Amplifier · Research Bundle

Prove It, Don't Assume It

Adversarial rigor for AI research

The Frame

AI research reads confident, but you can't check it

amplifier-bundle-research is a real, public Amplifier bundle whose stated purpose is "Superpowers for scientific rigor — defensible written artifacts." It exists to make rigor measurable, not assumed.

And the first proof it's real is that you can count it →

Proof · It Ships

This is real, public scaffolding — not a promise

The bundle (repo michaeljabbour/amplifier-bundle-research, public) is concrete and shipping: version 0.9.0, 69 commits between 2026-04-22 and 2026-06-08.

So the argument rests on something you can inspect. Now the problem it's built to defeat →

16

agent files on disk

12

discipline behaviors

18

non-smoke recipe YAMLs

6

experiment-integrity tool modules

The Trap

Confident prose fills the silence with plausible text

Where evidence is missing, the model doesn't stop — it produces authoritative-sounding prose with no sources, no confidence levels, and no named gaps. You have no way to falsify it.

The bundle answers with two mechanisms. First: bind every claim to a source →

Half One · Sourced Specialists

literature-scout grounds novelty instead of asserting it

The bundle splits work across specialist agents, each a lazy-loading context sink over a K-Dense skill. literature-scout must run an agentic Semantic Scholar loop — not claim novelty from memory.

Sources answer "is it true?" Next: an adversary asks "prove it" →

Formulate a query
Read the top-10 abstracts
Decide: novel or refine
Repeat up to 8 rounds, return a structured novelty verdict

Half Two · Adversarial Scoring

The honest-critic argues AGAINST the work

It issues graded findings with severity levels. Quality is contested, not self-declared — BLOCK items must be resolved before /draft.

But a graded gate is still a judgment. The turn: make judgment a number →

BLOCK
must clear before /draft — e.g. "causal claim unsupported by design"
WARN
flagged for attention
NOTE
recorded, non-blocking

The Turn · Quality Becomes a Number

Quality stops being a vibe and becomes a measured metric

Judgment-based runs use a cross-vendor ensemble of at least 3 judges from different vendor families — because same-vendor judging measured 3–10 percentage points more lenient. Every loop round yields an explicit 5-axis weighted score in [0,1].

Measured, not assumed. Does it actually change answers? →

≥3

cross-vendor judges (different vendor families)

3–10pp

same-vendor leniency bias, measured

5

weighted axes in the satisfaction score

0.50

default closure_target on [0,1]

The Payoff · It Changed the Answer

The honest-critic "earned its keep three times"

In the sleep-consolidation dogfood study, the adversary caught a RAG tautology, a format confound, and two genuine code bugs — and each correction changed a conclusion.

Corrected (n=192 paired): creative reorganize/derive consolidation HURTS (−28pp; fabricates ~50–75% of facts).

Caught a RAG tautology
Caught a format confound
Caught two genuine code bugs
Each correction changed a conclusion (n=192 paired)

The Takeaway

Don't trust confident output — measure it

Ground every claim in sources, score it adversarially, and make the number defensible. Quality becomes measured, provable, and improvable — not assumed.

That's the pattern amplifier-bundle-research turns into code.

Sources

Sources & Research Methodology

Primary source: michaeljabbour/amplifier-bundle-research (public), cloned from git@github.com:michaeljabbour/amplifier-bundle-research.git. Every claim re-derived from actual git / gh / grep / ls / sed / wc output.

Feature status: Released and actively iterated — bundle.md version 0.9.0; 69 commits between 2026-04-22 and 2026-06-08.

Commands run:

gh repo view michaeljabbour/amplifier-bundle-research --json name,description,isPrivate
ls agents/*.md | wc -l → 16 · ls behaviors/*.md | wc -l → 12 · ls recipes/*.yaml | grep -v _smoke | wc -l → 18 · ls -d modules/tool-* | wc -l → 6
sed -n '26,28p' bundle.md (context-sink specialists)
grep -n -i 'semantic scholar|8 round|novelty verdict' agents/literature-scout.md
grep -n -i 'BLOCK|WARN|NOTE|SEVERITY' agents/honest-critic.md
sed -n '93,140p' recipes/autonomous-experiment-promotion-gate.yaml (≥3 cross-vendor judges, Cohen's kappa)
grep -n -i 'leniency|same-vendor' behaviors/cross-vendor-judge.md (3–10pp)
sed -n '1,40p' context/orchestrated-loop-judge-rubric.md (5-axis score, closure_target 0.50)
grep -n -i 'earned its keep|RAG tautology|n=192|-28pp' docs/case-studies/sleep-consolidation-study.md
git log --oneline | wc -l → 69 · grep 'version:' bundle.md → 0.9.0

Gaps / not independently reproduced: The 3–10pp leniency figure, the ~1.5–2 point LLM-judge inflation (n=500), and the sleep-consolidation n=192 / −28pp / ~50–75% figures are quoted verbatim from the repo's own prose; the fact-checker confirmed the repo asserts them but did not re-run the underlying experiments. README lists 14 specialist agents while agents/ holds 16 .md files (adds figure-critic and experiment-runner); both counts appear literally in the repo.

Primary contributor: Michael J. Jabbour — 56 of the commits (Amplifier automation accounts for the remainder).