Rendered from docs/obligations/0237-eighteen-findings-that-the-paid-layer-campaign-of-run-20261003-1026-left-in-the-probe-recorder-and-the-campaign-driver.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

Eighteen findings that the paid layer campaign of run 20261003-1026 left in the probe recorder and the campaign driver

Context

The builders, verifiers and maintainers of #1474, #1472 and #1659 in run 20261003-1026 wrote these findings as intake lines. Each one is about the probe recorder, the campaign driver or the records they write. None names a reader outside this repository, so the product owner ruled each one RECORD at the end-of-run pass on 2026-10-09. The intake lines are 8 to 10, 21, 22, 30 to 43, 48, 51 and 54 to 57. Lines that state one finding twice are one item here.

Each item was checked on 66dec0f9 at the file and the symbol it names. The surviving mutants are as the verifiers reported them, and the product owner checked that the code each one names is still there. Intake line 49 is an item of HW-OBL-0234, so this record does not repeat it. The next campaign is #1700, and it meets every item below before it spends.

Obligation

What the recorder writes.

• A session that writes auto-memory files under its configuration directory has those paths listed as produced artifacts. step_seed_produced in tools/probe/probe-transform.sh reads every write call of the log and does not exclude the configuration directory. The paths carry no findings key, so the grader cannot grade the session. The evaluation of the campaign states that one session, L7-documentation-absent-p3-r4, was discarded and drawn again for this reason.
• A transcript under docs/probe-runs/ does not say why a repetition is missing. A job that a ceiling refused and a session that ended unrecorded leave the same gap. A recorder line that lists the refused and the unrecorded jobs would explain each gap. Lines 5, 7, 14 and 18 of the campaign have such gaps.
• tools/probe/probe-record.sh states the count of refused provider fetches only in a sentence of record.md, "It refused %s of that kind". clean in tools/probe/campaign-drive.sh matches that sentence, so a change of wording holds the parallel ramp. A provider_refusals: field would remove the match.
• tools/probe/probe-transform.sh refuses a stream that does not parse with exit 4, so such a session never records. The "were not counted" sentences of tools/probe/probe-record.sh are then reached at status 0 only when two jq -s reads disagree.

What the driver does.

• run_job and the end of tools/probe/campaign.sh exit 7 both when a cap or a ceiling refused a job and when a recorder failed. tools/probe/campaign-drive.sh must tell the two apart from slice/refused/ and slice/started/. A distinct exit code for a batch that ended on refusals alone would let a wrapper read the code.
• A resumed batch accepts a --repetitions that differs from the count its job list was made at. A batch of 354 jobs at 177 repetitions resumed with --repetitions 30 rewrites <out>/repetitions to 30 and keeps 354 jobs.
• When a batch ends capped, tools/probe/campaign-drive.sh exits 8 and leaves a.done unwritten, so a relaunch enters batch A again. No verb marks a capped batch accepted. In #1659 a person wrote a.done by hand.
• build_engine in tools/probe/campaign.sh runs a plain cargo build, not tools/hw-cargo. A batch started without CARGO_TARGET_DIR builds an unthrottled target under engine/target.
• campaign.sh --dry-run prices a pooled line at the power calculation. For the regression tier of #1474 it printed 354 sessions at 88.50 dollars, while the run planned the tier's one repetition, 8 sessions in all.

What no record states.

• A batch may raise --repetitions above a tier's count only up to the count the dry run priced. Only the header of tools/probe/campaign.sh states this rule. It rests on the parent's ruling to merge PR #1685 and on the owner's approval of 2,142 dollars. No decision or obligation record under docs/ states it.
• The campaign ceiling moved from 30000 to 205200 cents. Only the comment above the campaign tier in .headwater/probe.yml states the move and its reason. HW-DR-0076 states no figure.
• Before slice 1 of #1659, tools/probe/campaign.sh built no tree for a component arm. The arms no-hook, no-skills, no-claude-md and mcp ran the present tree while each transcript said the arm's change was applied. No committed result is affected, because none of those arms had a committed run, and no record states the defect.
• None of the 8 regression probes declares a string under leaks: in .headwater/probe.yml. So seal.sh --leak prints undeclared for each and can find no leak in any of them.

Tests that hold less than the code does.

• Mutant E of tools/probe/campaign.sh survives. Removing the --repetitions check leaves tools/probe/probe-record-fixtures.sh green, though 0, 010, -1, abc and 1.5 exit 2 by hand. The same gap in tools/probe/campaign-dry-run.sh is inferred and was not run.
• Four mutants of tools/probe/campaign-drive.sh have no case. N19 ignores what was spent before on a shared cap. N16 writes the canary range high before low. N1 refuses a session at exactly the canary bound. N13 caps a retry that recorded one new session.
• tools/probe/campaign-drive.sh has no case against four things: the real tools/probe/campaign.sh, a live record.md with refusals, the overshoot at a bound of 300, and BSD date. Its header says exit 5 means only that campaign.sh refused the plan. campaign.sh also exits 5 when --repetitions is above the priced count.
• Mutants D2 and U1 of tools/probe/campaign.sh have no case. D2 points the dirty-tree check at the batch's output directory and not at the repository root, and the suite's stub answers for any directory. U1 makes --unrecorded accept any status but 0. A case with a status-10 session and a cost file should exit 2.

What governs the tools.

• No document declares governs onto tools/probe/campaign-drive.sh, tools/probe/campaign-drive-fixtures.sh, tools/probe/probe-transform.sh or tools/engine/build-declaration-fixtures.sh. Spec 15 governs tools/probe/probe-record.sh and tools/probe/egress-proxy.py only. HW-OBL-0235 lists campaign.sh, campaign-dry-run.sh and the other probe files of the run before, so the two records together name the ungoverned set. The adjudication of slice 1c of #1659 ruled that the edge waits on a probe how-to or on the next document that describes the driver.

Discharge

Each item discharges alone. A recorder or driver item discharges when the change merges with a case that fails without it. A record item discharges when a decision or obligation record states the rule or the defect. The governance item discharges when a document declares governs onto each file it names. The record discharges when every item has. An item that gains a reader outside this repository leaves this record for an issue, and the record says where it went.