothing is suggested for a reader to run, and I offer no figure for how frequently a delivered machine differs from its plan.

What the documents share, and the three fault lines

The agreement is about provenance. Every official interface read here is an export of current kernel state rather than a certified statement of what was sold [S2][S10][S11]. Every accessible contract in this set places risk, backup responsibility and the burden of using the right channel on the customer [S14][S15]. And the editorial pages describe the parts a plan contains without describing any procedure for checking them after delivery [S12][S13][S16][S17]. Three different kinds of document, all stopping at the same place.

The first fault line runs through memory. The kernel documentation defines a physical-memory counter, the cgroup documentation defines an enforcement limit, and a catalogue prints an allocation without saying which of those a buyer should expect to see [S1][S3][S12]. A record that does not name its field sits on the wrong side of that line.

The second runs through the port. The kernel documents a link attribute with a validity precondition [S7], the utility treats the same setting as configurable [S8], and the vendor pages make the setting contingent on a traffic policy [S12][S15]. One word, three different regimes: a driver property, a configurable setting, and a policy-dependent allowance. None of the three says how much traffic actually moved.

The third is the sharpest. The official CPU documentation admits that virtual hardware can be misreported by the very interfaces it describes [S5], while the commercial documents disclaim warranties about whether the service meets a customer's requirements [S14][S15]. Read together, the party supplying the readout warns that it may be inaccurate and the party supplying the plan declines to guarantee the outcome. I am calling that juxtaposition a construction of my own, made from documents that never address each other, and I am not attributing intent to either side [S5][S14][S15]. It does explain why a two-sided record matters more here than a confident single reading.

A fourth, quieter divergence concerns what counts as a claim. The contract excludes claims resting solely on website, system and lookup results and routes support through a ticket with the right category [S14]. The editorial pages offer no claim path at all [S16][S17]. The buyer's note and the provider's reply are the only pieces that both sides can see, and neither is written in the shopping comparison.

Once you have checked evidence for a delivered-spec mismatch, the public VPS uptime records page checks an availability claim against published incident history.

While this page deals with evidence for a delivered-spec mismatch, the what a cheap VPS does not include page lists the add-ons that turn a cheap headline price into a larger bill.

A record sheet, filled in before anything is sent

What this sheet cannot supply is a verdict. It cannot establish that a readout is accurate, because the official documentation itself leaves room for virtual hardware to be misreported [S5]. It cannot make a private file binding, because the terms exclude claims resting only on system results [S14]. It cannot recover a plan page that was edited before it was saved [S12][S13]. Those are limits of the method, not missing steps in it.

Limits: pages that would not open, and claims left unmade

Several documents that would have been useful did not open today. The second provider's plan page returned HTTP 403 [U1]. Its service-level agreement returned HTTP 403 [U2], as did its terms of service [U3]. A manual page for a hardware-reporting utility returned HTTP 404 [U4]. The dedicated attribute documentation for block-device queues returned HTTP 404 on every path tried [U5], so this page relies on the block-device utility's own column documentation instead [S6]. A knowledgebase index on the budget provider returned HTTP 404 [U6]. I quote nothing from them, I draw no policy conclusion from them, and no earlier note about them stands in for what they might have said.

I read the comparison pages only to record which parts of a plan they describe. Nothing of their sentence structure appears above. The per-dollar worksheet on this site reports no self-run result of its own; this page does not revisit its subject, contradict its position, or supply the figure it says is missing. The two sit on different questions: what a published price buys, and what can be written down about what arrived.

The narrow conclusion

A delivered machine makes its own state readable, but readable is not the same as agreed. The useful artefact is not a value; it is a pairing — a plan wording held next to a named field, with the definition of that field attached, with a moment attached, and with a written exchange that both sides can see [S1][S2][S12][S14]. Reproducibility here is a property of the path and the sequence, not of the characters in a single readout.

That is a smaller answer than "prove the plan was wrong", and the documents support no stronger one. The interfaces described above were built to expose state [S11]; the contracts were written to allocate risk and channel requests [S14][S15]; the editorial pages were written to describe parts [S16][S17]. None of the three document types performs that joining. The buyer is the first party in a position to do it, and the cheapest tier is where a missing pairing is least likely to be noticed, because the plan card fits on a screen and the record does not.

Frequently Asked Questions

Does a screenshot of one command count as evidence?

Not on its own here. It shows a readout without naming the field's definition, the plan wording it is held against, or the moment either was captured [S1][S2][S12][S13]. The terms also shut out claims whose only basis is whatever the provider's website, system and lookup tools returned, so a one-sided capture is not the form the agreement contemplates [S14]. Keep it as raw material, then add the saved plan text, a field-by-field note and the ticket reply [S14].

Which memory field should go into the record?

Both, labelled separately. The kernel documentation defines MemTotal as total usable physical memory minus reserved bits and kernel code, and describes MemAvailable as an estimate [S1][S2]. The cgroup interface defines a separate file as the hard limit on memory usage [S3]. Writing one number without saying which definition produced it leaves the central ambiguity unresolved [S1][S3]. No allocation figure is asserted on this page.

What if the link setting does not match the plan page?

Check the preconditions before treating it as a discrepancy. The kernel documentation limits that attribute to interfaces implementing a particular driver query method, so a virtual interface may not expose it at all [S7]. The utility documents the setting as something that can be displayed and configured [S8], and the plan and policy pages make it subject to a traffic-based adjustment [S12][S15]. Record the absence as an absence, save the plan wording, and raise the question through the channel the terms name [S14].

Sources

All pages read on 2026-11-11 unless noted. HTTP status is recorded as returned on that day.

I opened the three comparison sites and the vendor pages only to record what they disclose and what their contracts require. No continuous prose from them survives above, and nothing here comes from a transcript. Fixed field names and interface names inside clauses are kept as printed. Every cross-document reading above is labelled as an interpretation or an inference and carries its chain. No measurement was taken for this page, no tool is suggested to the reader, no figure for how frequently a delivered machine differs from its plan is offered, and no ranking or analytics figure is used.

Marcus Reid
Systems Architect & VPS Infrastructure Analyst

Marcus has been working with virtualized server infrastructure since 2015, deploying and benchmarking VPS and cloud instances across DigitalOcean, Vultr, Hetzner, OVHcloud, Contabo, RackNerd, Hostinger, and the major hyperscalers. His work focuses on the gap between what providers advertise and what actually happens when you deploy a real workload. He verifies every price and specification against official provider pages before publication.