Confidence, absence and editorial status

Every published figure separates four questions: does the quantity have a number, how strong is the evidence for that number, what is the number’s basis, and has an editor approved it? They answer different questions and are stored separately.

The profile table shows Evidence and Basis in separate columns. A number can be reported but measured, or confirmed but theoretical. Neither label replaces the other.

Basis: how the figure was obtained

Basis is not a quality ranking. A theoretical peak FP32 rate is exactly as legitimate as a measured one. It answers a different question and belongs to a different comparability group. Status is always displayed with a figure.

derived is reserved for figures this project computed. A figure a magazine derived in 1987 is estimated and cites that magazine; only numbers produced by the pipeline’s own formula registry are derived, because only those can be recomputed and re-checked on every build.

estimated carries two further requirements, because an estimate is only worth publishing when the reasoning behind it is visible. It must use the estimate-from-formula method, which keeps it out of the comparability group of figures somebody measured or specified; and it must name a research record carrying the formula or the procedure it was computed by. The sourcing threshold is the ordinary one: one tier A source or two agreeing tier B sources. An estimate whose method cannot be cited is not a low-confidence figure but an absent one.

Basis does not decide whether a multiplier may be produced; the comparability group, evidence level and editorial status do. But a multiplier built on an estimate says so in its generated caveat, so a result never reads as firmer than what it was computed from.

Method, provenance and evidence stage

The confidence status says what kind of number it is. Three further fields say how it was produced, who produced it, and whether the hardware existed yet. Only the first changes what the figure may be compared with.

• Method comes from a closed registry, methods-v1, in the same way units and metrics do. It names the operation behind the figure: a nominal clock, a theoretical peak, a design capacity, a published SPEC run, a wall measurement. It is part of the comparability group, so a theoretical peak never divides into a measured result. Before the registry existed the field was free text, and two curators describing the same kind of figure could split the catalog along a seam that said nothing about the hardware.
• Provenance — one of vendor, independent or community, recording who stated the figure. It is deliberately not part of the comparability group. A manufacturer’s silence must not cut its hardware off from every number somebody else published: an independently established clock for a chip whose maker never quoted one belongs in the same row as the clocks that were quoted.
• Evidence stage — either shipped or pre-launch, recording whether the hardware had reached buyers when the figure was published. A pre-launch figure must stay provisional, which is what keeps it out of every automatic multiplier. It is not filed under a method of its own, because that would only hide it in a separate row: the announced number is shown beside the shipped one, with the reason no multiplier follows written underneath.

Sourcing rules apply identically whoever the source is. Community documentation is not a lower tier by definition. A wiki is tier C because it aggregates, not because enthusiasts wrote it; a carefully sourced independent teardown is tier B on its evidence.

Evidence level: confirmed, reported and rumored

Every numeric record has an evidence level. It records the best support for the specific value, not whether a manufacturer happened to publish a specification. A manufacturer’s silence cannot turn an independently reported number into an unknown.

Tier C can support only a rumored value. Repeating the same claim across aggregators does not promote it. Sources must support the value itself: an otherwise authoritative document that says nothing about the quantity does not meet the tier A threshold.

reported and rumored values cannot feed automatic ratios, derived claims or promoted comparison presets. A comparison orders available figures as confirmed, reported, then rumored; an unresolved conflict does not select a number automatically.

Absence states

unknown, not-applicable and unverified are not evidence levels. They are absence states.

A missing figure is never zero, never an empty string and never omitted silently. It is one of three explicit states. The first two are findings about the machine and mean opposite things; the third is a statement about this project, and is the only one that is nobody’s fault but ours:

• unknown — the quantity exists, but a recorded search found no published candidate or no compatible candidate. Its unknownAudit names the review date, research routes, conclusion and a short summary. At least one route must be outside manufacturer documentation. Manufacturer silence alone cannot establish this.
• not-applicable — the quantity does not exist for this machine. The Commodore 64 has no floating-point unit, so its peak FP64 rate is not “0 FLOPS” and not “unknown”; the question does not apply.
• unverified — a document is cited for the quantity and nobody here has read the page yet. It contains no number and carries no note, because anything it said about that document would be invented, and it never claims the source is silent.

The third state exists because the first was doing two jobs at once. A figure recorded as unknown asserts that somebody looked; a backlog of unread manuals recorded the same way turned an unfinished job into a claim about what the world never published. Separating them means a reader can tell the Xbox Series X, whose full specification table contains no wattage at all, from a machine whose manual is sitting on a shelf here unopened. An unverified figure is counted as a gap rather than as coverage, so recording one can never make this catalog look more complete than it is.

Collapsing either into zero would be the single most damaging thing this dataset could do. Zero is a value: it sorts, it charts, it divides, and it would silently claim that a machine was measured and found to do nothing. Absent quantities carry no normalized twin, never enter a formula, and are rendered as text rather than as a bar of length zero.

An absence may carry a short note explaining it, which is displayed instead of the value.

Records that contain only documented absences remain public in the research section. They keep their source trails, but do not appear in the catalog, search, or new comparison choices until a numeric measurement is available.

When one machine answers twice

Some machines answer a question more than once, and neither answer is wrong. An Apple M1 runs four cores at 3.2 GHz and four at 2.064 GHz, at the same time; asking “what is the M1’s clock” has two correct replies and no single one.

A figure may therefore name the part of the machine it describes. The part is not part of the comparability group. A clock is a clock whichever cores hold it, so both figures reach the same row as every other machine’s clock, where a reader looking for the M1 expects to find it. What the part does is make the two figures two figures rather than a disagreement: they occupy separate slots, and validation keeps treating a genuine second value in one slot as a conflict.

The cost is that such a machine offers no multiplier for that quantity. There is nothing to divide, and the comparison says so in those words rather than leaving a cell that looks empty.

Editorial status: approved and provisional

Editorial status is about the review process, not about the hardware.

• approved — an editor has checked the figure against its sources and accepted it. Approved measurements must satisfy the full sourcing policy, and only approved records may feed automatic comparisons.
• provisional — the figure is recorded and visible, but not yet accepted. It might be sourced only from a tier C aggregator, it might be the subject of an unresolved conflict, or it might simply be awaiting review.

The consequences are enforced by the pipeline rather than by convention:

• A provisional record can never appear in an automatically generated multiplier. The ratio formula refuses provisional inputs outright, with provisional-record as the stated reason.
• A provisional record can never appear in an editorially promoted comparison preset.
• An approved measurement without sufficient sourcing fails validation and stops the build. It is not downgraded automatically. A person must source the record properly or mark it provisional.

Sources carry the same status for a different reason: an approved source is frozen against data:fetch. If the upstream document changes, the change is reported for review instead of being written into the manifest, so a silent edit upstream can never rewrite a figure here.

Conflicts

When two acceptable sources state different values for the same subject and metric, that is not resolved by picking the newer one or by averaging. Averaging two disputed figures produces a third figure that no source supports.

The competing values go into the conflict ledger with their evidence, and an editor records a decision:

• accepted — one candidate is chosen, with a written rationale and a date. The rationale is part of the public record.
• unresolved — the disagreement stands. Validation warns on every run, and figures for that subject stay provisional until someone settles it.

An unresolved conflict remains visible instead of becoming a silent omission.