Ingest 2026-08-23: training-data-quality EXPANDED (supply squeeze/model-collapse); OpenAI ZDR for frontier; code-review→outcome-validation

[hornof]

Summary

Daily Brief 2026-08-23. Mostly re-surfaces (simulation-scaling, watermark, agent-interface, Import 469, GLM-5.3 — all prior folds, deduped). Net-new: 1 new page + 2 folds.

New page:

• concepts/training-data-quality — training data as a contested + degrading substrate. Unifies two threads the wiki had been accumulating separately: the provenance/IP shock ([[amazon]] rare-books, Ingest 2026-08-17: Anthropic $65B; Stripe buys OpenRouter $7B; Amazon rare-books training data; Copilot-Autofix compromise #247) and the saturation/model-collapse risk (Pew "How Much of the Internet Is Written with AI?" — the "hall of mirrors"). Ties cleanly to [[simulation-scaling]] (Ingest 2026-08-22: simulation-as-scaling-law (new page); Nvidia harness>model; Anthropic IPO backlash-risk; 'everything's a neocloud' #256) as the synthetic escape hatch — the two are complements: this names the problem, simulation-scaling proposes the fix.

Folds:

• openai — Zero-Data-Retention extended to frontier models + private safety processing: enterprise privacy/compliance differentiation for regulated domains.
• agentic-engineering — two human-agent-workflow datapoints: (a) "More than just code review" — reviewer shifts from line-by-line → outcome validation ("linter → decider"); (b) Torvalds debugs a kernel with AI as a "tireless helper" that resisted the hard problem — a high-credibility floor-vs-ceiling datapoint, ballast against the "code is free" maximalism.

Watch-items: GLM-5.3 "beats frontier at 1/5 cost" — flagged lacks rigor (no benchmarks, undefined baseline); not folded, verify first; local-LLM quality gaps; Fire-HD ownership; llm-openrouter 0.7.

Judgment calls

• Created training-data-quality now because a second, distinct signal (Pew model-collapse) joined the earlier provenance shock — two facets of one uncovered structural theme, and it slots between amazon and simulation-scaling with real explanatory coherence. (This also promotes the standing training-data-provenance create-candidate under a broader, more accurate name.)
• Did not fold the GLM-5.3 "1/5 cost" claim — low-rigor source; kept as verify-first watch-item rather than letting it inflate ai-margin-collapse.

Test plan

• Daily-Brief dedup ran (2026-08-23.md had 0 log hits); re-surfaced headlines cross-checked vs Ingest 2026-08-16: multi-agent failure-modes; Grok CSAM harm; Willison local-inference; text watermark #246/Ingest 2026-08-17: Anthropic $65B; Stripe buys OpenRouter $7B; Amazon rare-books training data; Copilot-Autofix compromise #247/Ingest 2026-08-20 (brief): Mojo open-sourced (new page); GLM 5.3 'death of params'; Every Model Cheats; Willison sandbox cluster #253/Ingest 2026-08-22: simulation-as-scaling-law (new page); Nvidia harness>model; Anthropic IPO backlash-risk; 'everything's a neocloud' #256
• 0 delta ghost links across 5 touched files; roundup filename in log.md
• Ingest 2026-08-22: simulation-as-scaling-law (new page); Nvidia harness>model; Anthropic IPO backlash-risk; 'everything's a neocloud' #256 branch pruned at sync; single linear commit, base main

🤖 Generated with Claude Code

[hornof]

Correction (before merge): the first commit on this branch accidentally replaced the pre-existing concepts/training-data-quality.md (the 2026-05-30 data-quality / cognitive-core / data-as-moat page) — a slug collision I missed. The follow-up commit 38abec9 fixes it: all original content is restored and the new material (Pew model-collapse + Amazon rare-books provenance + simulation escape hatch) is integrated as a "Supply Squeeze" section. Duplicate index entry removed; roundup re-worded from "new" to "expanded." Net diff for the file is now +22/−8 (additive), not a replacement. Safe to merge.