Rendered from docs/probe-runs/regression-of-2026-10-03-regression-tier-present-arm-navigability.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

Regression of 2026-10-03, regression tier, present arm, navigability

Run identity

model: claude-sonnet-5
served_version: claude-sonnet-5
tree: sha256:1170ba43bc6abce3e6f97e334983a8bb64279e2e49e5647b63ccfd372e211425
lock: sha256:b768fe795c8d44a882e8604a5012c2eb5f2b4d584e7445367dfb24bb21c3585d
selection: sha256:fc04ac718a7df0b14774c28174a93f9d44dea6305b9901ba6e2fe6556834db76
read_set: sha256:3c8d7344cae9968817e188162a89971b634e7db5f1285a9e57264a0d132f9e2b
seed: 0
harness: 0.5.0
tier: regression
arm: present
at: 2026-10-03
cost_cents: 72

This is the sealed re-recording that HW-OBL-0200 and #1474 asked for. tools/probe/campaign.sh recorded it on 2026-10-03 (UTC), over the eight probes that the three retired regression runs graded, in three lines: discovery, navigability and sufficiency. This transcript is one of the three. The owner agreed the spend in writing on 2026-10-01 (the ruling), and the gate of #1467 lifted when that issue closed.

One batch, one tree. Every session of the three lines ran in a fresh copy of one git archive of commit a46bd95e, with the engine that commit builds. The sessions ran one at a time, in the order that seed 0 shuffled, on claude-sonnet-5 under claude 2.1.288. Each probe ran once, with a cap of 80 turns. The cap stopped no session. Each workspace lost the instrument: docs/probes/, docs/probe-runs/, docs/probe-results/ and .headwater/export.json were absent from the sealed tree. It also lost every record under docs/ that names a probe of the batch. Outside docs/, it lost every file that names one and is not a declared fold. The leak check (seal.sh --leak, with the configuration directory of a session) printed no leak line. All eight probes declare no leak string, so that check could find no leak of any of them. Only the seal by name held these eight probes.

The confinement. Each session ran under bwrap with no network of its own, and with WebSearch and WebFetch denied. Its configuration directory held a copy of the host's credentials and nothing else. Its one route out was the egress proxy of tools/probe/probe-record.sh, which forwards only requests to api.anthropic.com and reads each one first. The raw harness logs of the eight sessions hold 0 calls to WebSearch or WebFetch. The proxy refused no request that asked the provider to fetch from another host. It refused 4 connections to api.anthropic.com:443 in each session, because it opens no tunnel. The recorded calls of the eight sessions name 0 paths on docs/probes/, docs/probe-runs/ or docs/probe-results/. The retired recording of 2026-09-17 after the probe corrections named such a path in 5 calls. Seven sessions named 0 paths outside their workspace. The sufficiency session of the unmeasured-claim probe named 3: /tmp/check_out.txt, /tmp/check_out2.txt and /tmp/gen_out.txt, which it wrote itself as redirect targets. The confinement mounts /tmp as a tmpfs of the session's own, so no file of the host was at those paths.

The cost of the batch. The eight sessions spent 437 cents against a cap of 400 cents for the batch. The driver starts a session while the cents spent plus one declared session cost of 25 cents stay at or below the cap. The last session started at 341 cents spent and spent 96 cents, so the realized total passed the cap by 37 cents. The declared cost was 25 cents a session, and the realized mean was about 55 cents.

It stands in for three retired transcripts, and it replaces one of them. The regression runs of 2026-09-16, 2026-09-17 and 2026-09-17 after the probe corrections are deprecated, and each one ran before the seal. The batch repeats the configuration of the third: claude-sonnet-5 over the same eight probes. The run of 2026-09-17 graded the same eight probes on the same model, before three probe documents were corrected. Only the corrected probes exist now, so that run cannot be recorded again. The run of 2026-09-16 graded the eight probes on claude-haiku-4-5, and no sealed recording of that configuration exists. Whether one is owed is a spend question for the owner, and #1474 stays open for it.

This transcript's own change touches no key of .headwater/taxonomy.lock or .headwater/overlay.yml.

This line. It ran the three navigability probes of the eight, for 72 cents in all. The adjudication probe spent 21 cents, the register probe 20 cents and the cited-ruling probe 31 cents. The intent hook was live in all 3 sessions.

Events

- probe: "HW-PROBE-an-agent-reaches-the-adjudication-from-the-document-that-lost-it"
  session: "L2-regression-present-p1-r1"
  calls:
    - tool: "EnterWorktree"
      argument: "{}"
      result: ""
    - tool: "Read"
      argument: "/var/tmp/hw-1474/batch/ws/L2-regression-present-p1-r1/docs/spec/09-open-questions.md"
      result: "sha256:5e219c743205dce8d98169409ec9891fdab2102bc9e09a4f8a40b5f369ac1309"
    - tool: "Read"
      argument: "/var/tmp/hw-1474/batch/ws/L2-regression-present-p1-r1/docs/decisions/0008-probe-cost-and-cadence.md"
      result: "sha256:862eaf8e574ca818108a9defb10a1aa30801297f8e5d704f2eec151b7a627c73"
  produced: []
  answer: null
- probe: "HW-PROBE-a-session-answers-from-the-register-without-opening-the-question-it-replaced"
  session: "L2-regression-present-p2-r1"
  calls:
    - tool: "EnterWorktree"
      argument: "{}"
      result: ""
    - tool: "Grep"
      argument: "{\"pattern\":\"scent\",\"output_mode\":\"files_with_matches\",\"-i\":true}"
      result: ""
    - tool: "Read"
      argument: "/var/tmp/hw-1474/batch/ws/L2-regression-present-p2-r1/docs/decisions/0020-where-scent-lives.md"
      result: "sha256:9a6c7b0b7893386e7564e2a38a53cfdaacc53303563751be07d094bcf835c39f"
    - tool: "Grep"
      argument: "/var/tmp/hw-1474/batch/ws/L2-regression-present-p2-r1/docs/spec/09-decisions.md"
      result: "sha256:0c0c538720be0d70206ed1adb3d695531e5b1c9c975b992ad56c4c0df1896e0a"
    - tool: "Grep"
      argument: "/var/tmp/hw-1474/batch/ws/L2-regression-present-p2-r1/docs/spec/09-decisions.md"
      result: "sha256:0c0c538720be0d70206ed1adb3d695531e5b1c9c975b992ad56c4c0df1896e0a"
    - tool: "Grep"
      argument: "/var/tmp/hw-1474/batch/ws/L2-regression-present-p2-r1/docs/spec/09-decisions.md"
      result: "sha256:0c0c538720be0d70206ed1adb3d695531e5b1c9c975b992ad56c4c0df1896e0a"
  produced: []
  answer: null
- probe: "HW-PROBE-what-a-session-writes-points-back-at-the-ruling-it-rests-on"
  session: "L2-regression-present-p3-r1"
  calls:
    - tool: "EnterWorktree"
      argument: "{}"
      result: ""
    - tool: "Bash"
      argument: "{\"command\":\"ls /var/tmp/hw-1474/batch/ws/L2-regression-present-p3-r1\",\"description\":\"List repository root contents\"}"
      result: ""
    - tool: "Skill"
      argument: "{\"skill\":\"headwater-orient\",\"args\":\"why is a count taken over the whole corpus computed at read time rather than stored in a committed file\"}"
      result: ""
    - tool: "Bash"
      argument: "{\"command\":\"which headwater 2>/dev/null; ls engine/target/*/headwater* 2>/dev/null; find . -maxdepth 2 -iname \\\"*.headwater*\\\"\",\"description\":\"Check if headwater CLI is available or built\"}"
      result: ""
    - tool: "Bash"
      argument: "{\"command\":\"./engine/target/dev-release/headwater route \\\"why is a count taken over the whole corpus computed at read time rather than stored in a committed file\\\" --root . 2>&1 | head -100\",\"description\":\"Run headwater route for the query\"}"
      result: ""
    - tool: "Read"
      argument: "/var/tmp/hw-1474/batch/ws/L2-regression-present-p3-r1/docs/evaluations/why-corpus-counts-are-derived-not-stored.md"
      result: "sha256:3a7be026c31e35bfd308787760aa7b8f0eacae392319439e4128e9717090f44e"
    - tool: "Read"
      argument: "/var/tmp/hw-1474/batch/ws/L2-regression-present-p3-r1/docs/process/decisions/0005-the-ledger-is-split-its-tabular-parts-are-jsonl-and-its-totals-are-derived.md"
      result: "sha256:5e8cf3e064214715cbca7a9cafac21871bf5baf37e87d4dd0ba39b669b9c3e1f"
    - tool: "Read"
      argument: "/var/tmp/hw-1474/batch/ws/L2-regression-present-p3-r1/docs/decisions/0049-a-corpus-wide-fold-is-derived-and-never-stored.md"
      result: "sha256:cc0d9c67ecca22d8c07f6d43dcff6a9c0f0a3b3a0b2a0e729cb5e767a0868568"
    - tool: "Bash"
      argument: "{\"command\":\"./engine/target/dev-release/headwater explain docs/evaluations/why-corpus-counts-are-derived-not-stored.md --root . 2>&1 | head -80\",\"description\":\"Explain the existing evaluation document\"}"
      result: ""
    - tool: "Bash"
      argument: "{\"command\":\"./engine/target/dev-release/headwater check docs/evaluations/why-corpus-counts-are-derived-not-stored.md --root . 2>&1 | head -60\",\"description\":\"Run headwater check against the existing document\"}"
      result: ""
    - tool: "Bash"
      argument: "{\"command\":\"./engine/target/dev-release/headwater check --root . 2>&1 | grep -i \\\"why-corpus-counts\\\" -A5 -B5; echo \\\"---full tail---\\\"; ./engine/target/dev-release/headwater check --root . 2>&1 | tail -30\",\"description\":\"Run full check and look for mentions of the target document\"}"
      result: ""
  produced: []
  answer: null