Count task directories per stage: find tasks/working -mindepth 1 -maxdepth 1 -type d | wc -l
(and the same for tasks/reviewing, tasks/revising, tasks/accepted). Expect: 63,
3, 4, 1 — 71 task directories total.

Run GATE_FAST=1 _build/gate-task.sh <dir> against every one of those 71 directories and
tally results. Expect: all 71 runs end with the exact line GATE-TASK PASS, and zero ✗
lines appear anywhere across the combined output — verified true on 2026-08-07 (71/71 pass,
0 failures).

Run it once more, alone, on the one unit that has actually cleared platform acceptance:
GATE_FAST=1 _build/gate-task.sh tasks/accepted/ball-valve-parametric. Expect: exactly
51 lines starting ✓ followed by GATE-TASK PASS.

Final check: grep the accepted unit's task.toml for category = "Hardware" and
subcategory = "CAD" — both present, matching the mission in company/CHARTER.md.

Run source _build/project.env && eval "$ATOMS_DONE_CMD" — this is the literal ground-truth
command _build/progress.sh, _build/assess.sh, and the dashboard all consume. Expect (per
Article II, the real number of complete task dirs on disk): a number at or near 71.
Currently fails: prints 0.

Run the shipped dashboard directly: bash _build/progress.sh. Expect: a fleet percentage
reflecting real completed atoms. Currently prints exactly fleet 0.0% (0/9 tasks) —
confirmed via the shipped script itself, not a re-implementation.

Diagnose: ATOMS_DONE_CMD walks "$UNITS_ROOT"/*/ — one level below tasks/, i.e.
tasks/working/, tasks/reviewing/, tasks/revising/, tasks/accepted/ themselves — and
checks each of THOSE four directories for a top-level task.toml. None of the four stage
folders is itself a task, so the count is always 0. Real units live two levels down
(tasks/<stage>/<slug>/) since the corpus was reorganized into stage folders after
project.env was authored (comment on line 8 still says "each unit = one task dir
tasks/<task-name>/ directly"). Expect (corrected two-level walk): 71.

Final check: grep -n ATOMS_DONE_CMD _build/assess.sh — confirm assess.sh (the per-unit
audit tool) evaluates the same broken command, so this is not an isolated display bug: every
fleet-percentage or per-unit ground-truth read taken today through the shipped tooling
under-reports.

Run _build/gate-task.sh tasks/reviewing/ball-valve-downsize (no GATE_FAST). Expect: 57
✓ lines including ✓ docker-build-environment, ✓ docker-build-tests,
✓ agent-image-isolation, ✓ oracle-runner-reward-aware, ✓ oracle-verifier-roundtrip,
ending GATE-TASK PASS — verified true 2026-08-07.

Run _build/gate-task.sh tasks/reviewing/eir-dc-motor. Expect: same shape, ending
GATE-TASK PASS — verified true 2026-08-07.

Run _build/gate-task.sh tasks/reviewing/pipe-parts-scad. Expect: same shape, ending
GATE-TASK PASS — verified true 2026-08-07.

Final check: all three reviewing-stage units are fully green on the automated gate. Nothing
here explains why they are not yet accepted — per company/NOTES.md the gap is human/SME/Founder
review bandwidth, not a broken build (see F8 for whether that note's own dates are current).

Run _build/oracles/ball-valve-downsize/gate-oracle.sh directly (bypassing the gate-task.sh
wrapper). Expect: the run exercises the full battery — primary solution ×3, an alternate
valid profile ×3, a frozen naive probe (fails), the unsolved base state (fails), three sliver
near-miss profiles (each fails exactly one named pytest check — e.g. test_watertightness or
test_exterior_overhang), every measured-property mutation (each fails), and one
formatting-only variation (passes) — closing with the exact line ROUNDTRIP PASS: primary and
alternate solutions passed three times each; the base and frozen naive states failed; all
three sliver near-misses failed exactly one named test; every measured-property mutation
failed; formatting-only variation passed. Verified true 2026-08-07.

Grep the run's own trace for the reward-file convention rather than trusting exit codes:
Expect lines of the shape mutation-watertightness: container-status=0 reward.txt=0 (a
mutation that fails STILL exits the container cleanly — status 0 — but writes a losing reward)
and format-only: container-status=0 reward.txt=1 for the passing formatting-only run —
proving the Harbor reward contract is judged by the reward FILE's value, never the container's
exit code alone.

Final check: grep -q 'reward\.txt' _build/oracles/ball-valve-downsize/gate-oracle.sh — the
runner itself reads the reward file rather than only checking exit codes (this is exactly what
gate-task.sh's own oracle-runner-reward-aware check enforces).

Run docker run --rm --entrypoint sh fbta[openai credential redacted] -c 'find / -maxdepth 4 \(
-name solve.sh -o -name test.sh -o -path "*/solution/*" -o -path "*/tests/test*" \)
2>/dev/null'. Expect: empty output — nothing found. Verified true 2026-08-07.

Run docker run --rm --entrypoint sh fbta[openai credential redacted] -c 'ls -la /workspace'.
Expect: only the agent's working material (verified contents: data and release
directories) — no solution or tests directory visible.

Run docker run --rm --entrypoint sh fbta[openai credential redacted] -c 'ls /tests' — the
SEPARATE verifier image. Expect: test.sh and test_outputs.py both present — verified
true. This confirms isolation is directional: the verifier legitimately holds the answer key,
the agent provably does not.

Final check: re-read task.toml's [verifier] section — environment_mode = "separate" is
present (this is the contract F5 is proving in practice, not just on paper).

Read company/NOTES.md's founding-state entry: "Chief suite DESIGNED-BUT-UNBOOTED. All four
chief seats carry status: UNPROVEN in staffing.yaml... No seat packet in company/seats/ has
been written yet — packets are Step-6 artifacts." Note the claim before checking it.

Run grep -n "status:" company/staffing.yaml. Expect per the NOTES claim: UNPROVEN on
all five chief roles (ceo + 4 department chiefs). Currently contradicted: every one reads
status: PROVEN-PARKED.

Run find company/seats -type f. Expect per the NOTES claim: no files. Currently
contradicted: five .packet.md files exist (one per chief token —
terminus3-ceo/.packet.md, terminus3-eng-chief/.packet.md, terminus3-ops-chief/.packet.md,
terminus3-product-chief/.packet.md, terminus3-quality-chief/.packet.md), plus
terminus3-ceo/PARK-terminus3-ceo.md.

Final check: the correct behaviour is that company/NOTES.md is either corrected or dated
per-entry so a reader can tell which sentences are still live — right now, someone trusting the
NOTES snapshot instead of staffing.yaml/company/seats/ directly would believe the company has
not booted its leadership when it demonstrably has (all five chiefs PROVEN-PARKED with packets on
disk).

Run curl -s "http://127.0.0.1:8787/api/kanban?project=terminus3". Expect:
"reachable": true and an In Progress column with cards. Verified true 2026-08-07.

Cross-check every chief_token in company/staffing.yaml and every file under _handoff/
against the board's card list. Expect: a /continue <token> card for each of
terminus3-ceo, terminus3-eng-chief, terminus3-ops-chief, terminus3-product-chief,
terminus3-quality-chief, plus the working-session tokens overnight-44,
terminus-batch2, terminus-batch3 — all present, verified true.

Check the other columns. Expect (honest-empty convention): Backlog,
Currently Running, In Development, Testing, Done, and Deployed all report 0 cards
— verified true; nothing has been pushed through those stages yet.

Final check: the board's own "refreshed" timestamp read Jul 28, 2026 while it was still
correctly listing an overnight-44 card whose last activity is Aug 6 — worth confirming with
Observify's own refresh cadence directly (NOT verified here as a defect; flagged as unresolved,
since terminus3-side evidence alone cannot tell staleness from a display label that means
something else).

Read the claim in company/CONTRACTS.md §8 / company/NOTES.md: "downsize idle from 08/04,
eir from 08/05 -> nudge ladder fires ~08/18 (day 14)."

Run find tasks/reviewing/ball-valve-downsize -type f -exec stat -f "%Sm" -t "%Y-%m-%d" {} \;
| sort -r | head -1. Expect per the claim (last touched 08/04): a date on or before
2026-08-04. Currently reads 2026-08-07 (same day this plan was written).

Repeat for tasks/reviewing/eir-dc-motor. Expect per the claim (last touched 08/05): a
date on or before 2026-08-05. Currently reads 2026-08-07.

Final check: this does not prove the escalation math is wrong — "idle" plausibly means
reviewer/SME silence, not file inactivity, and file writes from a build/audit pass would not
reset that clock. It proves the claim is NOT verifiable from file mtimes alone, so the ~08/18
nudge date should be re-derived from whatever signal "idle" actually means (last reviewer comment,
not last file write) before anyone treats day 14 as a hard trigger.

[sensitive content omitted]
[sensitive content omitted]
one of the 71 units correctly tagged. Verified true 2026-08-07.

Count OpenSCAD-bonus-eligible units — metadata.languages containing openscad — across all
71 task.toml files. Expect: 60 of 71. Verified true 2026-08-07.

Final check: per company/CHARTER.md's stated payout ($300 flat + $50 OpenSCAD bonus per
accepted task), if all 71 units were accepted as-is the derived gross would be
71 × $300 + 60 × $50 = $21,300 + $3,000 = $24,300 — stated here as an arithmetic derivation from
the corpus + the charter's own numbers, explicitly NOT a platform-confirmed acceptance figure
(only 1 of 71 has actually been accepted to date).

Run ls -la stlFiles.zip. Expect: present, ~2.7 GB, Jul 31 timestamp.

Run ls -ld extracted. Expect: present.

Run ls -ld think-like-fable. Expect: present.

Run ls -la "Peer Coaches Frontier Bench Spec Doc .md" (note the double space before .md
in the real filename). Expect: present, 17318 bytes, Jul 29 timestamp.

Run ls -la _build/inventory/manifest.yaml. Expect: present.

Final check: none of the five were modified running this test pass — re-run step 1's ls -la
and confirm the size/timestamp is unchanged from what was recorded when this plan was authored
(do not touch, edit, or delete any of them — CONTRACTS §2 hard no-touch, project constitution
P-IV).