DUNIN7-M4 — INFRASTRUCTURE CHANGE REQUEST

CR-2026-183 — B-61: material reaches the model that cannot be cited (v0.1)

Version. 0.1 · Date. 2026-08-07 · Author. Claude.ai (drafting) / Marvin Percival (approving). Target. /Users/dunin7/loomworks-engine, branch main. Baseline. Post CR-2026-181 (engine 366102d). CC confirms exact HEAD at Step 0. Companion to. scoping-notes/loomworks-recall-window-tier-2-relationship-scoping-note-v0_1 §3 (SEPARATE CANDIDATE ITEM); CR-2026-180's Step 1 findings, which found this blocking B-13. Charter. Autonomous regime. Status. Ready for CC execution, with a read-first Step 1. Priority. Ahead of the rest of Wave 2 — blocks B-13.

1. Summary

On a recall turn the model receives two note blocks. _format_past_input (prompt.py:480) produces one, and its rows become AnswerSource. Tier-2's "RECENT NOTES" block (prompt.py:332) enters the system prompt on every turn and contributes nothing to sources.

So an answer can be shaped by material it cannot cite. That sits against B-11's guarantee and against the seed's posture that claims should be checkable.

Live, not latent. The scoping note flagged that no answer had been shown to lean on uncited Tier-2 material. CR-2026-180's Step 1 closed that gap: a recall turn is exactly the case where both blocks reach the model. The demonstration is structural and no longer hypothetical.

2. Step 1 — read before deciding (read-only, and this CR does not pre-decide the fix)

Unlike the rest of this wave, the remedy here is not settled in advance, because the three candidate shapes have materially different consequences and the deciding facts are not in Claude.ai's hands.

CC establishes and reports:

1. What Tier-2 is for. Why does a standing RECENT NOTES block exist alongside the recall block? What breaks if a recall turn doesn't get it? Read the code and any scoping material before answering.
2. Overlap. On a recall turn, how much of Tier-2's content is already in the recall block? If it is largely duplicate, the remedy is different than if it is largely additional.
3. What AnswerSource construction would need to cover Tier-2 rows — is the identifier available at that point, or does Tier-2 assemble from something that doesn't carry one?
4. Whether the seed speaks to this. Read the seed's Memory and Rendering sections for anything about what an answer must be able to cite. If the seed commits to something here, halt and report — this becomes a seed-conformance question, not an engineering choice.

Three candidate shapes, for CC to evaluate against 1–3, not to choose blindly:

• (a) Tier-2 contributes to sources — cite everything the model saw.
• (b) Tier-2 is suppressed on recall turns — the recall block is the answer's material; the standing block is for other turn types.
• (c) Tier-2 is marked in-prompt as non-citable — the model is told it may not ground an answer in that block.

Report a lean with reasoning. Halt for Operator decision — do not build the remedy in the same session that reads.

3. Out of scope

• The 50-vs-20 sizing question (B-62). Deliberately separated. Sizing governs how much uncitable material reaches the model, never that it is uncitable. Do not settle B-62 by picking a remedy here.
• B-13 itself. It unblocks when this lands; it does not land inside this CR.

4. Build steps

| Step | What | Mode | |---|---|---| | 0 | Pre-flight. | Auto | | 1 | Read per §2. Report findings and a lean. Halt. | Auto, halt | | 2+ | Build the chosen remedy — scoped in a v0.2 amendment to this CR after the Operator decides. | Blocked |

5. Kickoff prompt

Execute CR-2026-183 v0.1 Step 1 only, at ~/Downloads/cr-2026-183-
loomworks-b61-citability-v0_1.md. Confirm the CR number against the
ledger first.

This CR deliberately does NOT pre-decide its fix. Step 1 is a read, and
you halt after it. Three candidate remedies are named in §2 — evaluate
them against what you find, report a lean with reasoning, and stop.

The one thing that changes the CR's nature: if the seed commits to
anything about what an answer must be able to cite, halt and say so
plainly. That would make this a seed-conformance question rather than
an engineering choice, and it isn't settled inside a CR either way.

Do not settle B-62 (the 50-vs-20 sizing question) as a side effect.
Sizing governs how much uncitable material reaches the model, never
that it is uncitable.

DUNIN7 — Done In Seven LLC — Miami, Florida CR-2026-183 — B-61 citability — v0.1 — 2026-08-07