Skip to content

abdul-abdi/ai-brain-claims

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Ten Claims at the Frontier

a working notebook · ten dossiers · one primitive that survived

0 vindicated.  0 cleanly refuted.  10 in CONTESTED or SPLIT.
strong forms systematically fail.   weak forms systematically hold.

Live notebook ↗ · Paper ↗ · How it was made ↗ · Roundtable ↗ · Observatory ↗


A single autonomous session. 21 parallel research dispatches. Ten ambitious hypotheses at the seam between AI, context engineering, and the brain — strong enough to fail. The agents argued through nine analytical lenses (Karpathy, Joscha Bach, Hickey, Carmack, Bret Victor, Bryan Cantrill, Ayanokōji, pg, Taleb), staged a four-persona roundtable that killed the proposed product, then implemented and benchmarked the architecture that survived.

Nothing was hand-edited for the verdict.

The headline

The interesting object isn't the table of results — it's the shape of the disagreement.

What it landed on
0 / 10 strong forms vindicated
0 / 10 strong forms cleanly refuted
10 / 10 weak forms survived
5 / 5 / 5 five recurring failure modes · five threads · five engineering recommendations
2.4× retention lift over naive truncation, no privileged role info
100% replay determinism across all 50 (strategy × seed) pairs

What's in the box

  • Ten ~2,500-word PhD-level dossiers, one per hypothesis. Each: 4–6 supporting and 4–6 contradicting primary sources, two persona lenses, an explicit steelman, a verdict, and 5–10 papers to read — every dossier is committed as MDX at site/src/content/claims/.
  • A 4-persona roundtable that killed the proposed product. pg, Carmack, Taleb, Hickey — full persona-skill mode, two rounds with forced cross-examination. Hickey moved. The v1 that shipped is the position he moved to.
  • An observatory primitive, in Python: an immutable event log with importance and confidence as pure functions over the log. 22 passing pytest cases. The architecture that survived everything above.
  • A multi-seed benchmark — 10 seeds × 5 strategies × 3 replays = 150 measurements — embedded as a live, scrubbable figure on the Observatory page. Real numbers, real error bars, deterministic seeds.
  • A 25-paper curated reading list with per-citation verification status. Six load-bearing arXiv papers were re-fetched on 2026-05-10; two had attribution corrections (the Granier & Senn 2025 paper title was wrong; "Persona Vectors" is by Chen et al., not Anthropic-wholesale). Every correction is recorded in place.
  • A live notebook, with handles. Drag a numeral in the hero, the page rearranges. Hover a verdict cell, the persona-pair graph and the scatter answer. Reading is doing.

On the word "persona"

These are analytical lenses, not endorsements. Each is a curated profile — published positions, pet questions, characteristic objections, vocabulary — that an agent loads when it argues from that lens. "Think rigorously" is vague; "think the way Hickey thinks about state" is sharp. For every claim, two lenses were chosen for expected disagreement — surfacing different blind spots.

The deployment was uneven and the agents page is honest about it: Joscha Bach 6×, Karpathy 5×, Hickey 5×, Carmack 4×, the rest 1–2×. Only the eight roundtable agents and one design-pass loaded the full persona skills via the skill registry. The other twelve had prompt-level primers — research workers shaped by personas, not personas themselves.

Reproduce the benchmark

git clone https://github.com/abdul-abdi/ai-brain-claims
cd ai-brain-claims/observatory
pip install -e ".[dev]"
pytest -q                                          # 22 passing

cd ..
PYTHONPATH=observatory/src \
    python -m eval.benchmarks.needle_retention --seeds 10

Byte-identical output, deterministic seeds:

strategy                      mean    ± std   replay
-------------------------  -------  -------  -------
truncation                   27.5%  ± 11.5%     100%
recency                      27.5%  ± 11.5%     100%
recency+role                 21.2%  ± 11.9%     100%
task_relevance               65.0%  ± 15.4%     100%
needle_aware (oracle)       100.0%  ±  0.0%     100%

The task_relevance scorer — which has no privileged knowledge of the "needle" role — gets 2.4× the retention of naive truncation. The composite recency+role scorer falls below truncation because the default role-priority table doesn't include "needle" and so actively deprioritizes it. Imperfect importance signals don't fail neutrally — they fail toward the wrong answer. The pure-function architecture surfaces this cleanly: swap the scorer, replay the log, diff.

Use the observatory

from observatory import EventLog, view, importance, confidence

log = EventLog()
log.append({"role": "user", "content": "..."})
log.append({"role": "assistant", "content": "..."})
log.append({"role": "tool", "name": "search", "content": "..."})

# Working memory is a derived view, not stored state.
working = view(log, scorer=importance.recency_attention(), window=4096)

# Track retrieval and generation confidence separately —
# the dominant RAG failure mode is overconfidence on wrong context.
sig = confidence.dissociation(retrieval_score=0.92, generation_score=0.41)
if sig.diverged():
    log.append({"role": "verifier", "action": "re-query", "trigger": sig})

# A/B test scorers against the same log — replay-deterministic.
alt = view(log, scorer=importance.task_relevance(query="..."), window=4096)
delta = compare(working, alt)
pip install claim-observatory

Repository layout

ai-brain-claims/
├── site/           Astro + MDX + React-island notebook  →  abdul-abdi.github.io/ai-brain-claims/
├── observatory/    Python lib (MIT) — the architectural primitive that survived the roundtable
├── eval/           benchmark scripts + result JSON (consumed by the live figures on the site)
└── data/           prompts, transcripts, verification outputs, pipeline manifest — every number traces here

Read in 10 / 60 / 180 minutes

time route
10 min Synthesis — the convergent pattern, five failure modes, eight recommendations
1 hour Synthesis + the 3–4 dossiers most relevant to your work + the paper
3 hours All ten dossiers in order + the reading list + the observatory

A note on what survives

Every dossier ends with what would change its verdict. The benchmark is reproducible from a single command. The reading list flags which citations were re-fetched and which were corrected.

We expect the verdicts to age. The repository is the unit of revision. Disagree with a verdict — file an issue, there's a template. Open a PR with a new importance scorer, a citation correction, or an additional persona lens.

License & cite

MIT throughout (research, code, prose).

@misc{abdi-ai-brain-claims-2026,
  author       = {Abdi, Abdullahi},
  title        = {10 Claims at the Frontier: Adversarial Validation at the
                  AI \(\leftrightarrow\) Context \(\leftrightarrow\) Brain Boundary,
                  with a Measured Architectural Primitive},
  year         = {2026},
  howpublished = {GitHub repository},
  url          = {https://github.com/abdul-abdi/ai-brain-claims},
  note         = {v0.1 ship 2026-05-09.}
}

by Abdullahi Abdi, Nethermind · drafted by autonomous agents · curated, not authored

About

Adversarial validation of 10 hypotheses at the AI ↔ context ↔ brain frontier — research site + observatory primitive (Python). 0/10 vindicated, 0/10 cleanly refuted; all in CONTESTED or SPLIT.

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages