{"id":31200,"date":"2026-06-12T19:19:17","date_gmt":"2026-06-12T17:19:17","guid":{"rendered":"https:\/\/wordlift.io\/blog\/en\/?p=31200"},"modified":"2026-06-12T19:19:22","modified_gmt":"2026-06-12T17:19:22","slug":"sangue-e-grafi","status":"publish","type":"post","link":"https:\/\/wordlift.io\/blog\/en\/sangue-e-grafi\/","title":{"rendered":"Sangue e Grafi: Teaching a Small Model to Read the Bloodline"},"content":{"rendered":"\n

A few weeks ago at SEO Week I argued that structure is the moat<\/strong>. This month I spent two weekends in the woods<\/em> proving it, literally. The Hugging Face Build Small Hackathon<\/a> has a track called An Adventure in Thousand Token Wood<\/em>, and the rules read like a manifesto we could have written ourselves: small models only<\/strong>, the AI must be load-bearing, joyful is the bar. I built an Italian inheritance mystery. It is called Sangue e Grafi<\/strong> \ud83e\ude78 Blood and Graphs \ud83d\udcca and you can play it at wor.ai\/sangue-e-grafi<\/a>.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

The weekend was mine; the thesis behind it is not. It belongs to a research direction that Chiara Carrozza<\/a><\/strong>, our Director of R&D, and I have been developing at WordLift Lab, ontologies as executable specifications of agent behavior<\/a>, not as documentation. The hackathon was its smallest, most joyful test.<\/p>\n\n\n\n

Here is the game. A patriarch dies. His will says something precise: “my estate goes to my eldest living biological grandchild.<\/em>” The narrative you read, however, is a soap opera. It spends three loving sentences on the daughter-in-law who managed the family finances for decades, who everyone agreed was the backbone of the family, who surely deserved<\/em> the estate. The actual heir gets one sentence: he is twenty-two, and he “also lived in the area.”<\/p>\n\n\n\n

We gave ten of these puzzles to frontier models. Gemini 2.5 Flash scored three out of ten. Gemini 3.5 Flash and 3.1 Pro scored six. And here is the detail that kept me up at night: on several puzzles, all three named the same wrong person<\/em>. Not random errors, a systematic failure. The models found the right family. They fell for the wrong relative.<\/p>\n\n\n\n

Semantic drift is the bug, and it is everywhere<\/h2>\n\n\n\n

Language models are extraordinary at semantic association. That is precisely the trap. A spouse, a caregiver, a child, and an heir all belong to the same family story, and a model trained on stories will follow the character the story makes salient. But inheritance is not decided by salience. It is decided by formal constraints: biological lineage and its direction, age, aliveness, exclusions, the exact clause of the will.<\/p>\n\n\n\n

We call this failure semantic drift<\/strong>: the model follows what is emotionally and semantically related, and silently drops what is formally required.<\/p>\n\n\n\n

If you run AI for an enterprise, you have already met this bug wearing different clothes. The product that is similar<\/em> but not compatible. The claim that is plausible<\/em> but not covered. The customer that is relevant<\/em> but not eligible. The document that is related<\/em> but not authoritative. At WordLift we see these failures weekly in the wild, and they all share one anatomy: the missing ingredient is never more language. It is constraint-preserving reasoning.<\/p>\n\n\n\n

This is the thesis we have been building toward across our research, from RLM-on-KG<\/a>, where a language model navigates a Knowledge Graph autonomously, to SEOcrate<\/a>, our small-model GRPO experiment for SEO reasoning. Sangue e Grafi welds the two together and adds the idea I am most excited about.<\/p>\n\n\n\n

The ontology is the referee<\/h2>\n\n\n\n

Most reinforcement learning on language models uses another language model as the judge. It is expensive, non-deterministic, and with delicious irony, vulnerable to the same semantic drift it is supposed to grade.<\/p>\n\n\n\n

We did something older and, I would argue, more honest. We wrote a small OWL ontology<\/a>, kinship.ttl<\/code>, that encodes the rules of the world: isBiologicalParentOf<\/code> is a directed property; isMarriedTo<\/code> is symmetric and never establishes lineage; a will clause declares which relation types it requires and which it excludes. Then we used those axioms as the reward function itself: turning knowledge graph reasoning from a prompt-engineering hope into a trainable behavior.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Follow a required property through the graph: rewarded. Traverse an excluded one: penalized. Check that a candidate is alive, compare ages explicitly: rewarded. And the crucial design choice, the bonus for a correct answer is gated<\/em>. The model earns it only if it demonstrably walked the graph to get there. A right answer reached by reading the soap opera is worth almost nothing. We are not paying for answers; we are paying for valid reasoning paths.<\/p>\n\n\n\n

The referee is deterministic, auditable, and costs nothing per evaluation. The ontology<\/a> stops being documentation and becomes executable training signal. If you have followed our work on Agent-Oriented Ontology Engineering<\/a>, you will recognize the pattern: this is what it looks like when an ontology graduates from describing a domain to enforcing<\/em> it.<\/p>\n\n\n\n

Six runs, failures included<\/h2>\n\n\n\n

I want to tell you about the training honestly, because the failures taught us more than the wins.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

We post-trained Gemma 4B with LoRA and GRPO on a Modal A100 each run costs about as much as a good dinner in Rome. (Fine-tuning models with structured data is an old habit of ours; we were doing it with GPT-3.5<\/a> when that still sounded exotic.) <\/p>\n\n\n\n