Skip to content
Merged
Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
47 commits
Select commit Hold shift + click to select a range
ed05b4f
feat(goal): reconstruct a coding run from FSD state alone (LAB-135)
claude Aug 18, 2026
50952ad
docs(goal): the verdict row for the committed tree, and the shell cou…
claude Aug 18, 2026
03b33ff
chore(goal): parenthesise a precedence expression and flatten the 404…
claude Aug 18, 2026
846f409
docs(goal): the sixth consecutive PASS (LAB-135)
claude Aug 18, 2026
49c6f36
fix(goal): four ways the proof could pass on a wrong account (LAB-135)
claude Aug 18, 2026
825c7bb
fix(goal): a tie is not an order, and a reused workstream is not an a…
claude Aug 18, 2026
2960395
fix(goal): an exemption must cover the case it excuses (LAB-135)
claude Aug 18, 2026
64e903b
fix(goal): make a pooled value unreachable from a per-run judgement (…
claude Aug 18, 2026
9394b5b
docs(goal): the verdict row, on the commit it ran against (LAB-135)
claude Aug 18, 2026
47b2366
fix(goal): a check that failed red on truth and green on the defect (…
claude Aug 18, 2026
e901b17
docs(goal): the verdict row, on the commit it ran against (LAB-135)
claude Aug 18, 2026
c223c2f
fix(goal): finish three half-applied rules and make the halving unwri…
claude Aug 18, 2026
ef61657
docs(goal): the final verdict row, on the commit it ran against (LAB-…
claude Aug 18, 2026
04356d9
fix(goal): widen an ordering claim to the set it names, and close the…
claude Aug 18, 2026
b2f67cb
docs(goal): the round-7 verdict row, on the commit it ran against (LA…
claude Aug 18, 2026
d597362
fix(goal): stop rejecting a run that edits a file it wrote, and stop …
claude Aug 18, 2026
5cd694f
docs(goal): the round-8 verdict row, on the commit it ran against (LA…
claude Aug 18, 2026
a621de4
fix(goal): delete a requirement no assertion reads, and name three un…
claude Aug 18, 2026
0f3f30a
docs(goal): the round-9 verdict row, on the commit it ran against (LA…
claude Aug 18, 2026
08f510e
fix(goal): stop rejecting two attempts at one unkeyable path (LAB-135)
claude Aug 18, 2026
79324ce
docs(goal): the round-10 verdict row, on the commit it ran against (L…
claude Aug 18, 2026
8e40279
docs(goal): correct a claim the next round falsified, in place (LAB-135)
claude Aug 18, 2026
33ea117
docs(goal): A3 and A7 are structurally satisfied on the real path (LA…
claude Aug 18, 2026
c5a33e3
test(goal): execute the A3 and A7 disclosures instead of remembering …
claude Aug 18, 2026
91297e3
docs(goal): name two P2s and raise the ambiguity class to six (LAB-135)
claude Aug 18, 2026
67c1137
docs(goal): correct the stopping line's framing and the limits miscou…
claude Aug 18, 2026
041ebf2
fix(goal): put A8's precomputed failure in the pre-dispatch gate (LAB…
claude Aug 18, 2026
c78f95c
docs(goal): verdict at 88a1bcb08, and the A8 gate's other direction (…
claude Aug 18, 2026
3a89ffb
fix(goal): branch A7's disclosure on the cursor count it computes (LA…
claude Aug 18, 2026
986a9dc
docs(goal): verdict at 74d628367 (LAB-135)
claude Aug 18, 2026
223d8ed
docs(goal): round 11 produced six findings, not four (LAB-135)
claude Aug 18, 2026
801afd4
fix(goal): reconcile mutations against gaps globally (LAB-135)
claude Aug 18, 2026
7df09ca
docs(goal): verdict at fbcc6eb3f, thirtieth consecutive PASS (LAB-135)
claude Aug 18, 2026
cc18372
fix(goal): grade harness ground truth, and stop rejecting ties nothin…
claude Aug 18, 2026
e01139b
docs(goal): verdict at 5f3849211, thirty-first consecutive PASS (LAB-…
claude Aug 18, 2026
fb6fffc
docs(goal): the one-member bound was falsified three times in one rou…
claude Aug 18, 2026
e637a96
fix(goal): report the shell branch as unknowable, not denied (LAB-135)
claude Aug 18, 2026
288e46a
docs(goal): verdict at 2f5760ddc, thirty-second consecutive PASS (LAB…
claude Aug 18, 2026
6038da6
fix(goal): compare tied candidates by verdict, not serialized identit…
claude Aug 18, 2026
ead22cc
docs(goal): verdict at 58006beb4, thirty-third consecutive PASS (LAB-…
claude Aug 18, 2026
24f7b53
docs(goal): record the head's CI state in the verdict row (LAB-135)
claude Aug 18, 2026
5cfa517
docs(goal): the two source scans are backstops, not closes (LAB-135)
claude Aug 18, 2026
c59442d
docs(goal): CI on the current head in the verdict row (LAB-135)
claude Aug 18, 2026
cc1516e
docs(goal): delete six remembered counts rather than update them (LAB…
claude Aug 18, 2026
0ec3787
fix(goal): grade the mirror of the harness ground-truth rule (LAB-135)
claude Aug 18, 2026
0aebd6c
docs(goal): verdict at fd0791dcd, the kind call, and the counts obser…
claude Aug 18, 2026
f46ee05
refactor(claude-code)!: name the background-work option `detached` (L…
claude Aug 19, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
fix(goal): put A8's precomputed failure in the pre-dispatch gate (LAB…
…-135)

A8 is a pure source scan, so its verdict is final before a run is
dispatched — and it was not in the gate, so a reader with a forbidden
import spent a full model-backed coding run to return a failure already
in hand. The verdict was right either way; what it cost was a run. That
is wrong-extent on the gate itself, which exists for exactly one reason.

Also names two findings rather than folding them: A5's ROWS arm can
certify a fractional account in three distinct ways (a dropped
lastOutcome, gap handling skipped whenever any row exists, and a
successful TaskCreate whose row vanished leaving neither row nor gap),
and the third is the one with nowhere to look — checking gaps beside the
surviving rows would not catch it, because a vanished row emits no gap.

The sort probe's adapter is now disposed in the finally, so a throw
before the read does not leak the handle.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MVcrTezeLMBQRVHRg2QX2f
  • Loading branch information
claude authored and jhoffner committed Aug 20, 2026
commit 041ebf2dad7d7e80d930a203bc2c4ac735126137
Original file line number Diff line number Diff line change
Expand Up @@ -152,6 +152,13 @@ be produced at all.
- Assertion 4 compared the report against the run's *first* activity, so `write, report, write`
passed: it rejected the world where all activity follows the report and accepted the one where
only some does. It now grades the last mutation.
- **Assertion 8's verdict was final before dispatch and was not in the pre-dispatch gate**, so a
reader with a forbidden import spent a full model-backed coding run to return a failure already
computed. **Not a correctness defect — the verdict was right either way — and it is listed here
rather than as a limit because what it costs is a run, not an answer.** It is wrong-extent on
the *gate*: that gate exists for one reason, never spend a run the goal cannot pass, and it left
out one of the failures it already held. Every other member of this class has been a rule
half-applied to code; this one was half-applied to the mechanism that protects the budget.
- Assertion 8 checked module specifiers only, and `process.cwd()` names no module. It now also
scans for bare globals and computed dynamic imports, over the shared `paths.mts` as well as the
reader.
Expand Down Expand Up @@ -308,19 +315,38 @@ not currently grade:

- `lastTouchedAt` on a row is read into no comparison, so a record may carry any timestamp.
- `previousStatus` on a plan row is derived and never graded.
- **A plan row's `lastOutcome` never reaches the grader, because the READER drops it.** The
collection exposes it and the recorder writes it, and `readAccount`'s plan-row mapping carries
`title`, `status` and `previousStatus` only — so a task with a confirmed title and status plus a
later `TaskUpdate` that failed grades `a5-ok`, certifying a transition the state says never
settled. **This is A3's shape one notch weaker, and the pair says more than either half does.**
- **A5's ROWS arm can certify a fractional account in THREE distinct ways, and that list is the
finding.** It is the weakest thing in this check, and no one of the three says so on its own:

1. **A plan row's `lastOutcome` never reaches the grader, because the READER drops it.** The
collection exposes it and the recorder writes it, and `readAccount`'s plan-row mapping
carries `title`, `status` and `previousStatus` only — so a task with a confirmed title and
status plus a later `TaskUpdate` that failed grades `a5-ok`, certifying a transition the
state says never settled.
2. **Gap handling is skipped whenever any plan row exists** — the entry below, kept there
because it was reported on its own, and repeated here because it belongs to this set.
3. **A successful `TaskCreate` whose row is lost leaves neither a row nor a gap**, and
`rows.length !== 0` bypasses loss accounting entirely, so `a5-ok`. The sharp part is the
reporter's own note: **checking gaps beside the surviving rows would not catch it.** A gap is
emitted when a successful create's item id is *unreadable*; a row that simply vanished
produces no gap, so the loss is invisible on both surfaces at once. Every other exemption in
this check works because the absent thing leaves a trace somewhere. This one has nowhere to
look, which is a different repair from the other two and worth knowing before anyone starts.

Taken together they are one statement: **the arm reports on the rows it can see, and has three
separate blind spots about the rows it cannot.** That is this goal's own thesis — a fractional
account certified as whole — appearing three times inside one assertion of it.

**Number 1 is also A3's shape one notch weaker, and that pair says more than either half does.**
A3 is an assertion measuring a property the layer below it has already made true; this is an
assertion blind to a field the layer below it recorded faithfully and the layer *above* the
store threw away. Both are the same question asked at the wrong altitude — *what happened to
this value between the writer and the check?* — and neither is visible from the grader, where
every review of this file has been looking. It stays a limit rather than a fix on two counts:
A5 can fail, so this is one dimension of blindness rather than an assertion that cannot fail;
and its ROWS arm does not execute on a real run at all (`toolCalls === 0`, FIX-1185), so
closing it would change no verdict until the driver does.
every review of this file has been looking. **All three stay limits rather than fixes on two
counts:** A5 can fail, so these are dimensions of blindness rather than an assertion that cannot
fail — which is the bar this PR folds at; and the ROWS arm does not execute on a real run at all
(`toolCalls === 0`, FIX-1185), so closing any of them would change no verdict until the driver
does.
- A `tool_output` whose `toolCall.name` is unreadable matches no tool table and is invisible to
A2 entirely — an existence cell, not a presence one.
- A row's `storageKey` and its `topic` are never checked against each other.
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -1269,6 +1269,13 @@ await runGoal(async () => {
`bare global is a second way in`,
}
: { id: "A8", status: "pass", because: "a8-ok", message: a8Seen.join("; ") };
// A8 is a pure source scan, so its verdict is FINAL here — before a run is
// dispatched. Leaving it out of the pre-dispatch gate cost a full
// model-backed coding run to return a failure already in hand. Not a
// correctness defect: the verdict was right either way. It is wrong-extent on
// the GATE, which exists for exactly one reason — never spend a run the goal
// cannot pass — and did not include one of the failures it already knew.
if (a8.status === "fail") failures.push(`A8/${a8.because} — ${a8.message}`);

// ══ Precondition 1b — calibration against a state whose account is known ══
const state = JSON.parse(readFileSync(fixturePath(import.meta.url, "known-state.json"), "utf8")) as KnownState;
Expand Down Expand Up @@ -1531,9 +1538,9 @@ await runGoal(async () => {
// section is now understating what the run proves — so it fails and says so,
// rather than letting the goal quietly go on under-reporting itself.
const sortProbeDir = mkdtempSync(join(tmpdir(), "state-readback-sortprobe-"));
const { sqliteStores: probeStores } = await import("@flow-state-dev/store-sqlite");
const adapter = probeStores({ filename: join(sortProbeDir, "sort-probe.sqlite") });
try {
const { sqliteStores: probeStores } = await import("@flow-state-dev/store-sqlite");
const adapter = probeStores({ filename: join(sortProbeDir, "sort-probe.sqlite") });
const registry = (await adapter.resolve(["primary"])) as unknown as {
request: {
set(id: string, value: unknown, expected: "any"): Promise<unknown>;
Expand Down Expand Up @@ -1593,8 +1600,8 @@ await runGoal(async () => {
`against a store whose read path is doing something else entirely`,
);
}
adapter.dispose?.();
} finally {
await adapter.dispose?.();
rmSync(sortProbeDir, { recursive: true, force: true });
}

Expand Down