Token efficiency, measured on 180 verified pgAdmin candidates: stage 2 costs
$0.0139 per candidate, 90% of the spend, because the conversation is re-sent on
every turn and grows with each tool result.
- Context files now precede the question in the hunter's first message. DeepSeek
bills a cached prefix at 1/30 of a fresh one, by prefix; with the question first,
every candidate asked about the same file (about 5 per file) broke the prefix on
its first line and paid for the file again. A test pins the order.
- The verifier runs each candidate at 3 steps and repeats it once at 6 only when
the hunt spent every step on tool calls without a verdict. On the seven declared
v2 sites, 3 steps found 5 and 8 steps found 6; most candidates conclude in 3.
- Per-candidate token usage (prompt, cached, completion, calls) and escalation
are recorded, and each case reports its cache share, so the next efficiency
decision is made on numbers.
Validation on the 180 pgAdmin candidates plus the 22 known sites (about $1)
waits for account credit: DeepSeek returns 402 at the moment.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo
FUXA failed every hunt with HTTP 400, and neither run could say why. The client
discarded the response body. With the body shown, the reason was "This model's
maximum context length...": FUXA ships minified bundles whose single lines run to
megabytes, and grep_repo returned a matched line whole.
- grep_repo and find_refs quote a matched line (240 characters), never the whole
line.
- Every tool result is capped at 16 x max_chars, with a truncation marker.
- OpenRouterChatClient errors carry the provider's response body (first 500
characters). That would also have exposed the rejected model id in one call
instead of eighteen days.
After the fix, FUXA candidates verify normally at $0.004-0.013 each. A test covers
the megabyte-line case.
verify_sinks.py: skip futures cancelled at the spend ceiling instead of reading
them (the run crashed with CancelledError).
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo
The final-answer turn requested the provider's JSON mode with the tools still
attached. DeepSeek (deepseek-flash) then answered with the format itself
(`{"type": "json_object"}`), wrapped the array (`{"type": "json_object",
"value": [...]}`), or emitted its tool-call markup as text. A hunt that had found
the vulnerability reported nothing.
Measured on the v2 sink verifier with the declared vulnerable functions: 0 of 7
confirmed in JSON mode, 5 of 7 asked plainly (6 of 7 at 8 steps). The model named
Medplum's actual redirect-URI prefix-match bug, LMDeploy's redirect bypass of
_is_safe_url, and ContextForge's unsandboxed Jinja2.
The final turn is now plain text without tools. The parser also reads a
`{"value": [...]}` wrapper. Two tests that pinned JSON mode now pin the plain
turn, and a new test reproduces the wrapped answer.
Invalidates as measurements: the hunter experiment on v2 (c4a0bd8) and the first
sink-verifier run. Both used this turn.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo
This is the deciding experiment of the strategy review, fixed before any v2 pin is
hunted.
- Hunter: the unchanged tool hunter (default prompt, tools and 4 steps) on
deepseek-flash.
- Scope: the ranker's top 20 regions for the case family, in four hunts. The model
never sees the fix diff or the declared sites.
- Families: a fixed keyword table maps the model's family to the protocol's three.
- Scoring: exactly protocol v2's scorer.
- Cost: a hard $10 ceiling.
hunt.py runs it and scores it (--score). A pilot on a v1 development case
(mongo-express) found the vulnerability at lib/bson.js:60, at $0.031 per pin,
alongside a lot of speculative noise.
Transport fix found by the pilot: the DeepSeek endpoint the hunter builds was
unpriced, so the cost meter read $0.00 for any number of calls. It is now priced.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo
Every model call has failed since 2026-09-10 (83ab533):
- The default detector and judge model was `deepseek-v4.1-flash`, the product name.
The API rejects it with HTTP 400 ("The supported API model names are
deepseek-flash, deepseek-v4-pro"). The judge and pipeline record an endpoint
failure as a degradation, never a verdict, so nothing failed loudly.
- `tool_hunter.resolve_hunter_client` resolved OpenRouter only. In this project
DeepSeek is the chat provider and the OpenRouter key is scoped to embeddings, so
the hunter could not reach a chat model at all; the call came back 401.
The defaults are now `deepseek-flash` (priced). The hunter resolves DeepSeek first
when its key is set, and OpenRouter stays as the explicit or fallback route.
Verified live: plain completion and a tool call through both the endpoint resolver
and the hunter client.
Not affected: the independent evaluations (finding_dump runs with
max_model_calls=0) and the 2026-09-06 hunter measurement, which predates the id.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo
Measured live against DeepSeek on one pair: the model spends its step budget
on tools and then answers in a paragraph that names the sink, the argument
and the line -- which parses as nothing. A detector that found the bug and
reported zero is indistinguishable from one that found nothing, and every
family in round zero would have baselined at zero for this reason alone.
The loop now takes one more turn when the hunt ends in prose or runs out of
steps, asking the provider for JSON and handing back the reasoning the model
already produced. It formats an answer that was reached; it never asks for a
new one, and a hunt that called no tool never gets there.
Same pair, same model, before and after: recall 0.00 -> 1.00, Youden 0.00 ->
1.00 over three runs per side.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG
Eight of nine group-3 tasks were rejected. Each defect was hiding a number,
and the numbers are why the loop was not learning:
- The fixed side was judged against the *vulnerable* side's function
ranges. 191 of 227 pairs with a resolvable labeled span (84%) have
different spans on the two sides, so a leak was recorded as silence,
always in the direction that flatters Youden.
- Round zero measured every family with the generalist. `family:unknown`
earns no credit against any real family, so every family but `unknown`
baselined at a structural zero and every noise floor with it.
- The cross-family sweep counted absolute failures instead of negative
flips. On a corpus where most pairs fail at round zero that rejects
every round for damage it did not do. Unscorable rows voted too.
- K runs was declared and never threaded; `pairs --k-runs` was parsed and
discarded, and its only test grepped `--help`.
- The meta-agent proposer was a stub, so `learning round` could never
propose anything. It now composes MetaAgent.evolve against a scratch
copy of the family directory, and `--proposals FILE` runs a recorded
sequence offline.
- `refused_holdout` was a vocabulary word no code path could produce, and
the covering test accepted either outcome.
- Cost was metered once, after everything had been paid for; a stage that
raised left no record at all.
- The corrected lever number was published beside the earlier one rather
than in place of it, and the test named for that property never checked.
- Refusals were `continue` statements: `Refusal.degradation()` existed and
nothing called it. A train-split identical twin was reported as
`split:train`, which is untrue.
Test repairs that each changed a number: two split tests measured an
`x = 1` fixture that teaches nothing whatever the rule says; two CLI
scoreboard tests asserted against all-zero output produced by a stub that
answers twice and puts its finding outside the labeled function.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG
Task 2.1. Tier one decides from what the repository already knows: a declared
family label, an obligation kind, a mechanism id, the rule a finding came from,
or the labeled weakness, and it never spends a model call. Tier two asks one
closed question in JSON mode with thinking disabled, and abstains when the
answer names an id outside the taxonomy or the reply's mean token
log-probability says the model was guessing. Tier three is abstention. Every
answer records which tier decided it.
Answers are multi-label and come back in taxonomy order, so two runs of the
same question compare equal. The classifier proposes and never writes a label,
and it owns validation of the labels the corpus carries, because the corpus
stores a family as an opaque string.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG
Task 1.5. Four tools join the three the hunter already had. Resolving a
definition by symbol returns its file, line span and body, decorators included,
so a detector stops guessing what a callee or a route handler does. Listing
entry points, flows and obligations answers with identifiers and line spans
only, so a structural tool can never become a second way to dump the
repository. Every path a caller supplies is clamped to the scan root first.
The loop now takes the system prompt, the user prompt, the files to start from,
the tool subset, the character limit and the tags its findings carry, each
defaulting to today's behaviour, and it replays the provider's reasoning field
on the next turn because the provider otherwise drops its own chain of thought
across a tool call. A detector configuration is exactly this set of knobs.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG
Task 1.2. Detectors, the classifier's model tier, the proposer and the judge
resolve one chat endpoint: an explicit override first, then the scripted client
flag, then an endpoint named in the config, then DeepSeek, then OpenRouter.
Prove-order retrieval keeps its own embedding endpoint untouched, because the
chat provider we default to offers none.
The adapter handles the provider's documented quirks: thinking is on by default
there and silently ignores temperature, so it is disabled explicitly; the
reasoning field is kept on the reply for the caller to replay across tool turns;
JSON mode has no schema and occasionally returns empty content, so one retry is
built in, never a loop; token log-probabilities are requested only when the
caller wants them, and their mean is what lets the classifier abstain.
Cost comes from the provider's own usage fields, cache hits priced separately,
and a model with no recorded price reports no cost rather than a made-up one.
The endpoint record deliberately holds no key, so it is safe to journal.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG
