# Debrief Ledger

> Read several editions of a recurring technical event's debrief notes for what keeps going wrong, then write the
> learning log the next team inherits. Findings that keep coming back are separated from one-offs, grouped into
> themes, counted across editions, held as a candidate or a pattern, and turned into one deliberate change routed to
> its owner. Built on the SkillSafe skill @samber/event-continuous-improvement, from
> samber/dev-event-organizer-skills (MIT), by Samuel Berthe. The log format, the occurrence bar, the menus and the
> templates are the skill's own construction, not established organizer practice; only MLH's keeping-records guide,
> DevOpsDays' organizing.md and Fraser and Mancl (2023) are published practice. Every threshold is self-set.

https://debrief-ledger.skillsafe.ai/

## What is free
- The editions parsed from pasted notes or dropped files ("## 2024", "Edition 3", "Code Fair 2025 debrief", "---"), ordered by year or edition number where every heading carries one.
- Every entry through the signal gate: a candidate only when current, reproducible and generalizing; expected work (handled at no cost), closed (the condition is gone), too thin to file, or a thing that went well - filed with a reason, never counted.
- Names and handles taken out of the record and rewritten as conditions ("no owner and no backup for this step"); nothing sent or exported carries them.
- Themes grouped by what would prevent them, counted by the editions they appear in - one-off, candidate (two), pattern (three) - out of the editions held, with safety exposures dropping the bar.
- Routing: operational to event-planning-timeline, demand to event-market-fit, size ceilings to event-growth-strategy, recurring team load to event-team-structure, a named vendor that recurs to event-vendor-sourcing; a cost trend stays here.
- The three ranked menus (log persistence posture, trend-detection depth, log-audit cadence) re-ranked against the interview, each move naming the answer behind it, ruled-out rungs deleted.
- One change slot per theme (at most three), the KPI shortlist (2-3, each with an owner and an action), the closure pass on last cycle's changes, the Keep/Update/Consolidate/Replace/Delete audit and its one pass threshold (zero unclassified entries), a verdict, a draft learning log (.md), themes (.csv) and everything (.json).

## The metered lanes
Two tasks on gpt-terra, metered per run, routed by `task`. Input string fields: `task` (trends or log), `event`, `logs` and `facts`.
- `trends` replies with `summary`, `honesty`, `verdict`, `themes`, `changes`, `not_ours`, `menus`, `kpis`, `closure` and `caveats`; the page holds the verdict, every recurring theme's count, label and route, the change slots, the hand-offs, the picks and the shortlist to its own.
- `log` replies with `title`, `honesty`, `entries`, `not_counted`, `changes`, `opener`, `audit`, `outcome_log` and `handover_note`; the page requires every theme with its first-seen edition, count, label, status and audit call, owners that are roles, the opener as given and the zero-unclassified threshold.

## Links
- API tutorial: https://debrief-ledger.skillsafe.ai/api.html
- Source skill: https://skillsafe.ai/skill/@samber/event-continuous-improvement