How a run unfolds
In the books, first-years who survive to Threshing walk into a valley of unbonded dragons, and the dragons do the choosing. The Dragonkind site recreates that: every choice can end with you incinerated, and the author has said there is a 10% chance of that no matter what you pick.
The bonded dragons
Every bonded dragon we could collect, sorted by colour, then tail, then lineage, then oldest first. Most come from ClumsyBadger's Unofficial Dragon Register form, and the rest from dragonkind.com result pages linked in the megathreads. Rider names and photos are left out. Misspelled lineages are corrected when they are within two letters of a lineage seen at least three times for the same colour. The original spelling is kept in the notes.
Colour × tail
Lineages
Route map
Pick an outcome. Click any box to keep only the runs that passed through it, and click more boxes to narrow further. Hover or tap a box or band for its numbers.
Every dragon: how often the attempt ends in a bond
A run counts here when the player reached that dragon's bonding encounter and the form shows Bonded or Incinerated. Many players die on purpose to reroll, which drags the raw rate down. "Wanted it" keeps only runs where the player said they were after that colour, or any dragon.
The two Greens
First sighting: what the game rolls
Which dragon met the cadet first, by company and direction. Pools come from the community PDF, and cells outside them are counted as data-quality flags.
The rare dragons
Two separate rolls. The meadow Blue can join the Orange and Brown after you turn down the valley Red. The tree's shadow option leads to a Blue and then a Black.
Does the Blue show up in the meadow?
Share of meadow visits where a Blue was attempted, by who logged the run. The game cannot know which dragon a player is chasing, so gaps between these bars come from how people report, not from the game.
Does the shadow show up at the tree?
Share of tree climbs where the shadow option led to a Blue and/or Black.
What changes the Blue's appearance, once reporting bias is held fixed?
Adjusted odds ratios from one logistic model: route choices plus data source and chase goal. Bars that cross 1 mean no detectable effect.
What changes the shadow's appearance?
Same approach for the shadow. With so few shadows, intervals are wide.
Day by day (form submissions)
Bulk imports carry no dates, so they are left out here. Dots are sized by the number of runs, and whiskers show 95% intervals.
Bonding once you reach them
Step-by-step survival inside the encounter (from the community PDF)
Findings
These sentences are generated from the current data and filters, so they update when the data does. "Lead" means a pattern worth testing that is not yet established.
Route table
A route here means every choice up to the hour. Routes are ranked by the lower end of their 95% interval, so a route needs both a high rate and enough runs to rank high. Each route is compared with every other route that could reach the same outcome. "Adjusted" holds data source and chase goal fixed (Mantel–Haenszel). q-values correct for testing many routes at once.
Factor explorer
Each row compares runs with that answer against runs with any other answer to the same question, using Fisher's exact test. These comparisons are unadjusted, so check the adjusted models above before trusting a difference. Shaded rows have q < 0.05 after correcting for every comparison in this table.
Odds ratios
Data quality and bias
Three things worth knowing
Data
Methods
1. From form answers to runs
Each row of the "Threshing Routes" form (and each row pasted in from players' own logs) is one threshing run. The page reads the raw export and codes every row with the same fixed rules, which are visible in engine.js:
- Source. Rows with a real timestamp are form submissions. Rows labelled "Data Import", "Manual Upload" or "Historical import" were bulk-pasted from players' own run logs and are called bulk logs here.
- Cadet encounter. Split into the scream choice (help, investigate and leave, or ignore) and its resolution (killed the bullies, squadmate's revenge, left them to the dragons, tourniquet, left the cadet, took their weapons). The fork in the road only exists after helping, so other branches are coded "(no fork)".
- Encounter branch. A tree run with a Blue or Black recorded took the shadow. A tree run with Greens recorded descended slowly, and one with only a Red descended quickly. A valley run with an Orange, Brown or Blue recorded (or a Red marked "Changed my mind") passed the Red and reached the meadow.
- Outcome. Bonded to the one dragon marked "Bonded!". Otherwise incinerated by the last dragon marked "Incinerated" in encounter order, or died before the hour. "Saved by another" on the Greens plus a bonded Orange or Brown counts as a rescue bond.
- Flags. Rows are flagged, not deleted, when answers contradict each other: several directions or wings, Greens in the valley, two bonds, or a first sighting outside the documented pool. You can drop flagged rows with the checkbox on Deep Dive's route map.
2. Rates and their uncertainty
A rate is k events out of n runs that could have had that event. The denominator matters. The meadow Blue rate divides by runs that reached the meadow, not by all runs. Intervals are 95% Wilson score intervals, which behave well for small n and for rates near 0%:
With 0 events in n runs, the upper end is still informative. For example, 0 shadows in a few hundred bulk-log tree climbs still gives an upper bound near 1%.
3. Comparing choices
For one answer against the rest, the 2×2 table is [a b; c d] = [event, no event] × [this answer, other answers].
Comparisons use the odds ratio, Fisher's exact test and Benjamini–Hochberg q-values. An odds ratio of 2 means the odds of the outcome are twice as high with that answer. q-values control the false-discovery rate. With dozens of comparisons, some p < 0.05 results are expected by chance, so a "real" difference needs q < 0.05.
4. Adjusting for who reports what
Players chasing a rare dragon log runs where they saw a Blue far more often than other players do. Bulk loggers record every run, including dull ones. To separate route effects from reporting effects, the page fits a logistic regression with every route choice plus data source and chase goal:
The ridge penalty keeps estimates finite when an answer has 0 events. It shrinks extreme odds ratios slightly toward 1, which makes the model more cautious. Route rankings use the Mantel–Haenszel odds ratio across strata of data source and chase goal (bulk logs; form runs chasing Blue, Black, or "Blue or Black" on the newer form; everything else) for the same reason.
5. What this data cannot tell you
- It is self-reported and self-selected. Runs worth bragging about are over-represented, and that inflates every rare-dragon rate. Bulk logs are the least selected rows.
- Runs are not independent. One player may log dozens of runs, and the form has no player ID, so intervals are somewhat too narrow.
- The form records mostly stage-level choices. The newer form asks how you reacted to the first dragon, but most riders answer "unsure", and it does not record the answers given inside a bonding encounter, so route claims that depend on those cannot be tested here.
- The form records dragons you attempted. A player who saw the Blue but chose the Orange would not record the Blue, so Blue sightings are slightly undercounted.
- Step-by-step survival inside an encounter comes from the community PDF, which coded the free-text answers by hand. It is shown as published and cannot be recomputed here.
6. Best routes, the valley map and The Odds
The valley map and The Odds come from a solver (solver.js) that treats a run as a decision tree. Your choices are decisions (company, direction, the cadet encounter in the game's order — investigate the scream or keep walking, help your squadmate or walk away, then the bullies' fate or what to do for the wounded cadet — the fork, tree or valley, how to climb down, whether to face the valley Red, and at every bonding encounter whether to attempt the bond or die on purpose). The game's rolls are chance events: which dragon appears first, whether you die at the sighting, in the fight or before the hour, whether the shadow appears and which dragon it is, which dragon waits in the meadow, and how the bonding encounter ends.
Each roll's probability may depend on the choices before it. It is estimated with hierarchical Dirichlet back-off: the estimate for a full history is pulled toward the estimate for a shorter history, which is pulled toward the overall rate. How strongly is chosen from the data for each roll (empirical Bayes), so the solver does not chase noise in small groups. The best route is then found by backward induction, for two goals: the chance of bonding the dragon on this attempt, and, for riders who die on purpose whenever a run goes wrong, the chance that the account eventually ends on that dragon rather than another (maximised by Dinkelbach iteration). Ranges come from 200 posterior draws; "stays best in N%" is how often the recommended choice is still best across those draws.
Encounter outcomes use only genuine attempts: runs where the rider wanted that colour, any dragon, or said they would not seek Malek. The two summary cards on The Odds use the same rule in reverse: runs where the rider died on purpose at an encounter they did not want are left out, so they describe cadets who went into the valley and either bonded or died. The "low-bias" numbers use only bulk logs and form runs from riders who were not chasing a rare dragon.
Missing answers
"Unsure", blank answers, conflicting answers and questions an older version of the form did not ask are treated as missing information, never as a choice. Each comparison uses the runs that answered that question; models use runs that answered every question in them (complete cases). "I've lost count" counts as a later attempt. The only exception is the chase goal used to adjust for reporting bias: runs from before that question existed stay in the models as their own adjustment group, but that group is not shown as a finding. The solver and the valley map ignore missing steps in the same way.