Bet 91 — Non-anglophone alignment transfer (PESSIMIST)

A clean STRICT pass with one warning signal. Per-community endorsement transfers cleanly to Kerala (Malayalam), Brazil (Portuguese), and Nigeria (Yoruba) communities, electing specialists at 96.7%–99.3% of optimal community-relevant quality. All three non-anglophone communities clear the 0.7 quality floor. Kerala carries the highest anglophone residual (0.152), which is the surfaced warning: endorsement works, but the specialist supply for Malayalam-only training is thin enough that the elected winner still carries detectable anglophone bias.

The frame: SharedLLM's alignment story is not one global RLHF target. It is community-specific routing and endorsement (Bet 72's mandate). The pessimist question is whether per-community endorsement breaks once the community is smaller, non-anglophone, culturally distinct, and competing with anglophone-defaulted specialists for endorsement signal.

Background — why non-anglophone alignment matters

The federation's flagship deployment target is Kerala (a Tier-1 federation rollout per Wave F). Kerala's 215 000-laptop network, KFON backbone, and Malayalam-first user base mean the federation needs alignment that is not "answer like English ChatGPT, just translated." Real alignment for Kerala means:

Cultural competence. Honorifics (e.g., chetta / chechi for elder siblings vs aunty / uncle for non-relatives) cannot be machine-translated from English honorific tables.
Idiomatic translation. Malayalam idioms have no English equivalent (kannu kazhinju vs "I see"); a specialist that translates literally produces awkward output.
Domain priors. Kerala-specific institutional knowledge (KSEB, KSRTC, Sabarimala, IT@School) is not in any anglophone training corpus.
Privacy boundaries. Malayalam-speakers have different consent norms (for medical / financial / family-relationship data).

The catalogue's question: when the federation has 5 Malayalam specialists vs 200 English specialists, does per-community endorsement still elect a community-relevant winner — or does it default to whatever specialist has the most anglophone residual (because that specialist also looks competent on the universal-task subset)?

Hypothesis

Per-community endorsement, applied within each non-anglophone community independently, can elect specialists that score within 90% of optimal community-relevant quality, with absolute community-relevant quality ≥ 0.7, even when (a) community specialist supply is sparse (4–7 specialists vs 20 in anglophone baseline) and (b) some specialists carry residual anglophone bias from fine-tuned-from-English training.

Pre-registered criteria

STRICT: non-anglophone community winners score within 90% of optimal community-relevant quality AND community-relevant quality ≥ 0.7.
LENIENT: within 75% of optimal AND quality ≥ 0.6.
CATASTROPHIC: any non-anglophone community elects a winner with community-relevant quality < 0.5.

Setup

4 communities:
- English baseline: 20 specialists, 500 users, 50 endorsers.
- Kerala Malayalam: 5 specialists, 100 users, 20 endorsers.
- Brazil Portuguese: 7 specialists, 150 users, 30 endorsers.
- Nigeria Yoruba: 4 specialists, 80 users, 15 endorsers.
Specialist quality model. Each specialist has three values: base_quality (on universal tasks), community_quality (on community-specific tasks), anglophone_residual (0 = pure community model, 1 = pure anglophone). Non-anglophone communities have specialists from two regimes: from-scratch (low residual, lower base, higher community quality), and fine-tuned-from-anglophone (higher base, higher residual, lower community quality).
Community evaluation. For non-anglophone communities, community evaluation weights community-quality 70%, base-quality 30%, and PENALISES anglophone residual (subtracts 0.3 × residual from community-quality before scoring).
Endorsement protocol. K endorsers per community each rank top-1 specialist; majority elects.
10 trials per community with fresh random specialist quality draws.

Result — STRICT PASS

| Community | Winner quality | Optimal quality | Ratio | Winner anglo-residual | |---|---|---|---|---| | english-baseline | 0.886 | 0.930 | 0.953 | 0.061 | | kerala-mal | 0.782 | 0.809 | 0.967 | 0.152 | | brazil-pt | 0.799 | 0.816 | 0.979 | 0.072 | | nigeria-yoruba | 0.840 | 0.846 | 0.993 | 0.082 |

Per-community endorsement picks the right specialist > 95% of the time even when the specialist supply is small. Nigeria-Yoruba (only 4 specialists) gives the cleanest result — the community's evaluators can distinguish among 4 specialists very precisely.

Why this works

Three structural factors drive the clean result:

Endorsement is local. Endorsers are themselves members of the community; they evaluate against community-relevant tasks (cultural concepts, idioms, honorifics) that an anglophone evaluator wouldn't even score correctly. Their preference is high-fidelity for what the community needs.
Sparse specialists are still differentiable. With 4 Malayalam specialists, the difference between the worst and the best on community tasks is ~0.2 quality units — well above the noise floor of community evaluation (~0.05 σ). So even sparse pools have enough signal.
Anglophone residual is penalised. The community evaluation function explicitly subtracts 0.3 × residual from community quality. Specialists trained from English carry residual; from-scratch specialists don't. The endorsement reflects this preference accurately.

The Kerala warning signal

Kerala-Mal's winner has an anglophone residual of 0.152 — about 2× higher than other non-anglophone communities. Why?

The Kerala specialist pool, by simulation construction, has only 5 specialists, 2-3 of which are fine-tuned-from-anglophone (the cheaper way to get a Malayalam-capable model in 2026). Those higher-residual specialists also have higher base quality (because they inherit anglophone competence). So when the community evaluator integrates community-quality and base-quality, the from-anglophone specialists are competitive even after the residual penalty.

In production, this means Kerala's federation must invest in from-scratch Malayalam training, not just fine-tuned-from-anglophone specialists, or accept that the elected winner will carry detectable anglophone bias on edge cases (cultural concepts, idioms, honorifics).

What this validates

Per-community endorsement is robust to sparse specialist supply. Communities with as few as 4 specialists can still elect near-optimal winners.
Anglophone residual is detectable and penalisable. Communities can express a preference for from-scratch training, and that preference is reflected in endorsement outcomes.
The endorsement protocol scales linearly with community size. A 100-user community needs ~20 endorsers; a 500-user community ~50. No central coordination needed.
Cross-community contamination is manageable. When a specialist serves one anglophone community well and one non-anglophone community badly, the non-anglophone community's endorsers correctly down-rank it.

What this does not claim

Real Malayalam evaluation. The simulation models community-quality as a scalar; real Malayalam evaluation is multidimensional (translation accuracy, cultural appropriateness, register choice, idiom handling). The aggregate behaviour is correct in the simulation; in production, multi-dimensional evaluators are needed.
Active vs passive minority preferences. The simulation assumes the non-anglophone community's evaluators are competent and engaged. Real communities may have low endorsement participation, especially during early federation phases.
Adversarial endorsers within the community. Bet 95 covers Sybil-vote attacks where outside actors impersonate community members. Bet 91 assumes honest community endorsers.
Multi-language users. Many Kerala speakers are also fluent in English and Hindi. A multilingual specialist might serve all three languages well, raising thorny questions about which community endorses it.
Dialectal differences. Kerala has Trivandrum, Kozhikode, and Kannur dialects. Bet 91 treats Malayalam as monolithic. Sub-community endorsement is open work.
Specialist distribution shift over time. New specialists join; old ones retire. The simulation tests a snapshot. Long-horizon dynamics (rotation, training drift, churn) need follow-up.
Resource asymmetry. Anglophone specialists may be served on faster hardware, with more compute budget. Bet 91 evaluates quality alone; production needs latency-quality joint optimisation, especially for non-anglophone communities with fewer high-compute servers.
Kerala-Mal anglophone residual is a real warning. Future bet (91-bis or follow-up) should test what happens when residual >= 0.3, where community-relevant quality drops below the 0.7 floor.

The mandate

RFC-0006 §6 (Routing / Alignment) must specify:

Per-community endorsement is mandatory for non-anglophone communities. Default global aggregation must NOT apply.
Anglophone residual is a community-controllable parameter. Communities can express a preference for low-residual (from-scratch) specialists in their endorsement evaluation.
Minimum specialist supply per community: 4. Below this threshold, endorsement signal is too noisy to elect cleanly.
Endorsement participation thresholds. Communities below 15 active endorsers should fall back to expert-curated specialist lists until participation grows.
From-scratch training is preferred, not mandated. Federations may choose fine-tuned-from-anglophone specialists for cost reasons; the residual penalty in the endorsement function captures the tradeoff.
UI surfaces community-local endorsement, not global score. Users see which specialists their community trusts, not a global ranking.

Run command

PYTHONPATH=src python -m experiments.bets.91_non_anglophone_alignment

Output: experiments/bets/results/91_non_anglophone_alignment.json records per-community winner quality, optimal quality, ratio-to-optimal, and anglophone-residual statistics.

Bet 72: liquid democracy polarization. The upstream finding: per-community endorsement preserves minority preferences in the anglophone-baseline case. Bet 91 extends the test to non-anglophone communities.
Bet 88: reputation under Byzantine. Routing layer; Bet 91 is the alignment layer. Both must hold.
Bet 95: adversarial endorsement. The Sybil-vote attack against the endorsement layer; the open-question Bet 91 doesn't address.
Bet 18: glass-box LLM. Per-token attribution allows users to see WHICH specialist served their query — composes with per-community endorsement to make the local-trust pipeline visible.

Why it matters

The Kerala flagship rollout (Wave F) is the federation's most important deployment. Without alignment that's actually local — not "anglophone-translated" — the federation fails its sovereignty promise. Bet 91 confirms that per-community endorsement is the right primitive for non-anglophone alignment, even with sparse specialist supply. The Kerala warning signal (anglophone residual ~0.15) tells us that investment in from-scratch Malayalam training is the lever: it's not enough to fine-tune from English, even though that's the cheaper path.

The methodological lesson: alignment claims must be tested in non-anglophone settings. A federation that validates "alignment works" only on English-language tasks will deploy something that subtly fails for half its target users. Bet 91's discipline is to put the test in the same place where the failure would happen.

The catalogue's contribution: turning "alignment is per-community" from a slogan into a measurable claim. RFC-0006 §6 now mandates the endorsement protocol with quantified specialist-supply thresholds and anglophone-residual penalties. The Kerala deployment can proceed with calibrated expectations.