Also written asdreampathrolloutimagined futureensemble
The model that, from a moment in a real replay, imagines 32 possible versions of the next 60 seconds.
What it is
It is E7 used as a generator. Given everything one team could see up to a moment in a replay (fog of war included), it writes out what could happen next, five seconds at a time, twelve steps, and does this 32 times. Each version is a “dream”. Counting across the 32 dreams gives probabilities, for example “this hero dies within 60 s in 9 of 32”.
flowchart LR
R[Real replay up to the moment] --> E[E7 dreamer]
E -->|5 s step x 12| D1[Dream 1]
E --> D2[Dream 2]
E --> D3[... Dream 32]
D1 --> C[Count across the 32 dreams]
D2 --> C
D3 --> C
C --> P[Chance: this hero dies within 60 s, e.g. 9 of 32]
From one moment, 32 imagined futures, counted into a chance.
Example from this project
From minute 10 of a test game, 32 dreams take about 15 minutes on the VPS CPU. The coach tools read them through tools/dreamer_coach.py.
Why it matters here
The coach is meant to use these imagined futures as evidence: “if you stay here, you die in about a third of the futures”.
2026-11-03 11:42Claude (Opus): added a pele figure: From one moment, 32 imagined futures, counted into a chance.
2026-11-03 11:40import from concepts.md: added as new
NewDota projectAdded 2026-11-03
moment
Also written asprefixboundarycut
The point in a replay where the dreamer starts; it sees only what came before (the notes call it a “prefix”).
What it is
The replay is cut at a time divisible by five seconds. Everything before the cut is the input. Hidden enemies stay hidden: their position and HP are blank, as they would be for the player.
Example from this project
--doc 0 --at 60 means game 0, cut at 60 seconds. The final test used 240 moments from the first 20 minutes of 24 games.
Why it matters here
The coach will always speak about a moment. What the dreamer could see at that moment limits what it can know.
2026-11-03 11:40import from concepts.md: added as new
NewDota projectAdded 2026-11-03
anchored risk
Also written asboundary-anchored riskpersistent riskanchored at the boundary
The dreamer fixes each hero’s death risk for the whole 60 seconds at the starting moment; later imagined actions do not change it.
What it is
When the dreamer starts, it decides how likely each hero is to die in the next 60 seconds and when. What happens later inside the dream, including a different move we impose later on, does not update that risk.
Example from this project
Forcing a different move for the first 5 seconds does change the outcome. A different plan for seconds 30 to 60 would barely change the death risk, because the risk was fixed at the start.
Why it matters here
The dreamer can compare one immediate move, but it cannot yet judge a plan made of several moves. That is an open problem before a move-advising coach.
2026-11-03 11:40import from concepts.md: added as new
NewDota projectAdded 2026-11-03
sensitivity is not benefit
Also written asaction sensitivitycausal benefit
A forecast that changes when we change a move shows the model reacts to the move, not that the move is better.
What it is
The comparison tool runs the same moment twice, once with a different 5-second move for one of our heroes, and reports the difference. That difference is what the model believes, not proof about real games.
Example from this project
In the final test, changing one hero’s move target changed that hero’s next position in 15 of 32dreams. That shows the model listens to the order. It does not show the new order was a good idea.
Why it matters here
A coach that says “do X instead” needs evidence that X wins more often in real games, not only that the model’s forecast moves.
2026-11-03 11:40import from concepts.md: added as new
NewDota projectAdded 2026-11-03
hard contradiction
Also written ashard conflictimpossibility
A dream that breaks the game’s rules, such as a dead hero acting or a tower dying twice.
What it is
The checks read every dream and flag impossible sequences: actions by a dead hero, a building destroyed twice, a respawn without a death, and similar.
Example from this project
The final test had 7,680 dreams and zero detected hard contradictions. One dream ended early because the game ended in it.
Why it matters here
A coach cannot use a future that could not happen. Zero means none of the known kinds were found; the notes say it does not rule out every possible impossibility.
2026-11-03 11:40import from concepts.md: added as new
NewMachine learningAdded 2026-11-03
log loss
Also written asloglossNLLnegative log-likelihood
A score for probability forecasts that punishes confident wrong answers hard. Lower is better.
What it is
For each forecast, take the probability the model gave to what actually happened, and take minus its logarithm. Average over all forecasts. Saying 90% for something that happens costs little (0.11). Saying 10% for something that happens costs a lot (2.3). Always saying 50% costs 0.69.
Example from this project
On the final 24 test games, the dreamer’s 60-second death forecasts for our own heroes scored 0.4751; the existing direct death predictor scored 0.4768.
Why it matters here
It is how we decide whether the dreamer’s probabilities can be trusted as much as the predictor we already had.
2026-11-03 11:40import from concepts.md: added as new
NewMachine learningAdded 2026-11-03
not worse by more than a margin
Also written asnon-inferioritynon-inferiority marginupper95
A test that a new model is at most a small, agreed amount worse than the old one (the notes say “non-inferiority”).
What it is
We do not try to prove the new model is better, only that it is not meaningfully worse. We agree the margin first (here 0.01 of log loss). Then we compute a range that, with 95% confidence, contains the true difference. If even the worst end of that range is inside the margin, the test passes.
Example from this project
The dreamer’s difference from the old predictor was between -0.0127 (better) and +0.0087 (worse). The worst case, +0.0087, is below 0.01, so it passed. The range also includes zero, so it is not shown to be better.
Why it matters here
It is the standard the dreamer passed. “Passed” means “about as good”, not “better”.
2026-11-03 11:40import from concepts.md: added as new
NewMachine learningAdded 2026-11-03
even-slice sampling
Also written asstratified samplingstratifiedkeyed-stratified
Choosing the random draws for the 32 dreams from evenly spread slices, so a few dreams give a steadier estimate (the notes say “stratified”).
What it is
With 32 fully random dreams, by bad luck all of them can miss a rare death. Instead, each group of 16 dreams takes its random numbers from 16 even slices of the range, one per slice, so both rare and common outcomes get their share. The cost: dreams in the same group are no longer independent.
Example from this project
The 32 dreams are two groups of 16. The two groups are independent of each other; the dreams inside a group are not.
Why it matters here
Simple error bars that treat the 32 dreams as 32 independent coin flips are wrong for this dreamer. The notes forbid them.
2026-11-03 11:40import from concepts.md: added as new
NewMachine learningAdded 2026-11-03
exposed games
Also written asexposed cohortexposed acceptance cohort
Test games we have already looked at; they cannot be used again to choose or tune a model.
What it is
Once we have seen the results on a set of test games, every later choice we make could quietly fit those games. So they are spent: future decisions need games nobody has looked at.
Example from this project
The 24 games of the final dreamer test are now exposed, and so are the 24 of the earlier failed test.
Why it matters here
The next real test of anything built on the dreamer needs new, untouched games.
2026-11-03 11:40import from concepts.md: added as new
KnownWorking with agentsAdded 2026-11-03
stop list
Also written asstop rulestop condition
The results, written down before a long run starts, that make the agent stop and show Baro.
What it is
Before a long piece of work, the agent and Baro agree on the steps, where the work ends, the cost limit, and a short list of results that would make the agent stop early. Anything not on the list, the agent handles itself.
Example from this project
“If the new model is worse than E7 on the hardest 10% of games, stop and show me” is a stop-list item. “The first run crashed on a missing folder” is not: the agent fixes it and goes on.
Why it matters here
It lets an agent work for hours without waiting for an answer, while the decisions that matter still go to Baro.
2026-11-03 11:40import from concepts.md: added as known
KnownWorking with agentsAdded 2026-11-03
decide before and after, not during
Also written asforks rounddecision list
Baro’s decisions happen before a run (expected forks answered in advance) and after it (a list of what the agent decided), not while it runs.
What it is
Before Baro leaves, the agent lists the decision points it expects, each with the options and its recommendation, and Baro answers them. During the run the agent decides only things that can be undone, and writes each one down. Afterwards, the report lists every decision the agent took: what it chose, the other option, the evidence, and how hard it is to undo. Baro overturns what he disagrees with.
Example from this project
Fork answered in advance: “if the bigger model costs more than twice as much per step, stay with the small one.” Decision taken during the night and listed in the morning: “I used batch size 8 instead of 16 because 16 ran out of memory; to undo, rerun step 3 on a bigger GPU.”
Why it matters here
It reconciles overnight autonomy with Baro staying the engineering lead, and the morning list is where he actually thinks about the work instead of saying one big “go ahead”.
2026-11-03 11:40import from concepts.md: added as known
KnownWorking with agentsAdded 2026-11-03
can-be-undone line
Also written asreversible decisionone-way doortwo-way door
The line between decisions an agent may take alone (cheap to undo) and decisions that stay with Baro.
What it is
Baro decides: the model’s design and goal, how results are judged, the data format, spending beyond the agreed limit, and deleting anything. The agent decides, and writes down: settings, retries, analysis choices, the order of roadmap steps (announced on the dashboard), and which helper model does what. When a run hits one of Baro’s decisions and he is away, the agent parks that branch, does other work, and never sits waiting.
Example from this project
Retrying a crashed run on another machine: the agent’s call. Changing the test so a model “passes”: always Baro’s, and never during an unwatched run.
Why it matters here
Overturning a decision that can be undone costs little, so the agent can move fast there.
2026-11-03 11:40import from concepts.md: added as known
KnownWorking with agentsAdded 2026-11-03
blind test
Also written asblinded comparisonblind judge
Comparing versions while the judge does not know which version produced which result.
What it is
The outputs are shuffled and renamed (R01, R02, …) and the key is kept in a separate file. The judge scores them, and only then is the key used to see which version won.
Example from this project
On 2026-11-03, 45 replies from five models were scored by an Opus judge that did not know which instruction file each reply came from.
Why it matters here
A judge that knows which version is “ours” tends to prefer it, even without meaning to.
2026-11-03 11:40import from concepts.md: added as known
KnownWorking with agentsAdded 2026-11-03
literal rule following
Also written asover-applying rules
Agents carry out a written rule every time, whether or not it is needed in that situation.
Example from this project
A rule asking for “1-2 ideas worth looking up” produced a reading list in every one of the 15 runs that had it, across five different models. A rule about introducing new ideas made Gemini add a “new concepts” section.
Why it matters here
Every line in an instruction file changes every reply. Write rules only for failures that really repeat, keep them few, and describe the reader’s need rather than an action to perform.