Information-Theoretic Bootstrap engine for quantum gravity theory-space exclusions.

A localhost research platform that constrains the space of possible quantum gravity theories by simultaneously imposing every well-established consistency condition we can encode — amplitude bootstrap, holographic-information bounds, gravitational universality, anomaly flow, computational complexity bounds — and asking which UV completions survive, how robust they are, and what experiments would tighten the picture.

The engine is at v1.20.0, 351 tests, 33 active constraints across 7 Wilson coefficients, 6 candidate-framework encoders, 27 iteration cycles documented, and a non-empty intersection of all 31 constraints found by numerical search at toy precision. See the full research report.

git clone https://github.com/hassard0/itb-engine
cd itb-engine
python -m venv .venv && .\.venv\Scripts\Activate.ps1
pip install -e ".[dev]"
pytest                              # all 351 tests
itb serve                           # localhost web app
itb check --g4 0.5 --g6 0.4         # CLI feasibility check
itb research-agent --iterations 5   # LLM-powered Dr. M. (needs ANTHROPIC_API_KEY)

# OR run Dr. M. on a local LLM (Gemma 4 / llama.cpp / Ollama / vLLM):
itb research-agent --backend local --iterations 3 \
                   --base-url http://192.168.4.193:8080 \
                   --model gemma-4-26b-a4b-it

The 5-year-old version: three new things the engine found

Imagine you have a bunch of guesses for how gravity works at the tiniest sizes. Each guess is a recipe. We built a robot that knows lots of "rules" — things scientists already figured out that any correct recipe has to follow. The robot tries each guess against all the rules and tells us which ones break.

After running the robot for a long time and looking at what it learned, here are the three most surprising things it told us:

1. The "no-no list" we use to rule out theories was missing the most important rules

For a long time, scientists used two big lists of rules:

• Bouncing rules — about how particles smash together and bounce
• Hugging rules — about how regions of space share information

The robot found that almost none of those rules actually do anything when applied to the popular gravity guesses. They're all easily passed. The rules that actually keep us out of trouble are a different list called the "swampland rules" — rules about which theories are even allowed to exist in a universe with gravity at all.

Implication: The field has been spending a lot of time on the bouncing rules and the hugging rules. The robot is saying: maybe spend more time on the swampland rules. That's where the real fences are.

2. One famous gravity guess fails one specific sharing rule — and only that one

There's a gravity guess called Loop Quantum Gravity (LQG). The robot tested it against many "information sharing" rules from holography. We expected LQG to either pass them all or fail them all.

What actually happened: LQG fails the simplest sharing rule (the one with three regions, called "n=3 monogamy") but passes the harder ones (with four or five regions). The break is very specific — it's not that LQG is broken in general, it's that LQG is incompatible with one exact form of holographic information sharing.

Implication: Critics of LQG have been saying "LQG is non-holographic" without specifying how. The robot has now pointed at a specific equation and said "this one breaks; the others don't." That's a much more precise complaint than "non-holographic," and one a researcher could verify or refute against the actual published forms.

3. The experiment scientists should run next has changed

Before the swampland rules were turned on, the robot ranked experiments and said: "Look at gravitational waves with super precision — measure if they twist as they travel." That was experiment #1.

After the swampland rules were turned on, the robot's ranking flipped:

LIGO gravity-wave twist measurements dropped from #1 to #5. The new top experiments are:

1. Looking at the leftover light from the Big Bang very carefully (CMB-S4)
2. Putting tiny mirrors in two places at once and watching gravity make them choose (Bouwmeester optomechanical experiments)

Implication: If the swampland rules are correct (and they're at least plausible), the field's experimental priorities should reorder. CMB precision and macroscopic optomechanical collapse tests should be weighted above gravitational-wave birefringence updates.

One important caveat the 5-year-old should also hear

The robot is using toy versions of all the rules — close to the real ones in shape, but with simplified numbers. Imagine the robot is using a paper map of a city that has the right streets but the wrong house numbers. The map is good enough to find big patterns ("the swampland district has all the fences") but not good enough to tell you whether your specific friend's house is fenced in. To get exact answers, someone with the real published numbers would need to fix the map.

That's the next step. The robot is ready; the map upgrade is the work.

What the engine does, in one paragraph

Given a parameterized gravitational EFT (Wilson coefficients g_4, g_6, g_8 for matter; g_R², g_R³, g_R²_parity, g_R³_parity for graviton sector), the engine asks whether a candidate theory satisfies every encoded consistency constraint. If yes, it computes how robust the theory is (fragility, signed-distance margins). If no, it reports which physical principle eliminates it. Across 24 constraints spanning amplitude bootstrap (Caron-Huot dispersion bounds, parity-decomposed positivity), information-theoretic (Bekenstein, BNOSSW MMI, holographic subadditivity), and gravitational universality (anomaly inflow, EFT validity, Susskind/Lloyd complexity bound), the engine produces ranked experimental priorities, framework comparisons, and intersection-search results that target where to look next for new physics.

Architecture

itb-engine/
├── src/itb/
│   ├── theory.py                       Wilson-coefficient theory dataclass
│   ├── constraints/                    24 constraint modules (A/B/C classes)
│   ├── frameworks/                     4 candidate-framework encoders
│   ├── engine.py                       Constraint feasibility check
│   ├── mapper.py                       2D parameter sweeps
│   ├── voxel.py                        3D voxel sweeps
│   ├── adversarial.py                  Analytic-center search
│   ├── path_distance.py                Shortest path through allowed region
│   ├── completeness.py                 Allowed-region boundedness check
│   ├── fragility.py                    Distance-to-violation per cell
│   ├── importance.py                   Per-constraint redundancy ranking
│   ├── duality.py                      Cross-class IoU computation
│   ├── phase_components.py             Disconnected-component detection
│   ├── sensitivity.py                  Bayesian feasibility probability
│   ├── fisher.py                       Fisher metric on theory space
│   ├── observables.py                  Observable interface
│   ├── fingerprint.py                  Pairwise framework fingerprint
│   ├── first_disagreement.py           Per-pair best discriminating observable
│   ├── experiment_priority.py          Ranked experiment list
│   ├── intersection_search.py          scipy-driven all-constraint optimum
│   ├── battery.py                      Full-battery markdown report
│   ├── scenarios.py                    Pre-baked scenario variants
│   ├── report.py                       Multi-framework comparison
│   ├── plotting.py                     Plotly figure builders
│   ├── cli.py                          itb command
│   └── api/server.py                   FastAPI web app
├── frontend/                           Plain HTML + Plotly UI
├── tests/                              ~302 tests across all modules
└── docs/
    ├── superpowers/
    │   ├── specs/                      Original design specs (v0.1, v0.2)
    │   ├── plans/                      Implementation plans
    │   └── notes/                      Theoretical research log
    └── results/                        Computed research artifacts (per iteration)

Constraints currently encoded (24)

Class A — Amplitude bootstrap (12)

• scalar_positivity_g4 — Adams-Arkani-Hamed-Dubovsky-Nicolis-Rattazzi 2006
• scalar_positivity_g6 — same family, next order
• scalar_positivity_g8 — Caron-Huot dispersion tower next-next order
• scalar_convexity_g6_vs_g4 — g_6 ≥ g_4², next-order forward dispersion
• dispersion_tower_g6_squared_bound — g_6² ≤ g_4·g_8, chained Cauchy-Schwarz
• graviton_mixed_positivity — Caron-Huot-Mazac-Rastelli-Simmons-Duffin 2021
• cubic_curvature_positivity — g_R³ ≥ 0
• cubic_graviton_matter_bound — g_R³ ≤ κ·g_4²
• parity_violating_positivity — |g_R²|² + |g_R²_parity|² ≤ κ·g_4·g_6
• left_handed_graviton_positivity — polarization-decomposed
• right_handed_graviton_positivity — polarization-decomposed
• parity_violating_cubic_bound — |g_R³|² + |g_R³_parity|² ≤ κ·g_4²
• causality_bound — Adams et al causality / de Rham-Tolley

Class B — Information-theoretic (4)

• bekenstein_tight — g_R²² ≤ ½·g_4·g_6
• holographic_subadditivity — g_4 + g_6 ≥ g_R²
• bnossw_monogamy — g_4·g_6/(g_4+g_6) ≥ g_R²
• ligo_birefringence_bound — |g_R²_parity| ≤ 0.1 (LIGO/Virgo O3)
• ligo_graviton_mass_bound — g_R² ≤ 0.5 (LIGO O3 graviton dispersion)

Class C — Gravitational universality (7)

• eft_validity_box — |g_*| ≤ Λ cutoff
• anomaly_cancellation — g_4·g_6 - c·g_R²² = 0 ± tol
• weak_gravity_conjecture — g_R² ≤ α·√g_4
• generalized_anomaly_inflow — |g_R²_parity|² + 2·|g_R³_parity|² ≤ ρ·g_4·g_R²
• t_hooft_anomaly_matching — cubic/leading parity ratio bounded
• complexity_cutoff — Susskind/Lloyd weighted-L² aggregate bound

Candidate frameworks encoded (4)

LQG-induced fails on bnossw_monogamy (class B), strict-anomaly variants (class C), and complexity_cutoff (class C) — exactly the constraints LQG philosophically rejects (holographic, computational).

Headline result, honestly framed

After 18 iterations of building constraint structure, scipy-Nelder-Mead intersection search across the full 7-dimensional Wilson-coefficient space finds a non-empty common solution to all 24 constraints simultaneously:

g_4         ≈ 0.622   matter self-coupling
g_6         ≈ 0.395   next-order matter
g_8         ≈ 0.359   next-next-order
g_R²        ≈ 0.233   leading curvature coupling
g_R³        ≈ 0.151   cubic curvature
g_R²_parity ≈ 0       parity-conserving (driven to zero)
g_R³_parity ≈ 0       parity-conserving (driven to zero)

with worst-case constraint margin +0.0087. This is not any of the candidate frameworks — it's a new feasible point, parity-conserving, sitting between string-EFT and LQG-induced in coefficient space.

This is toy values across the board. The constraint forms are publication-grade-flavored simplifications; the exact prefactors are O(1) placeholders. The path to a real result goes through replacing each encoding with the literal published form.

See docs/results/2026-05-08-v1.8-honest-synthesis.md for the full reckoning.

Research artifacts (chronological)

• docs/results/2026-05-08-v0.8-baseline-report.md — first end-to-end full-battery analysis
• docs/results/2026-05-08-v1.0-publication-grade-report.md — dispersion tower + WGC + LIGO active
• docs/results/2026-05-08-v1.0-findings.md — what publication-grade encoding changed
• docs/results/2026-05-08-v1.1-bnossw-report.md — LQG fails BNOSSW MMI
• docs/results/2026-05-08-v1.2-cubic-curvature-report.md — cubic curvature pinches frameworks
• docs/results/2026-05-08-v1.3-experimental-priorities.md — first ranked experiment list
• docs/results/2026-05-08-v1.4-parity-violation-report.md — parity sector activated
• docs/results/2026-05-08-v1.4-experimental-priorities.md — GW birefringence becomes top priority
• docs/results/2026-05-08-v1.5-first-disagreement.md — high-s scattering most discriminating
• docs/results/2026-05-08-v1.6-anomaly-flow-report.md — anomaly matching reorders binding diagnostic
• docs/results/2026-05-08-v1.8-intersection-search.md — engine optimum found
• docs/results/2026-05-08-v1.8-honest-synthesis.md — corrected synthesis after 18 iterations

Plus scenario reports under docs/results/scenarios/.

Theoretical exploration logs

• docs/superpowers/specs/2026-05-07-itb-engine-design.md — initial design
• docs/superpowers/notes/2026-05-07-ideas-from-mvp-build.md — 7 research-direction ideas (v0.1 → v0.2)
• docs/superpowers/notes/2026-05-07-theorizing-new-models.md — 10 candidate QG model directions
• docs/superpowers/notes/2026-05-08-v02-learnings-and-new-ideas.md — 5 new ideas post-v0.2
• docs/superpowers/notes/2026-05-08-v04-learnings-and-v05-direction.md — 5 new ideas post-v0.4

Honest limitations

1. Toy values throughout. Every constraint uses simplified forms with O(1) placeholder prefactors. Real Caron-Huot 2024 numerical bounds, real BNOSSW inequalities for n=3 regions, real LIGO O3/O4 sensitivities in proper units would all move the engine optimum.
2. 7-coefficient EFT. Real gravitational EFT has dozens of operators. The architecture supports adding more; the encoding work is the limit.
3. 2D and 3D analyses. Higher-dimensional sweeps are computationally tractable but not yet routine.
4. MMI proxy form. The harmonic-mean BNOSSW form is structurally correct but not the literal published inequalities.

The architecture is research-grade. The encoding effort to make the result research-grade is weeks of literature-aware work, not minutes.

License

MIT. See LICENSE.