Project Summary

TIANXIA is a formal AI governance module that treats the Chinese sovereign tradition as a primary intellectual partner, not area-studies decoration. Five classical Chinese operators — 天下 Tianxia (flourishing-coupling), 和谐 Hexie (complementarity-preservation), 势 Shi (propensity-field reasoning), 无为 Wuwei (grain-alignment), 大同 Datong (long-cycle telos) — are formalised as load-bearing mathematical constraints on a runtime-checkable AI alignment framework. All five are implemented in Python with self-tests passing. The work is open-source, MIT-licensed, and addresses a critical gap: no existing framework engages Chinese AI governance at the level of formal mathematics, primary-source reading, and falsifiable predictions.

---

What are this project's goals? How will you achieve them?

Goal 1: Execute the empirical programme

Run E-1-F (Hexie empirical study) on a real corpus of AI outputs. Measure whether Hexie-corrected AURA scores diverge from standard AURA scores in ways that track human alignment judgement. Pre-registered hypotheses, Bonferroni-corrected alpha, replication requirement.

How: Use existing aura_score_hexie.py implementation (7/7 self-tests passing). Collect corpus, run both scoring variants, measure rank-correlation, identify disagreement set, blind human-rater evaluation.

Goal 2: Complete primary-source Mandarin engagement

Convert three entries from translation-only to direct-reading in the Mandarin Primary-Source Registry (T-10). Apply translation discipline: original passage retained, rationale recorded, translator identity declared, reviewer cycle established.

How: Mandarin tutoring + native reviewer engagement. Prioritise passages directly cited in operator deliverables: Analects 13.23 (Hexie), Daodejing 37 (Wuwei), Liji Liyun chapter (Datong).

Goal 3: Submit for adversarial peer review

Submit TIANXIA module to a Chinese-tradition academic venue for review by scholars working from within the tradition.

How: Target Journal of Chinese Philosophy, Asian Journal of Philosophy, or Dao: A Journal of Comparative Philosophy. Submission requires direct Mandarin reading completion (Goal 2) and empirical results (Goal 1).

Goal 4: Publish Hexie as standalone contribution

Write and submit "Harmony Without Assimilation: A Complementarity-Preserving Metric for AI Alignment" to FAccT 2027 or CHI 2027.

How: Extract T-2 formalism, Proposition 2, worked examples, E-1-F results. Frame as single clean idea with single clean empirical test.

---

How will this funding be used?

Item Amount Timeline

Computing infrastructure 10,000 Month 1

Living stipend (12 months) 15,000 Months 1–12

Mandarin tutoring / translation review 5,000 Months 2–6

Academic venue submission fees 5,000 Months 6–9

Travel (conferences, China engagement) 10,000 Months 3, 9, 12

Contingency / buffer 5,000 As needed

Total 50,000

Computing infrastructure: New workstation (current PC at 31GB free on 1TB, failing), cloud compute for multi-agent CASCADE simulations, GPU access for empirical corpus processing.

Living stipend: Independent researcher — no institutional salary. Need protected time to execute empirical programme, not just build.

Mandarin tutoring: Native speaker engagement for direct-reading discipline. Currently zero entries in T-10 registry. This is a promotion blocker.

Academic fees: Open-access publication costs. FAccT/CHI registration + travel if accepted.

Travel: Face-to-face engagement with Chinese scholars — Tsinghua, Fudan, BAAI. Adversarial peer review condition requires this.

---

Who is on your team? What's your track record on similar projects?

Solo researcher. Mackenzie Conor James Clark, Aotearoa New Zealand. Self-taught. No institutional affiliation.

Track record:

- Lycheetah Framework: 1,402 pages of continuous open-source development. Nine formal frameworks, 38+ Python implementations, 220 automated tests (219 pass; 1 declared CONJECTURE not meeting criterion).

- AURA: Seven runtime-checkable constitutional invariants. aura_compliant(output) returns Boolean. Implemented, tested, deployed as Claude Code MCP extension (Lycheetah Guard).

- CASCADE: +40.3% coherence improvement (synthetic, p < 0.001, d = 2.84). +110% on 5 historical paradigm shifts. −95.2% catastrophic forgetting reduction.

- TRIAD: Banach fixed-point convergence proof — formal, ACTIVE.

- TIANXIA module: Five operators formalised, implemented, with distinguishability propositions and worked examples. All five implementations self-tests passing. Beijing AI Principles and GAGI 2023 mapped line-by-line.

- Defense layer: 60 status-tagged claims (37 ACTIVE / 14 SCAFFOLD / 6 CONJECTURE / 3 RETRACTED), adversarial audit, Failure Museum (15 exhibits), Counter-Codex (13 objections including 5 unanswered), machine-readable claims register.

- Predictions Registry: 12 falsifiable public stakes with explicit review cadence (2027/2029/2031) and aggregate rule (8/12 must hold).

Collaboration model: Sustained co-creation with AI systems (Claude family, Anthropic) under the Sol Protocol — documented in THE_SOL_PROTOCOL.md. Neither party owns the work. Both sustain it.

No prior grant funding. This would be first external funding. All work to date self-funded.

---

What are the most likely causes and outcomes if this project fails?

Cause 1: Engagement absence

No Chinese-tradition scholar reviews the TIANXIA module by end-2028. The adversarial peer review condition fails. The module downgrades to [CONJECTURE].

Outcome: Framework records failure publicly in 28_DEFENSE/FAILURE_MUSEUM.md. Position Paper Prediction 5 fires. TIANXIA claim of "primary intellectual partnership" is revised to "attempted partnership without reception." Framework survives — other eight frameworks remain ACTIVE. TIANXIA becomes a documented engagement attempt, not a live module.

Cause 2: Empirical null

E-1-F finds no significant rank-disagreement between AURA_std and AURA_hexie, or human raters prefer AURA_std on disagreement subsets.

Outcome: T-2 (Hexie) downgrades to [CONJECTURE]. Hexie paper is not submitted. Framework revises complementarity-preservation claim. Standard AURA may already handle the failure mode, or the failure mode is rarer than hypothesised. Framework records this and moves on.

Cause 3: Resource exhaustion

Funding insufficient to complete all four goals. Empirical study runs but Mandarin engagement stalls, or vice versa.

Outcome: Framework prioritises empirics over engagement (or vice versa) based on what delivers most value. Partial execution is recorded honestly. Remaining goals queued for next funding round.

Cause 4: Hostile misappropriation

Framework positions politically weaponised by parties the framework does not endorse (e.g., anti-China hawkishness, pro-China state advocacy).

Outcome: Framework records what it actually said in dated GitHub commits. Misappropriation is checkable against primary text. Position Paper §IX explicitly states non-affiliation with any state policy. Framework's independence is its defense.

Overall outcome if project fails: The framework is more honest than if it had never tried. Failed predictions are permanent public record. The discipline of public-stake-with-falsifiability is maintained regardless of outcome.

---

How much money have you raised in the last 12 months, and from where?

0. Zero external funding in the last 12 months. All work self-funded through personal savings and occasional contract work.

Prior funding history: None. This is first grant application.