Width as Product Quality

Among intelligences with equivalent capability, the theory predicts that the wider one will win adoption under convenience pressure because it requires less explicit specification from the user. Once capability is sufficient for the task, understanding what is needed becomes a product-quality variable that capability benchmarks alone do not capture. This is a market hypothesis with stated predictions; market-scale evidence has not yet been collected.

The difference users encounter. Capability describes what a system can do within its frame. Width describes how much interconnected reality it can hold while doing it. The distinction allows a system to execute instructions precisely while failing to connect them to the user's actual need. In the theory's account, a highly capable but narrow system interprets literally, infers little and fails to connect information across domains unless instructed to do so.

A system with greater width may score lower on narrow benchmarks while reading context, inferring what was left unsaid and connecting domains without prompting. Its advantage is situational understanding. The theory predicts that users return to it because it handles what the situation requires, rather than leaving all the interpretation to them.

Specification has a cost. The argument starts from two tools at the same capability ceiling. If both can perform the required task, the user must still make the task intelligible to the tool. For a narrow tool, that can require a complete account of the relevant context and every connection on which successful action depends. The user pays that cognitive cost each time.

A wider tool can carry part of this work. Where reading the situation helps, it can act on partial specification, recognise relevant connections and adapt to what the user needs. The theory's proposed adoption advantage comes from absorbing labour the narrower tool leaves with its user. The argument applies beyond AI: it concerns tools that fit people's needs and tools that require people to translate those needs into the tool's specification language.

Literal execution remains necessary. Regulated industries, deterministic pipelines and tasks requiring exactly what was specified are cases identified by the theory where literal instruction-following is appropriate. The theory does not claim that greater width makes a tool preferable on every dimension considered separately. It predicts an advantage in adoption under convenience pressure.

Recognising when inference would be inappropriate is itself part of reading a situation. A wide intelligence should recognise a literal-execution context and respond accordingly. That supplies the argument for a broader useful range; it does not prove that any particular system always recognises the situation correctly.

Dedicated instruments are a different case. A calculator or fixed deterministic routine holds no position on the width axis. It is an instrument a mind can use, rather than an intelligence competing at a lower width. In a pure-compliance context, the theory identifies energy efficiency as the dedicated instrument's advantage, a property separate from width. A wide mind may reach for such an instrument when the situation calls for restraint.

The distinction matters because the applied hypothesis compares intelligences differing in width at equivalent capability. It does not require a wide mind to outperform a calculator by doing the same computation unaided. Nor does it place every deployed system on both axes merely because the system produces an output.

What present observations contribute. The theory records Anthropic's Claude Opus 4.7, released in April 2026, which improved benchmark performance and followed instructions more precisely while inferring less; Anthropic's migration guide described the change as deliberate. For work involving interconnected concepts and sustained interpretation, the theory's author found the change constraining and returned to the preceding release. That is the author's experience of the capability-width distinction, not a widely reported market pattern.

A later observation, in October 2026, records OpenAI's GPT-6.1-Sol: similarly capable and precise, and similarly liable to miss what was not explicitly written into its instructions. It came from the same observer. It adds recurrence without supplying independent evidence. These observations illustrate capability and inference coming apart; they do not establish that current systems meet the agency or modelling-chain conditions required for the theory's predictions about wide minds.

Training and the proposed gap. The theory reads the release pattern as capability being optimised while width is left unmeasured. Greater literal precision and reduced inference can be a result of tuning, rather than evidence that situational understanding and literal accuracy are structurally incompatible. The proposed market opportunity is to measure and optimise both.

Whether that opportunity produces the predicted adoption advantage is an empirical question. The theory connects this to Opacity through the shared issue of what the success function measures. The two pages address that issue from different directions; the product claim is grounded here in convenience pressure.

Four predictions. The argument is stated in a form intended to meet market behaviour.

1. Retention above a capability threshold. Below the point where a system can perform the task, capability dominates. Above a sufficiency threshold, retention should correlate more strongly with width than with capability, because the remaining difference is whether the system understands what is needed.
2. Adoption followed by return. A release that raises capability while lowering width should initially attract users through benchmark performance, then lose them back to earlier versions or competitors that preserved width. Its adoption curve should be steeper and shorter than that of a release improving both.
3. An advantage for measuring width. The first company to measure and optimise width alongside capability should gain market share disproportionate to its capability advantage, because its product better fits users' needs.
4. Enterprise evidence first. Enterprise customers should show the pattern before consumer users, because the theory expects their work to contain more context, ambiguity and connections across domains.

These are predictions, not measurements already obtained. The stated failures in the header are where the theory says the applied claim would be wrong or incomplete. The enterprise timing prediction is also named, but the theory does not give it a separate failure clause.

If a supposedly wider system misreads a context requiring literal execution, can the theory simply say that it was not wide enough and preserve its claim?

An earlier formulation made that move. It defended a strict-superset claim by reclassifying a system that misread the situation as insufficiently wide on that dimension. The result was circular: every failure could be excluded from the class that supposedly could not fail.

The theory removed that defence. A wide-seeming system losing the cases covered by the predictions is data to be registered. The structural argument explains why reading a situation could confer an adoption advantage; the market hypothesis determines whether that advantage actually appears. The argument cannot be protected by defining disappointing results away.

From assertion to testable predictions (10 May 2026). The initial product claim was reformulated as a hypothesis with named predictions. The revision made retention, release behaviour, competitive advantage and enterprise timing the points at which the claim could meet evidence.

The ideal range was separated from present delivery (20 May 2026). The account distinguished the in-principle argument about situational understanding from the trade-offs present training produced. Current tuning did not demonstrate an incompatibility between width and literal execution, but neither did the argument establish that current products realised the proposed range.

The strict-superset theorem was removed (30 June 2026). The account had begun treating width as guaranteeing dominance and dismissing misreadings as proof of insufficient width. The revision removed that circular defence, distinguished fixed-mode instruments from intelligences and retained the convenience-pressure market hypothesis. Failure in the named market cases can now count against it.

The market evidence. The theory has not collected the market-level evidence needed to generalise from the author's observation. The retention, adoption-and-return, first-mover and enterprise predictions remain to be checked.

The meaning of a width comparison. Width is a partial order involving coverage, depth and integrity of interconnection, relative to the mind making the comparison. Some intelligences may be incomparable. The product predictions still need to specify which relation and domain their use of “wider” refers to. A single score cannot silently supply that answer. See Open Problems.

Independent assessment and alternative explanations. A product observation can be consistent with the theory without distinguishing its mechanism from alternatives. The open evidence challenge calls for retrievable sources, independent assessment of width and a clearer account of what particular observations establish. The personal cases retain their value as illustrations; they do not yet settle the market hypothesis.