New corpora added to tv (following the repo's
playlist→captions→clean pipeline): SAIA (@saia_ar) 2026 uploads — 37
videos, SAIAConf 2026, Spanish, ~116k words; Pragmatic Engineer 2026
uploads — 42 videos (30 podcast interviews + 12 Pragmatic Summit talks),
English, ~620k words. Raw .vtt excluded per convention. Caveat: yt-dlp
approximate_date is ±1 month — re-verified real upload
dates at the 2025/2026 boundary (dropped a6aYliF6Mq4, really
2025-12-23). Hit persistent HTTP 429s on caption downloads; patience +
per-language retries got all 79 primary-language transcripts.
saia-landscape-report.md (fc7a324): Argentina AI
scene from SAIAConf 2026 — six narratives (PoC
explosion/12%-reach-production, culture>tech, cities-as-vanguard,
admitted replacement/Pergolini, data-as-moat, assets+window) and gap
taxonomy ranked by evidence: talento/educación (312) dominant; capital
(33) the quiet gap; ecosistema (321) is the #2 keyword category.
Per-talk notes with quotes in
analysis/data/saia_notes.md.
prageng-landscape-report.md (93993b8):
production-stage discourse — bottleneck moved from writing to trusting
code (Hejlsberg et al.), measurement discipline (Tacho: 92.6% adoption,
26.9% AI-authored merged code, ~10% flat savings, incident divergence),
AI amplifies org-as-is (MIT GenAI Divide, Fowler/Beck retreat), 6× cost
growth (Uber), fundamentals up but cognitive-erosion risk (Osmani);
regulation essentially absent (112/620k words). Notes in
analysis/data/prageng_notes.md.
Rule-of-5 review (16b1c5c) caught a CRITICAL:
PragEng actor map listed 2025-corpus guests (Shopify/Thawar, Guthrie,
Kalra) from memory — corrected against playlist JSON; Booch relabeled
IBM. Also: Gartner attribution softened, SAIA gap numbering synced to
counts, deep-read coverage corrected to 11, verification status +
single-coder limitation + ASR-noise policy documented in both caveats.
Verdict: READY after fixes.
Fix pass (6b962f6): layer-1 scan scripts committed
(analysis/layer1_saia.py, layer1_prageng.py),
closing the review's residual auditability item. Re-run exposed two
regex bugs in the original ad-hoc SAIA scan (ley substring
matches, accented país never matching normalized text):
regulación 38→32, ecosistema 260→321; report + notes synced. PragEng
totals reproduce exactly.
Next: the cross-corpus comparison (Argentine
adoption-stage vs production-stage discourse) is the obvious follow-up
and would need full spec discipline (sealed hypotheses, joint codebook,
second coder) per analysis-spec.md conventions; alternative smaller
step: run the versioned layer-1 scripts as a regression baseline before
any codebook work.
Session: 19:53 (talleres+tv,
close)
Rule-of-5 review of the PragEng/SAIA leverage pass
(talleres 9eab127): verified all 20 video IDs against tv
playlists + 6 quotes against raw ASR transcripts. Caught a HIGH misquote
— "la IA vino a jubilar tareas" dropped Biazzi's negation and inverted
his meaning; restored verbatim. Also: pandemia pseudo-quote replaced
with verbatim, una única verdad fragment, C-36 figures added to the
open-gaps spot-check queue, feedback-log intake entry corrected
pre-commit (append-only overclaim avoided). Verdict READY. typos hook:
_typos.toml allowlist grew (KSkcgIYQy0U, importante)
Corpus-extension candidates (tv
7c3e666): 15 corpora/people tiered P1–P3, question-first from all
research to date (both landscape reports' actor maps, talleres
deep-research batch, verify-before-use gaps). P1: DORA/Forsgren/DX
primaries, Husain/Shankar evals, LABI (Meeting 2025 BA + IdEM, es-native
scientific corpus), Haase/napari/QuPath tool primaries, Latent Space.
Anti-candidates + method checklist included. README indexed
Beads epic + tickets (tv): epic tv-oio
+ 11 children (tv-oio.1 probe = critical path gate; .2–.6 P1; .7–.10 P2;
.11 P3 backlog; .6/.9 also gated on tv-3b9). Each carries talleres
integration acceptance (teaching-topics C-xx + evidence-map +
feedback-log)
Issue-review (rule-of-5) over the ticket set: 1
HIGH fixed — tv-oio.8 wrote into the version-stamp table owned by
tv-oio.5 with no dep (bd dep add tv-oio.8 tv-oio.5);
metadata.files reassigned (candidates doc → probe/epic only; analysis
tickets now claim planned outputs: labi-landscape-report.md etc.);
base_commit 7c3e666 anchored on all 12; quantified meters added to
.3/.4/.6/.9 (≥10 verbatim spot-checks + 100% video-ID grep + provenance
header — encoding the rule-of-5 lesson); .5 anti-goal (stamps cite
maintainer-primary only); .11 trigger-only acceptance. Verdict
READY_TO_WORK. Dolt pushed
Next: claim tv-oio.1 (yt-dlp probe, ~60m) to
unblock the wave; talleres-side: C-34–C-43 fold into 73/74/75/82 designs
at their next design pass; CABA §7 decisions still awaiting Sasha
Diagnosed why Jellyfin didn't pick up Chinatown:
server runs as user quadlet (systemd-jellyfin, ports
8096/8920), media mount ~/PROJECTOR → /media; the only
library was a TV "Shows" type rooted at /media with no
Movies library, and the loose chinatown-*.mkv at
/media root was ignored (TV libs expect
Series/Season/Episode).
Restructured to
~/PROJECTOR/Movies/Chinatown (1974)/Chinatown (1974).mkv +
moved all 6 series into ~/PROJECTOR/Shows/, repointed the
Shows library via media.mblink.
Two follow-on bugs: (1) TV lib rooted at
/media swallowed /media/Movies as a bogus
"Home Movies" series; (2) my echo left a trailing newline
in media.mblink → path /media/Shows\n →
library silently empty, restarts didn't re-trigger a scan.
Fix: Sasha deleted + recreated the Shows library
via the web UI — cleanest path. DB (read-only sqlite on
BaseItems) now shows Chinatown as a proper Movie + all 6
series with seasons/episodes.
Next: redundant Chinatown copy in
~/Downloads/MOVIES/DONE/ (~1.8 GB) could be removed;
consider adding a Movies-type convention to the
to-projector pipeline so future films land in
Movies/Film (Year)/ directly.