---
title: Leaf Reweight
family: lifetime-objective
summary: Turns the nineteen weights of the reference search's hand-written board-scoring function into run-time data, so that refitted weight vectors can be tested with nothing else about the search changed.
status: Exploratory
evidence: development
generated: false
---

## What this is

A bounded correction around fair D4 that makes the frozen leaf's **nineteen
weights runtime data instead of compile-time constants**, and then tests the
reweightings that
[`finding-10` Addendum A](/docs/exploratory/finding-10-suite-validation) says
predict achievable clears far better than the frozen vector does.

Nothing else changes. The feature extractor is
`drop7::fair_only_horizon::extractFairFeatures`, called by name from the
unmodified reference; the chance stratification, canonicalization, cache keying,
column order, work accounting and terminal utility all come from the frozen code
by way of `../risk-calibration/search.cpp`, which is already proved
decision-identical to the reference. The only degree of freedom is the weight
vector.

<GameTreeFigure seed={0x5eed1007} moves={20} caption="Reweighting changes only the numbers on the outcome boards at the bottom of this tree; the chance branches, their probabilities and the search above them are untouched. That is why a refit can be tested with nothing else about the search changed." />

## The CHECK gate

With the frozen weight vector the local leaf must return values whose raw
`uint64_t` bit patterns are identical to `frozen::fairLeaf`, and the search must
select identical columns and spend identical work. The dot-product accumulation
order is preserved term for term, because floating-point addition is not
associative and bit-identity depends on it.

```sh
./approaches/lifetime-objective/leaf-reweight/build.sh
B=./build/lifetime/leaf-reweight

# 1. leaf bit patterns on real fair-play boards and their chance expansions
$B --leaf-check --depth 3 --chance-samples 7 --max-work 16000000 \
   --seed-start 0xa5278000 --check-games 24 --check-moves 400
$B --leaf-check --depth 4 --chance-samples 7 --max-work 16000000 \
   --seed-start 0xa5278100 --check-games 3 --check-moves 45

# 2. the frozen reference driver, at its own configuration
$B --reference-parity --seed-start 0xa5278200 --check-games 4 --check-moves 60

# 3. weights-as-data vs weights-as-constants, columns AND work, at the
#    configuration the evaluation actually uses
$B --self-parity --depth 4 --chance-samples 5 --max-work 3200000 \
   --seed-start 0xa5278400 --check-games 3 --check-moves 60
$B --self-parity --depth 4 --chance-samples 7 --max-work 16000000 \
   --seed-start 0xa5278300 --check-games 3 --check-moves 22
```

## Setting weights

```sh
# one constant at a time
$B --arm t1-rough --depth 4 --chance-samples 7 --max-work 16000000 \
   --weight roughness=560 --seed-start 0xa5279000 --games 64 --threads 10

# a whole vector from a file of "name value" lines, plus an optional bias
$B --arm t2 --weights approaches/lifetime-objective/leaf-reweight/weights-refit-fair-achievableClears-a0p5.txt ...
```

`--max-work 16000000` is mandatory at seven chance strata. Worst-case depth-4
work there is 11,892,398; leaving the frozen 3,200,000 bound in place silently
degrades the search to a completed depth 3, which has already produced one wrong
conclusion in this programme
([`finding-05`](/docs/exploratory/finding-05-chance-strata)).

## Analysis scripts

| script | what it does |
| --- | --- |
| `refit.py` | fits the leaf's 19 features to the achievable-clear label, rescales the fitted direction into leaf units matched to the frozen leaf's mean and standard deviation, and writes `weights-refit-*.txt` at several frozen↔fitted interpolation weights |
| `tier1.py` | per-term Tier-1 anchors, and the cross-origin transfer table that decides whether the fitted direction is a synthetic-position artifact |
| `sweeptable.py` | tabulates a directory of cohort artifacts against `frozen`, paired by seed |
| `compare.py` | the reported cohort table: quantiles, flow rates, occupancy, paired deltas with a one-sided 95% bootstrap lower bound over whole games, W-T-L, and the lower-tail listing |

## Result

See [`finding-14`](/docs/exploratory/finding-14-leaf-reweight).

## Sources

- `search.cpp`, `build.sh`, `sweep.sh`, `evaluate.sh`
- `refit.py`, `tier1.py`, `sweeptable.py`, `compare.py`