Fixed
-
Caption deduplication can no longer invent a second person out of one flapping face. Field validation caught a phantom that survived every earlier fix: "A person walks across a paved walkway, then Nico stands near a green bush outside the house" — one person in frame, narrated as two. Neither the summarizer nor face recognition was at fault. The camera captured the same human twice in a row with byte-identical captions; recognition named him on the first frame and, one frame later at a hard angle, returned "Unknown Person" whose nearest gallery match was a different enrolled person just past the merge bound, so the identity merge correctly refused to consolidate them. Deduplication then collapsed the two identical captions into one frame and merged their identity lists, handing the summarizer a single frame carrying two identities — which is precisely the evidence its single-actor bias and the verified-single-person constraint both stand down for, since a frame that really shows two people must be allowed to say so. The two-actor summary was faithful to a prompt that had been given a fact no frame ever contained. Identities merged across a run of duplicate frames are now capped at the largest number of faces any single contributing frame actually held, so a temporal union can no longer masquerade as people standing together. Excess entries are dropped weakest-evidence first (empty, then "None", then "Indeterminate", then "Unknown Person", matched in every legacy gallery spelling), and a second real name is never dropped — two enrolled people across a run stay ambiguous between one flapping human and a genuine pair, and erasing someone who is present is a worse failure than naming them imprecisely. A frame that genuinely recorded two faces keeps both, and recognized-people lists can only get more precise, never emptier, so no Sentinel rule changes its firing. (#543)
-
The repeated-scene sentinel no longer leaks into summaries as a person's action. A frame the vision model answered with the control reply "Scene unchanged." is normally dropped, but it is deliberately kept when face recognition found someone the model missed — otherwise the detection would vanish from the summary, the sensor and the notification. The frame was kept carrying that control reply verbatim as its scene description, so the summarizer was handed a frame that said nothing had happened alongside an identity it was required to mention, and wrote around the contradiction: "A man in a gray shirt walks out of an open door onto the porch, then stands still as Lindo." The frame is still kept, now under a neutral stand-in caption instead of the control reply, mirroring the treatment the empty-and-failed-caption path already used for the same reason. That stand-in has to continue the subject rather than introduce one: the first wording, "A person is present.", was an indefinite introduction, and placed after a frame that had already introduced the same human the summarizer read it as somebody new arriving — "A man in a dark shirt stands on the porch at dusk, then remains there as a person becomes visible nearby", one human in frame, narrated as two. The summary prompt's introduce-once rule could not undo it, because that rule governs what the model writes while this caption is what the model is given. The caption now continues the subject, still states that a person is there so the detection survives into the summary even when the vision model missed them entirely, and still makes no claim about motion, since the sentinel means the model saw no change. Describing only the scene was measured and rejected: it dropped the person from the summary in three runs out of ten whenever recognition had no name, which is the erasure this branch exists to prevent, and it would also have stopped the frame's snapshot from being eligible as the notification image. Consecutive sentinel frames holding a person share that caption and collapse into one frame, which the identity cap above keeps from claiming two people. (#543)
-
Camera summaries stop introducing one person twice. With face recognition unable to identify anyone — night infrared, subject facing away — the summarizer has no name and no verified-single-person block to work from, only its own rules. Those rules told it to assume one unknown person across separate frames for the purpose of counting, but said nothing about how to refer to that person, and every frame is captioned independently on purpose, so one man produced "a person" in one frame and "a man in shorts" in another. The summary stitched them together: "A person walks near the house entrance, then later a man in shorts stands at an open doorway with a black cat nearby" — two people to any reader, without ever using a plural for the counting rules to catch. The rules now cover reference as well as counting: when the frames describe one individual, that person is introduced once, by name when there is one and otherwise by the most specific description any frame gives, and every later mention continues the same subject instead of starting over. The verified-single-person block carries the same direction, which also removes the last case where it named the right person in the wrong place ("A person walks across a paved walkway, then Nico stands near a green bush"). Frames that genuinely show two people are still narrated as two. Measured against the local summarization model on the real prompts, ten samples per case: the unidentified-person summary went from zero clean runs to ten of ten, the verified-single-person summary from three of eight to ten of ten, and a genuine pair stayed correctly plural throughout. (#543)