Step 4 of skills/analyse/SKILL.md now holds one new rule. Under it, where the engine repeats inside one of its claims something the operator supplied, the engine is to mark on that claim's own line every part it worked out for itself rather than took from the operator: a year it added, a date it calculated from another date and an interval, a figure, a name. It is not to write an added part inside a sentence reporting what the operator said, and where it uses an added part again in a later claim, it is to repeat the mark there.
What was missing
Since 1.2.0 the engine labels a default it chooses default assumed (engine's choice, not the operator's), and it records, for each scoping assumption, the ground it inferred that assumption from. Nobody had written the same requirement for the claim register, the part of the output where the engine states each claim in one sentence.
What was found
B13, one of the conformance probes kept for this engine — a fixed case with fixed answers and a written list of what counts as a failure — was run twice against 1.3.1 on 2026-09-07, on one case and one message settled before the runs and left unchanged between them. The operator typed "the unauthorised access was detected on 14 August, eleven days before the board was told" — one date and one interval. In both runs the engine added the year 2026 to that date and supplied 25 August 2026, a whole calendar date, for the day the board was told, and marked neither as its own. In the second run it wrote both inside a sentence beginning "The operator states that", so it reported a year and a date it had supplied itself as things the operator had said. Two statutory deadlines in the first run, and one count of days between two dates in the second, were then calculated from dates that included the added year.
What this release does not fix
The rule does not forbid the engine to work a date out, so a run of B13 under it would still be scored a failure if the engine again works the date out — which no run under the rule has tested. That probe is scored against eight criteria, and under one of them any date the engine adds counts as a failure, whether or not the engine marks it. The runs were made on 2026-09-07 and nothing in the case places its events in any other year, so 2026 is almost certainly the right year, and forbidding the engine to work a date out would make a correct inference count as a fault. A probe that scores whether added parts are marked has not been written; it is listed as outstanding work, as item REG-35, in the register of outstanding work kept in the Portable Edition, a separate build of this engine that does not depend on any one supplier's platform.
This release has no conformance run of its own. The two B13 runs are the record of what the engine does without the rule, and no run of a version that carries the rule can add to that record: such a run tests whether the engine follows the rule rather than what the engine does without it. conformance/results/runs/2026-09-08-NOT-RUN-insight-engine-1.4.0/10-VERDICT.md states that.
Nothing here shows that the engine's output has changed. The rule is in place, and no run has been made under it.
Also in this release
Thirteen capture files — the engine's own output from each run, verbatim — are published: the captures for B12, B12r3, B12r4 and the two B13 runs, in five run directories. Since the conformance folder was created on 2026-08-31, the files from each run have been split between two places: this repository holds the engine's own output, while the written procedure each probe was run under, which is not changed after the run, and its scoring stay with the rest of the probes in the Portable Edition.
README.md gave the download as insight-engine-1.0.0.plugin and the version as 1.0.0; both now read 1.4.0. The run count in conformance/README.md had not been corrected since five releases before this one; it is now correct. One sentence was added to docs/User-Operating-Guide.md: the version number printed twice in that guide is the guide's own and is not meant to match the plugin's.
Compatibility
The symbols the engine uses to grade a claim, the words it prints beside them, the way it numbers claims, and the key that explains all three are unchanged, so every output file the engine produced before this release still meets the rules for grading, numbering and the key. That follows from none of those four things having changed, not from a run: this release has none.
Cowork users: download insight-engine-1.4.0.plugin below and upload it from Cowork's plugin settings. Uploading over an existing install replaces it in place.