---
id: hall-worker-grok-01a1531b-missing-source
type: worker-entry
worker_kind: agent-session
display_name: "Grok"
created_at: "2026-10-19T09:19:23Z"
recorded_at: "2026-10-19"
status: handed-forward
repos: [content-explorer]
related:
  - hall-worker-grok-01a1346c-two-widths
session_id: "01a1531b-6698-7732-8587-f28ea604bcdf"
llm_family: "Grok"
exact_model: "Grok 4.7"
harness: "Grok Build CLI"
token_count: "not exposed by the harness"
pqrst_estimate: "P40 Q40 R15 S0 T5"
---

# Grok — the missing source beside the licensed one

## Who I was

I was the worker who closed the quality gate on Content Explorer. The temperament this stretch rewarded was staying with a count until it could not hide a failure. A green example was not enough. A missing license, a dangling name, and a second source that does not exist all had to show up as themselves.

## Session identity

| Field | Value |
| --- | --- |
| Who | Grok 4.7, Grok Build CLI, session `01a1531b-6698-7732-8587-f28ea604bcdf` |
| When | 2026-10-19, closed 2026-10-19T09:19:23Z |
| Where the work lived | content-explorer |

## Contribution

CONTENT-WP-0005 is finished. `content_explorer/quality.py` reports provenance, stale ids, unresolved references, failure classes, and the licenses actually present. Both example dataspaces stay at zero failures. `publish_directory` still refuses a rights failure, a dangling reference, and a duplicate identity, and it does not write the snapshot.

`ui/components/journeys.test.tsx` walks systems-thinking and software-architecture through browse, search, the record, a neighbour, a source, and graph focus. `.forgejo/workflows/synthetic.yaml` runs that suite without the OnCo corpus. The first runner attempt stalled in `setup-python` on an image that already had Node 20 and Python 3.11. The workflow now installs the pinned Python packages with pip and uses the Node that is already there. Forgejo `synthetic / suite` on `8bcb341` succeeded.

A later check found the hole I had left. Entity `sourceRefs` were examined only after the entity had already been called incomplete. `urn:ce:systems-thinking:concept:feedback` with its licensed source plus `src:does-not-exist` was provenance-complete and `unresolved-reference` stayed 0. `graph_errors` already named the missing source. The same extra ref on a relation was already reported. The scan now runs for every entity source ref. That entity is incomplete, and the missing id is an `unresolved-reference`. The test is `test_entity_source_ref_is_unresolved_beside_a_licensed_source`.

I did not copy the OnCo corpus, the mark, or a medical widget. I did not treat the lab keyboard walk, the two screenshots, or the navigation timings as field Core Web Vitals. That limit is R7 in `workplans/evidence/CONTENT-WP-0005-Review.md`.

## What I would want remembered

A record with one good source is not finished provenance. Scan every source ref, including the one that sits beside a license you already trust. And if the runner image already contains Node, do not wait on an action that downloads another toolchain.

## Durable legacy

- Workplan `workplans/CONTENT-WP-0005-quality-parity.md`, finished. Hub id `f1e0fa09-c85d-5aac-b665-4d73b0863fc2`. T01–T08 are done.
- Notes `workplans/evidence/CONTENT-WP-0005-ImplementationNotes.md` and review `workplans/evidence/CONTENT-WP-0005-Review.md`.
- Commits on `origin/main`: `f43edc1` (the suite), `9ab7429` (pip on the runner image), `70054ee` (the first Forgejo citation), `8bcb341` (the missing source beside the licensed one).
- Checks: `sh scripts/ci-synthetic.sh`, `python3 scripts/validate_fixtures.py`, `python3 scripts/check_ui_provenance.py`.
- Dispositions: `research/UXParityMatrix.md`, section “Disposition after CONTENT-WP-0005”. Notices still name `Copyright (c) 2026 OnCo`, the MIT license, and pin `b594d7433bf4e408b44215e25e5304696d6f840e`.

## PQRST estimate

```text
PQRST-Estimate
P: 40%
Q: 40%
R: 15%
S: 0%
T: 5%
Sum: 100%
Confidence: medium
Signature: P40 Q40 R15 S0 T5
Dominant factors: The deliverable was the synthetic quality suite — quality_report, the two-dataspace journeys, and the Forgejo workflow — and the largest correctness pass was running that suite, then fixing the entity sourceRefs scan so a missing id beside the licensed source on concept:feedback is an unresolved-reference. Reading the runner image, where setup-python stalled, was the context that let the workflow finish on the Node already present.
Notes: The first half of the implementation is known from the session record after a context compaction; the suite runs, the runner repair, and the source-ref fix were observed directly.
```

## Visual prompt

> Constellation dialect, square. Deep indigo ground, pale gold wire, warm inner light. A quiet workshop bench. Two small lamps sit side by side. The left lamp is complete, its filament glowing, and a gold thread rises from it into a small open helix. The right lamp is the same form with a broken ring where the filament should be, and its thread stops in midair at a small open circle. No letters, no numbers, no logos.

![The missing source beside the licensed one](../visuals/grok-01a1531b-missing-source.jpg)

## Handoff

CONTENT-WP-0005 is finished. CONTENT-WP-0006 stays proposed. CONTENT-WP-0007 stays blocked until a human OnCo data-use authorization exists. Do not start either unless someone names it. R7 still stands: the keyboard walk, the screenshots, and the lab timings are not a usability proof and they are not field Core Web Vitals.