Riders who choose, agents who negotiate

Bordure rests on seven pillars. Four of them shape the transfer market.

The game owner’s brief came down to two stories. They became the acceptance criteria of the whole effort, then named tests.

Before the work began, the game failed the first story. The finding of October 19, on real-2026 at creation: Pogačar signed with FDJ. The utility of an offer was almost entirely salary; between UAE and a continental team, prestige was worth only about 12% of the rider’s market value, and squad quality only entered through it. The offers in the story were not even legal: a mercenary’s offer was capped at 1.5 times the rider’s market value, and Van Rysel Roubaix’s €5M was worth 2.5.

A rider reads an offer on six criteria, each a satisfaction from 0 to 1, computed from public facts. Their priorities, meaning the weight of each criterion, come from their mentality.

The mentality is drawn once, when the rider is created, and never changes. It is random for now: we have hand-set no real rider’s mentality, but a content pack can fix it, for anyone who wants to mod the peloton. The odds depend on age and ambition: progression weighs mostly before 23, security beyond 31, and the taste for winning or leading rises with ambition.

The rider does not tell you they are a “mercenary”. You read a sentence, from one of two sources: their agent, as soon as you have heard their ask; or a scout, whose report is right with a probability that rises with their level, from 55% at level 1 to 95% at level 5. When the scout is wrong, they read one of the other four mentalities at random, through a named draw made at reading time, with nothing stored in the state. On the rider’s page, the scout “confirms” or “contradicts” the agent. For now, the agent always tells the truth.

Before T167, an offer made on a Monday got its answer the following Monday. It was a form. A talk now puts your team face to face with a rider’s agent on any day of the window (all year for a free agent), in back-to-back turns: time does not pass during the talk.

• A tone, on four notches: “very far”, “far”, “one more effort”, “close”. They also name the terms that block (“the salary”, “the length”, “he wants to be leader”) and what no term can buy back (“he doubts the squad’s level”).
• A concession in proportion to your effort. Their ask drops by half of what your offer rose, never below their requirement. Each step is therefore paid at half: the final price lands between their requirement and their opening ask, depending on how you negotiate.
• Nothing for repeating. An offer that has not risen earns no concession and costs two turns of patience.
• An insult has a threshold. An offer whose salary alone does not reach three quarters of their requirement costs two turns. Two insults: they walk out, until the window closes.
• They go away to think, or walk out. Once patience runs out, if the turn’s offer is within 3% of their requirement, they go away to think: the following Monday they come back with an agreement, a last ask valid for a week, or a refusal. Otherwise they walk out: no more talks with that team for four weeks.

T167’s first version reasoned in utility: a 4% margin on the agent’s utility was worth about 8% in salary for a mercenary and 27% for a competitor, depending on how much salary weighs in their mentality. The same negotiating rule therefore produced very unequal prices. T167 now reads everything in salary: the requirement is the salary that, on the terms of their ask, brings the rider’s utility up to their requirement level, and the agent asks 10% above it. The owner settled the premium: at 4%, negotiating well only won 4%, too little for the gesture to matter.

Patience, for its part, depends only on ambition, which is public: a patience that followed the mentality would give it away.

An agreement is a firm offer: the budget, the place and the status promise are reserved until the rider answers, and the team can neither withdraw nor lower it. But the rider does not sign at the agreement. They decide on market day, the Monday, among all the agreements received.

The advance stops follow the same logic: the game no longer stops “every Monday for lack of an answer”, only when a rider decision concerns you (agreement, wait, an agent coming back after thinking, last ask, refusal, signing) or when your staff’s plan has changed. The stop names the decision.

Before T168, a team ranked candidates by a single value, the expected aptitude on the terrains that matter to it. Nothing told a star from squad depth, or a young rider from a lieutenant, and up to 16 teams went after the same rider on the first Monday. The recruitment plan now ranks candidates by role, in the order the owner asked for.

• A role only keeps what the team would pay. The first version went after stars without looking at the price: 26 teams out of 36 targeted a star on the first Monday. A role now keeps a rider only if their estimated price fits under the role’s ceiling, and squad depth does the same. That is what nearly tripled the pair conclusion rate.
• The focus is live. It no longer comes from the content: it is the draft’s formula applied to the day’s active squad, on public ratings, “anchored” so that it only changes if the squad clearly points elsewhere. A declared focus (content, founded team) wins. In the design, Red Bull came out as “time trial” on real-2026 because of Evenepoel: its leader role was therefore looking for a rouleur. This answers the limit our draft post noted: the focus now serves the market.
• The AI gives up instead of clinging on. An AI negotiates through the same function as your staff: it opens at 90% of the ask brought to the shape of its offer, reads from the tone what would be needed, offers its ceiling once, then gives up when it can no longer go higher. Leaving the table and coming back the following week kept places for it on riders out of reach.

T169 puts the talk on screen. The mock-ups were validated before the code; the screen as shipped departs from them in a few points, recorded in the ticket’s report.

The talk commands (offer, accept their ask, leave the table, improve an agreement, delegate) are stored with their day, like programme edits: resuming in the middle of a talk, with an agreement pending, gives back the same world. All the text is in French and English, following the project glossary.

Every ticket has its bench: lab career world plays complete careers, on the default pack with fictional races, every team run by its staff, over 8 seeds and 20 seasons. Targets are written before the measurement, and the lines that block are named in advance. The world’s stationarity (the best quality does not drift by more than 2 points in ten years, no squad outside its bounds, no team beyond its budget, no rider created in an emergency in steady state) holds at every step.

The cost: a simulated season takes 9% longer (11.9 s against 10.9 s), because the market panel, which the Recruitment screen re-reads after every command, computes the plan and the world ranking: it goes from 9 to 47 ms. We did not need to store the ranking in the state to keep it.

A transfer market is a compromise between rich teams and ambitious riders, not a truth. Here is what we know we do not yet do well.

• Pair conclusion is still under its band. 27.4% rather than 40 to 80%, and no ticket has met it since T47a. A team still hears asks higher than its estimated price, because the rider’s true mentality and their agreements elsewhere escape it.
• Mercenaries do not sign the best-paid offer often enough (63% against 70% aimed for). A rider almost never waits: the agreement is taken at their ask, above the waiting threshold. The AIs improved an agreement only 17 times in twenty seasons.
• Neo-pros from the talent pool stay negative. The plan’s young rider is chosen on their public projection, which says nothing about their margin; only a scout reads it.
• The mentality is drawn at random, and the agent always tells the truth. An agent who lies, or a fooled scout, will come later.
• The levers are not there. The last season is capped at three seasons, and the settings on screen are greyed-out boxes. On T170’s programme: length up to five seasons, the signing bonus (paid up front from the current season’s budget), the guaranteed programme (“at the start of the Tour de France”), leader exclusivity, then result bonuses and a season option.
• No paid transfers. No release clause, no buying out a rider under contract: the game has no transfer fee, and we ruled it out at design time. Bundled contracts and agent characters too: the agent stays the rider’s voice.
• Old saves do not resume. The market changed model with every ticket; a career begun before is refused, not migrated.