Gate semantic quality
I use structural heuristics to hold weak summaries out of downstream debriefs, so the artifact has to pass its own checks before it can influence the next one.
Executive intelligence system
Sortie
I built Sortie after reading a confident meeting summary I couldn’t fully trust and realizing a polished artifact without visible evidence was worse than no summary at all.
What it isAI-enabled executive intelligence system
What it is for
I use it to turn meeting transcripts, email, calendars, action state, and forecast signals into verified summaries, daily debriefs, and weekly retrospectives.
The problem
The summary had the names and actions right enough to sound authoritative, yet I couldn’t tell which claims came from the transcript, which came from accumulated context, and which were model inference. Once those distinctions disappeared, the next debrief inherited the same uncertainty.
I needed a durable record across meetings, email, action state, and forecast changes, with source authority preserved at every handoff and unknowns kept visible.
System output against fabricated inputs; no employer data.
Method: see How the evidence was made.The approach
This isn't a prompt wrapper. Sortie monitors my workday, processes raw transcripts, maintains action state, watches forecast-relevant changes, and produces daily and weekly synthesis under explicit rules for inference, preservation, and human review.
I direct the system in natural language and AI writes the code. Architecture, failure analysis, acceptance criteria, and the decision to ship stay mine.
With 771 meeting-summary files in the current vault snapshot, I have enough recurring patterns and hard cases to test whether a change improves the output or merely makes it sound more confident.
Design decisions
Treat silence as unknown
Before I call an action open, I scan recent interaction evidence for closure and report missing evidence as unconfirmed.
Separate human dispositions
I route context corrections through confirm, exclude, or transcription-error decisions because a plausible entity still needs an accountable disposition.
Recover the useful record
I take recovery snapshots before rewriting daily notes, restore missing meeting links, and verify repairs from disk. The system refuses a rewrite if it cannot save the recovery copy. External edits can still race with a write, so an independent audit checks the resulting record.
System output against fabricated inputs; no employer data.
Method: see How the evidence was made.
System output against fabricated inputs; no employer data.
Method: see How the evidence was made.Scale
Verified149tracked Python files
Verified1,733collected tests
Measured771meeting-summary files
Measured180daily debriefs
What remains unproven
The verified evidence here is repository and artifact scale. Tests were collected, not executed, for this refresh; file counts do not establish summary quality or time saved. The method is not proven at organizational scale until another operator or team uses it against a pre-period baseline and produces a measurable improvement in cycle time, quality, coverage, or cost.