Rigor as Enforced Habits

A bundle that won't let you overclaim

The Frame

LLMs write confident papers, not rigorous ones

It's trivial now to produce an academic-LOOKING paper for anyone who needs a defensible artifact — but the writing is confident, not rigorous: overclaiming, no methodology lockdown. amplifier-bundle-research (v0.9.0, MIT) answers by shipping scientific rigor as enforced habits, not advice.

First question: is this a real system, or a promise?

The Proof

This is a real, worked system — not slide-ware

The rigor is packaged and shipped: a fleet of specialist agents, enforced behaviors, pipeline modes, document recipes and venue knowledge — 32,870 lines of authored discipline in the repo.

So what exactly are all these habits defending against?

16

agent .md files

12

enforced behaviors

6

pipeline modes

20

recipe .yaml files

9

venue formats

32,870

lines of markdown

The Enemy

The failure mode is HARKing

An LLM will happily report a finding it stumbled on as if it had predicted it all along — motivated reasoning dressed up as a result. That's the specific enemy the enforced habits have to defeat.

The first tooth: take away the room to move the goalposts.

Tooth One — Lockdown

The methodology is hash-sealed before any data is seen

/study-plan writes a SHA256-locked preregistration.yaml — predictions, analysis plan, stopping rules — locked and timestamped. Its mode default action is block. It can't be quietly rewritten.

But a lock is only as good as the gate that enforces it.

# modes/plan.md — /study-plan
preregistration.yaml:
  predictions: {confirmatory, exploratory}
  analysis_plan: ...
  stopping_rules: ...
  honest_pivot_clause: true
  locked_sha256: <hash>
  locked_at: <timestamp>
default_action: block

Tooth Two — honest-pivot

On any divergence from the plan, the transition stops

honest-pivot is default-ON and fires at every mode boundary after /plan. When a draft diverges from the locked pre-registration it halts the mode transition — "there is no 'just continue' path." Disabling it needs an explicit --no-honest-pivot flag, logged in the artifact.

One loophole remains: who gets to grade the result?

Keep + label exploratoryFindings not in the plan carry the tag through to the artifact.
Timestamped amendmentDeviations are recorded, not erased.
Explicit disableOnly via a logged --no-honest-pivot flag.

Tooth Three — Cross-vendor critique

A claude-* system can't be graded by a claude-* judge

Same-vendor judging produces a measurable leniency bias: per behaviors/cross-vendor-judge.md, the claude-opus-4-7 judge graded claude-*-family agents 3 to 10 percentage points more leniently than gpt-5.5 did — across three Anthropic agents on the same 30 items. When the system is claude-*, the recipe HALTS rather than let a claude-* judge grade it.

Three teeth in place — now, do they actually catch anything real?

3–10

pp more lenient, same-vendor judge (PSE-PG14 Layer 2)

3

Anthropic agents compared

30

items, held constant

The Payoff

On a real study, the headline finding flipped TWICE

The bundle is dogfooded on a real sleep-consolidation study (porting arXiv:2605.26099, n=192 paired). Under its own honest-critic the headline flipped twice — each flip catching a false causal claim before it could ship.

Same discipline, whoever is making the artifact.

RAG tautology

The retriever was an exact-match oracle — a proxy that guaranteed its own result.

Format confound

Verbatim used sentences; the LLM used triples — not a fair comparison.

Two genuine code bugs

Each correction changed a conclusion. Corrected finding: faithful consolidation ≈ dumb verbatim (delta≈0, p=0.69–1.0).

The Takeaway

Same scaffolding, same honest critic — for whoever needs it

Patent attorney, policy analyst, founder, first-time researcher — everyone gets the same discipline, shaped to the artifact they're making. The rigor a decade of research training buys, now packaged as habits that can't be silently skipped.

Shipped · v0.9.0 · MIT

Sources

Research Methodology

Repository: michaeljabbour/amplifier-bundle-research (v0.9.0, MIT). Re-derived against a fresh clone at /tmp/abr, HEAD 79b64cb (2026-06-08).

Feature status: Shipped — lockdown, honest-pivot, cross-vendor-judge, autonomous experiment loop all present in-repo.

Commands run:

gh repo view michaeljabbour/amplifier-bundle-research --json name,description
git log --oneline | wc -l → 69 commits (2026-04-22 → 2026-06-08, span 47 days)
ls agents/*.md | wc -l → 16; ls behaviors/*.md | wc -l → 12; ls modes/*.md → 6; ls recipes/*.yaml | wc -l → 20
ls context/conference-formats/ context/venue-formats/ → 9 venue formats
find . -name '*.md' -not -path './.git/*' | xargs wc -l | tail -1 → 32,870 lines
cat modes/plan.md, cat behaviors/honest-pivot.md, cat behaviors/cross-vendor-judge.md
bash scripts/experiment-loop/tests/run_all.sh → ALL 10 helper tests PASS
cat docs/case-studies/sleep-consolidation-study.md

Gaps: The 3–10pp same-vendor leniency figure (PSE-PG14 Layer 2) and the sleep-consolidation numbers (n=192, p=0.69–1.0, arXiv:2605.26099) are the bundle's self-reported figures — the in-repo claim text was verified verbatim, but the underlying datasets/harness were not re-run here.

Primary contributors: Michael J. Jabbour (56 commits), Amplifier bots (13 commits).