Idea
Code review from facts about the compiled program
Which inputs reach a result, which fields go unread, and what changed after an edit? Yichus answers from the compiled Bosatsu program and checks a reviewer’s claims against those answers. You can run the checks yourself.
Code & run Change a price calculation
Worked example: four facts from a pricing module
A small demos/pricing module with planted problems:
struct Order(id: Int, qty: Int, discount_code: Int)
def quote(o: Order) -> Int:
_ = o # the order is thrown away...
42 # ...and the "computed" price is hard-coded
def shipping_fee(subtotal: Int, region: Int) -> Int:
match cmp_Int(region, 0):
case GT: add(subtotal, 5)
case _: add(subtotal, 5) # both branches identical: region decides nothing
def surge_bonus(x: Int) -> Int:
match cmp_Int(x, 1000):
case GT: x # looks input-dependent...
case _: 7 # ...the sampled inputs all take this branch
What the tool says about it (the complete captured agenda):
quote:
“takes arguments, none of whose data reaches the produced value; every
value the result can take originates in literals”
The price is fake. The tool offers no opinion; the fact is checkable.
Order.discount_code:
“no compiled extraction of the field exists in the supplied Bosatsu IR (reads: []); no declared API wire consumer was established; other engine or host consumers are outside this evidence; extraction absence alone does not justify removing the field”
The supplied Bosatsu never extracts the discount code. Check engine and host consumers before removing a field.
shipping_fee(region):
“the parameter only selects among branches; its data never reaches the
produced value”
Region picks a branch, but both branches compute the same thing.
surge_bonus:
“across 8 seeded runs (seed 6321458022832390279),
Demo/Pricing/surge_bonus always produced 7; bounded observation, not
proof; parameter(s) x statically reach the produced value and no
literal-result fact applies — the constancy is not statically
explained”
This static analysis does not establish the constancy, so seeded runs provide additional evidence.
None of them says “bug”; you decide which ones are. The run-based fact is a bounded observation of seeded runs, not proof.
A wrong review fails like a failing test
Three claims about the module, checked (complete report):
Benchmark: fact-tool review versus source review on planted defects
On a 46,922-line codebase with 26 planted defects (round 008), the fact-using AI reviewer found 25 of 26 and the source-reading reviewer found 23 of 26, both at about 100k tokens. With the caps removed, the source reader found all 26 at 1.85× the tokens. All the benchmark data, including the losses →
Run the pricing checks
Needs Node 20+ and jq; the pricing module is served by this site:
# fetch the module and the three claims from this site
curl -O https://yich.us/demos/pricing/pricing.bosatsu
curl -O https://yich.us/demos/pricing/report.bosatsu
curl -O https://yich.us/evidence/pricing/claims.json
# up to 500 fact cards
echo '{"verb":"agenda","limit":500}' | npx yichus@0.3.9 call api_explore \
--source pricing.bosatsu --source report.bosatsu --args -
# one fact type
echo '{"verb":"agenda","limit":500,"card_type":"dead-field"}' | npx yichus@0.3.9 call api_explore \
--source pricing.bosatsu --source report.bosatsu --args -
# check the three claims; --expect turns a refuted claim into exit code 1
printf '{"claims": %s}' "$(cat claims.json)" > claims-args.json
npx yichus@0.3.9 call api_verify_claims --source pricing.bosatsu \
--source report.bosatsu --args claims-args.json --expect /refuted=false
# fact-diff wants a "before": copy the module, then edit
# old/pricing.bosatsu to give quote a real computation.
# Compare that edited version with the shipped constant version:
mkdir -p old && cp pricing.bosatsu old/
jq -n --rawfile before old/pricing.bosatsu --rawfile after pricing.bosatsu \
'{before_sources: [{fileName: "pricing.bosatsu", source: $before}],
after_sources: [{fileName: "pricing.bosatsu", source: $after}]}' > diff-args.json
npx yichus@0.3.9 call api_fact_diff --args diff-args.json
Explore the facts live in the playground →
Early UI fixtures and experiments live in the archive.
Ask your agent about anything here; it has the detail.