Analysis, research and decision record for aprendIA, IRC's WhatsApp teacher-training bot (Nigeria, Ecuador, Niger). Private, and permanently so: this repository contains quoted conversation material and references teachers by pseudonymous identifier.
Split out of the former irc_aprendia on 11 September 2026. Its companion is
aprendia_pipeline, which holds the ETL
and the Shiny dashboard.
The dividing line is: if the pipeline cannot run correctly without it, it belongs in the pipeline repo. Everything that judges the output belongs here.
So event_schema/ here holds the design notes, validation records, the gold-standard tooling
and nine checks/ scripts — while the 01–06 build sequence and the LLM prompts it consumes
live in the pipeline repo. Likewise classification/, and onboarding_schema/ is built this
way from the outset rather than being prised apart later.
The codebook is not here either. codebook.md is canonical in the pipeline repo, because it
documents the fields that repo produces.
analysis/ |
Briefing notes, thirty persona conversation reads, the questions repository |
research/ |
Questions registry, triage doc, course taxonomy, product feedback |
process/ |
Templates for the three recurring deliverables |
docs/ |
Engineering plans, the field reliability survey, the split plan, the episode model note |
event_schema/ |
Design, validation results, gold standard, checks |
classification/ |
Session-type classification: phase 0, rerun policy, escape hatches |
qual_quant_synthesis/ |
A nine-phase qualitative/quantitative synthesis |
onboarding_schema/ |
In progress |
scripts/ |
Research instruments and one-offs — not pipeline stages |
Finished deliverables are not here. They are published to a Dropbox folder colleagues can actually open, because the audience for a briefing note does not use git. Draft here, publish there — the axis is state, not cadence.
data/ is a symlink to the pipeline repo's data/, so relative paths resolve. Two variables in
.Renviron, which is gitignored and must be recreated per clone:
aprendia_data_path=<the offline_data location holding the raw Telerivet repository>
aprendia_pipeline_path=<path to the aprendia_pipeline checkout>
The first points at data that is not anonymised. It sits outside every repository deliberately, and nothing here should read it.
Before quoting any figure, check data/run_manifest.csv — it records each pipeline run's
timestamp, commit, row counts and data window, so a number can say which data produced it.
The Year 1 goal is a learning and data environment: a set of fields that states, field by field,
what is reliable and until when, so that Year 2 inquiry can start from a question rather than
from re-verifying the instruments. Much of what is here is that verification work —
docs/field_reliability_survey.md grades every detector, and
qual_quant_synthesis/method_faults.md records thirteen method rules, each earned by getting
something wrong first.
Read those two before trusting any number in this repository.