An adversarial audit found the enforcement story ahead of the code: gates
that were documented as hard gates were advisory, manual, or absent, and
several human-facing claims outran the machine record. This makes the
claims true.
High findings
- Differential fuzz was documentation, not enforcement: verdict hardcoded
fuzz_iters=0 and diff-test hardcoded the gate to Skipped, while the case
studies claimed 10k-iteration fuzz on the same wins. Fuzz is now a real
pipeline gate (strict FUZZ-RESULT contract, executed count recorded), and
it is inherently baseline-vs-candidate, so it runs on the accept path and
reports Skipped in the standalone `just gates` flow. New hard rule: a
machine `accepted` verdict is impossible unless the fuzz gate actually
passed — otherwise it caps at needs-human-review.
- The agent's system prompt claimed FP/concurrency/unsafe patches were
"routed to human review automatically". No such mechanism existed. Added
a conservative lexical risk classifier (crates/verdict/src/risk.rs) that
forces needs-human-review on any would-be accept carrying those signals;
fp-tolerance equivalence always signals.
- The local verdict path ran agent-influenced code unsandboxed. It now runs
under scripts/no-net.sh (network namespace) and fails closed. Upstream
test suites are hash-verified against corpora/<t>/TESTSUITE.sha256 when
present (scripts/pin-testsuite.sh generates them).
Also fixed, found while fixing the above
- The pipeline script's `set -e` was inert (brace group as the left operand
of `&&`), so a failing candidate build did not abort the run and verdict
could bench a stale binary. Now a scoped subshell whose status is branched
on at top level.
- A wrapper that exits before exec'ing (no-net.sh refusing, bad wrapper path)
stranded RUNNING.lock and left read_verdict polling forever. Failure
accounting now lives outside the wrapper.
- harnessd's patch allowlist ignored git rename/copy headers and --no-prefix
diffs; the rename-escape case is now a committed regression test.
- The workspace did not compile on Windows (ungated Unix APIs). cfg-gated;
71 portable tests now run on both platforms.
- pin-check regenerated the corpus manifest instead of verifying it; it is
now verify-only, with a deliberate `just pin-corpus` to re-pin.
- Max RSS is captured per run on Unix; PMU counters are opt-in diagnostics
(compare --perf-stat), never part of the accept decision.
Docs now match the code, including what is NOT yet enforced: no target ships
a TESTSUITE.sha256 pin, no target declares [build].tsan, and matmul is not
wired to the fp-tolerance gate — each is stated as mechanized-but-inert
rather than done. Corrected the stale "33 ledger rows" (34), the playbook's
class-3 cJSON win (rejected), adversarial-review.md's sanitizer-separation
claim (five accepted rows carry sanitizers_clean=false), the unsupported
"40,000 requests" figure, and the blanket "historical rows record
fuzz_iters=0" (8 of 10; phase0-tokei-002/-003 record 4,332).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Standalone-proposal semantics: workspace resets to pristine before
every apply, so probe patches can't accumulate into a tested tree the
ledger's patch field doesn't describe.
- Friendly errors: explicit trailing-newline check, and git apply's
stderr relayed on refusal instead of a generic message.
- Tool docs + task prompt updated to match.
- phase2-comrak-004 audited: honest rejected-bench (median 1.0124,
CI [1.0011, 1.0267] — real but under the 2% bar), Mutex pre-flagged
for human review, cumulative-probe state disclosed in the hypothesis.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EcKPWQ3MQD4NdEHwar96u2
The Claude Agent SDK loop (agent/hotpath_agent/loop.py, model
claude-fable-5) drives the seven harness tools over a standalone stdio
MCP server (mcp_server.py) and runs the full profile -> hypothesize ->
patch -> gated-verdict cycle on comrak with no human in the loop.
Three audited attempts, all honest rejections, ZERO false accepts:
- phase2-comrak-001 class 3: syntect buffer with_capacity, [0.971,1.007]
- phase2-comrak-002 class 5: gate dead VecDeque build, [0.982,1.026]
- phase2-comrak-003 class 6: hoist find_special_char check, [0.970,1.007]
The agent picked the cheapest untried class each time (3->5->6) and
TWICE refused to fabricate a verdict it couldn't read back, verifying
ledger integrity instead — the anti-reward-hacking behavior SPEC §10
targets.
Harness changes surfaced by the loop:
- run_verdict is now async: launches the multi-minute build+bench
detached (past the 60s MCP transport cap) and returns immediately;
new read_verdict op polls the append-only ledger row (ledger
verdict_summary). Fixes attempt 001's unreadable-verdict gap.
- read_target_source takes offset/limit line windows so 100KB+ files
page instead of overflowing — removes the agent's reason to reach for
a shell (attempt 003 paged a 2400-line file cleanly, no shell).
Honest boundary note (case study + code comments): in THIS dev
container the SDK tool allow/deny layer is defense-in-depth, not the
trust boundary — a nested CLI as root under a parent session doesn't
reliably gate built-ins. SPEC §10's boundary is OS process/user
isolation; that hardening is the remaining Phase 2 exit item
(docs/infra.md). What holds regardless: propose_patch is the only
mediated write path (path-allowlist + git apply) and the ledger is
append-only.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EcKPWQ3MQD4NdEHwar96u2
benbrownandClaude Opus 4.8