Dumpster AI Research Map

Key jargon

Term	Plain-language meaning
Research intake	Uncurated material collected for later assessment.
Signal	A potentially useful idea, source lead, or pattern found in noisy material.
Triangulation	Checking a claim against independent or stronger evidence.
Promotion	Moving a supported conclusion from intake into durable curriculum material.

Key concepts

The dumpster is a lead generator, not an authority.
Preserve provenance and confidence while promoting only claims that survive verification.

Concept map

flowchart LR
    A["Inventory noisy intake"] --> B["Extract claims and source leads"]
    B --> C["Triangulate against primary evidence"]
    C --> D["Promote narrow supported lesson"]

/Users/tester/code/atoz/dumpster/AI contains 3,078 files and 45 curated summary/BLUF notes at migration time. It is an intake archive, not a publication source. The curriculum uses its strongest patterns while preserving caveats.

Intake	Useful curriculum contribution	Evidence status
`2026-07-31-github-rasbt-llms-from-scratch`	Tokenization → attention → GPT → pretraining → fine-tuning progression	Repository/README captured; code not executed in intake
`2026-07-31-arxiv-prompt-chaining-scholarly-report-generation`	Structured intermediate artifacts and stage-level workflow evaluation	Paper captured; narrow reported experiment
`2026-08-10-youtube-stanford-cs329a-planning-multi-step-reasoning`	Planning/search/parallel reasoning concepts and cost tradeoffs	Lecture transcript; follow primary papers for claims
`2026-07-25-x-0xclodex-agent-memory`	Graph/state hypothesis for durable agent coordination	Research lead; underlying primary source absent
`2026-07-25-x-0xcodila-agent-swarm`	Loops are not durable memory; explicit state matters	Research lead; social post/image incomplete
`2026-08-11-defcon34-taming-the-swarm-agentic`	Probabilistic agents need evidence-backed deterministic gates	Official conference deck; claims not reproduced
`2026-08-29-offsec-ai-harness-research`	Stage-level diagnostics and differential acceptance contracts	Captured 2026 papers; author claims, unreplicated
`2026-08-30-offsec-ai-harness-research`	Trace-backed grading instead of terminal success tokens	Captured paper; author claims, unreplicated
`2026-08-31-offsec-ai-harness-research`	Authorization-to-effect, outcome reconciliation, safe retry lineage	Captured paper; author claims, unreplicated
`2026-09-01-offsec-ai-harness-research`	Separate evaluation validity from execution authority	Captured papers; author claims, unreplicated
`2026-08-02-linkedin-akshay-pachaar-llm-production-7-stages`	Production lifecycle lead: train/adapt/optimize/serve/monitor	Social summary; numerical claims not used as facts
`2026-08-11-defcon34-one-chain-ai-infra`	MLOps control-plane and supply-chain threat-model lead	Short official deck extraction
`2026-08-11-defcon34-offguard-ai-gateway`	AI gateway, MCP, secret, SSRF, and guardrail threat-model lead	Conference research; product details require current advisories

Exclusion rules

Social-media claims never become settled facts without their primary source.
Repository popularity is not quality or safety evidence.
Author-reported benchmark results remain source claims until reproduced.
Security research contributes defensive architecture and evaluation lessons, not operational attack procedures.
Archived copies are not assumed current.