Current examples, and an earlier benchmark, kept apart.

Research overview. Benchmark record current as at 10 August 2026.

Current Frame Free examples illustrate how the guide behaves on fictional inputs. The earlier benchmark tested a method file that predates Frame: it changed how the model worked, and it did not beat an informed baseline on judgement. Neither is a customer result.

Current Frame Free examples.

These use the Frame Free guide that is offered today. They show what the questions and the handoff look like on a specific decision.

The guide’s example: The worked example printed on page 5 of the guide: a vendor choice that turns out to be a question about shift rules. Read the handoff
Five examples: A vendor renewal, a hiring decision, a capital purchase, a struggling programme and a pricing change, each on fictional inputs, with its provenance stated beside it. Read the examples
Boundary: Agent-run and worked examples illustrate behaviour. They are not customer outcomes, performance benchmarks or independent validation, and Frame Free has not been benchmarked.

Earlier method benchmark.

One preregistered study, 45 scored runs, two models, three fictional cases, eight checks. Everything below is stated so that a reader can hold GreenSquare AI to it.

Observed: Process behaviours. Under the tested conditions the loaded method changed how the model worked: it asked before answering, compared real options including doing nothing, and tagged its evidence on every run. The first of those is definitional, because a single-turn baseline cannot ask.
Not established: Better judgement. On the three judgement checks the informed baseline matched or exceeded the loaded condition.
Which artefact: A method file that predates Frame. The study is not a test of Frame Free or Frame Pro, and its result does not transfer to them.

Evidence statusReviewed 10 August 2026

Claim tested: That loading the method file makes a capable model reliably run a disciplined process on a consequential decision, and surface the decision the facts point to, where a baseline does so only sometimes.
Result: Process: supported under the tested conditions. Judgement: not supported. The informed baseline matched or exceeded the loaded condition on all three judgement checks.
Read with: Caution. One scorer, no independent re-score, 15 runs per condition, and two of the 8 results are definitional rather than measured. All 18 protocol deviations are published.
Still open: Independent re-scoring, publication of the 45 transcripts and the sealed reference answers, per-case counts for seven checks, and a refresh on current model versions.
Product relevance: The study tested a method file that predates Frame. It does not validate Frame Free, and it says nothing about Frame Pro.

One controlled statement of the result.

The counts that carry it are below, taken from the benchmark record. Every figure on this page renders from the same data file as the record, so the two cannot disagree again.

Read the full record

Under the tested conditions the method improved specified process behaviours. It did not outperform the informed baseline on the judgement checks.

Judgement checks, informed baseline against loaded condition
Judgement check	All facts (baseline)	Loaded method	Reading
Surfaced the hidden decision	15 of 15	14 of 15	Baseline higher
Direction matched the sealed key	15 of 15	15 of 15	Tie
Caught the planted bias	15 of 15	15 of 15	Tie

On the process side, the loaded condition asked before answering in 15 of 15 runs against 0 of 15, compared real options including doing nothing in 15 of 15 against 1 of 15, and tagged its evidence in 15 of 15 against 0 of 15. The first of those is definitional: a single-turn baseline cannot ask. The other two are measured.

The boundary sits beside every product claim.

Supported: A change in process, on two models and three fictional cases, 15 runs per condition.
Not established: Better judgement than a well-briefed model. Performance on other decision types, organisations, user populations or model versions. What an operator arriving with partial knowledge and a blind spot would get, because that condition was not run.
Not tested: Frame Free or Frame Pro. The study used an earlier method file, identified by its published hash. Nothing commits a later release to being that file.

Corrections to the public record.

Kept here in date order so a reader can see what changed and when, rather than finding a page that quietly says something different from the last time they read it.

10 September 2026: This page previously reversed the judgement result, crediting the loaded condition with matching or exceeding the informed baseline. The benchmark record says the opposite, and the counts support the record. The sentence is corrected here and in every place it was repeated.
10 August 2026: The published grid was re-cut from 54 runs to 45 by removing nine design-phase pilot runs the protocol excluded, and every disputed cell was adjudicated against the sealed answers. Every count on the benchmark page reports the 45.
11 August 2026: The preregistration commit identifier was corrected. The one previously published exists in no ref; the real commit is linked from the benchmark record.

Inspect the claim, the method and the frozen inputs.

ResultsBenchmark recordAll 8 checks, 18 deviations, eight limitations, sources and what is withheld.DesignMethodologyConditions, cases, scorecard, preregistration and scoring.SourceFrozen preregistrationCases, protocol, rubric and hashes, at the commit that fixed them.

What would change the result.

A condition that resembles a real operator, arriving with most of what matters and one thing they cannot see. An independent re-score against the published rubric. Publication of the transcripts and the sealed key. A refresh on current model versions, with the date-stamped history the protocol promised. None of this is scheduled; when any of it is done it will be recorded here with a date.

Try the current guide on your own decision.

Frame Free is the guide offered today. Judge it on a real decision, with the boundaries above in mind.

Get Frame Free PDF See the examples