Rapid Response Pipeline

The pipeline has two halves and they are not equally reproducible. Severity triage and summarisation are a language model making a judgement; the rest — which icon, which comp, which cells go in the spreadsheet After Effects renders from — is rules. Rules can run honestly in a browser, so they do. What you see below is rapid_response/mapper.py: _normalize() truncating summaries at 50 and 120 characters, _choose_icon() testing tags in its real order, and to_ae_rows() emitting one row per slide.

The tags below are the boundary. In production a model assigns them from the article body; here you set them yourself, which makes visible exactly where judgement stops and determinism starts. Everything to the right of that line is reproducible forever.

One breach story — Chrome moving to HTTPS by default — composed into a five-slide carousel: an alert, a statistic, the detail, the product line, the endcard. The headline, supporting line, icon, call to action and background image on every slide were chosen and placed by the system from a single summarization pass. The animation itself was built by designers months earlier and never touched again.

Ten stages, hourly, unattended. The spine runs left to right; the middle fans into the paper trail every run leaves behind; the end splits into two render rails that were both live — the same story payload driving a browser tool and After Effects.

Jitter has no API, so the daemon uses the editor like a person. The brand animation lived in a browser tool with no automation surface at all. Rebuilding it somewhere scriptable would have thrown away the designers' work, so instead a local service drives the real editor: it logs in once through a real Chromium window — MFA included — and keeps the session, then duplicates the template, selects a comp, and types each headline as genuine keyboard events. It reads every field back to verify what landed, because a dropped keystroke in a headline is invisible until it ships. When the tool has no API, resourcefulness is the skill.

Automation was allowed to touch exactly five layers. Per comp: headline1, headline2, buttonICON, buttonCTA, image — verified identical across all eighteen comps in the node map, with no escape hatch. Animation curves, easing, timing, layout, type and colour were not addressable by the pipeline, so they could not drift. Copy that would overflow was clamped upstream at composition time rather than breaking a layout downstream. This is what made the system safe to run unattended: it was structurally incapable of overwriting the brand system.

One payload, two render engines. The same composed story data drove the Jitter rail and the After Effects rail — a watcher picking up render sheets and calling aerender against the original template. Both were live; After Effects came first and kept working after Jitter became primary. A tool change was a new rail, not a new pipeline.

Worth being precise about: the severity gate is the part people expect to do the filtering, and it accounts for five rejections. Stories overwhelmingly die earlier — at the keyword sieve or at the fetch. The gate's real job was keeping model spend proportional to the news, not culling the feed.

Built solo for a real client context and run through late 2025. It is archived now — the point of this page is the preserved evidence, not a system waiting to be restarted. The class of problem it answers is the one where the volume or speed a brand needs exceeds what a manual team can produce, and the brand system still has to survive contact with the automation.