Show me the page

The 50 MGD target at the South Bay plant carries two dates in two public documents. Our own review caught the difference before the draft went anywhere.

On 23 October our review step did its job. We were checking a draft we had written about the South Bay International Wastewater Treatment Plant, the federal plant on the US-Mexico border that treats sewage from Tijuana. The draft said the plant would reach 50 MGD "by December 31, 2027," and it cited a November 2025 release.

We opened the release and found the sentence. It does give December 31, 2027 as a deadline, for something else: "December 31, 2027, is the deadline to complete all Minute 328 projects." Those are Mexico's projects. On the plant itself, the same release says only that the US Section of the International Boundary and Water Commission (USIBWC) is "starting to plan for an accelerated 50 mgd expansion of SBIWTP."

I believe every fact an AI system gives you about a utility should come with its source, the date of that source, and a label saying whether the document stated it or someone derived it. The South Bay check is the reason I hold that view as firmly as I do: the page showed us the date before anyone else read it.

Where the date came from

The December 2027 date for 50 MGD is in the public domain, in a different document. In a statement on 24 July 2025, the USIBWC Commissioner said: "We also have speeded up our originally announced plans to double the plant's capacity to at least 50 MGD. This is now scheduled to be done by December 2027." There is no day in that sentence.

The State Department's FY2027 budget request, dated 20 April 2026, asks for an increase of $242.2 million "to complete rehabilitation and expansion of the SBIWTP to 50 mgd secondary processing, with a 75 mgd peaking factor by 2031."

So one headline number carries two dates in two public documents: December 2027 in the Commissioner's 2025 statement, and 2031 in the 2026 budget request. The draft had taken an exact day from a deadline for Mexico's projects and attached it to the capacity of the US plant, and opening the page is what showed it.

I do not know yet whether the two documents describe the same scope. The budget request names secondary processing and a peaking factor. The 2025 statement says "at least 50 MGD" and does not name a treatment level. That is a question for USIBWC, and I would want its answer before writing that the date slipped.

Each of the three documents is accurate on its own terms. The risk sits in one step: joining a number from one page to a date from another. Anyone who summarizes public records takes that step many times a day, and so does every AI tool that writes a briefing.

One 50 MGD target, three documents, two dates for the plant. Each bar is as wide as the date its document names: a month, a day, a year. The draft took “50 MGD” from A and the exact day from B, which belongs to Mexico's projects. From the USIBWC Commissioner's statement (24 Jul 2025), the EPA first quarterly update (21 Nov 2025) and the State Department FY2027 budget request (20 Apr 2026).

Why shared memory makes it worse

In a Forbes column by Sandy Carter on The Enterprise Brain (11 August 2026), Carter frames AI in three stages: an agent that keeps nothing between tasks, a loop that runs one workflow, and a brain that shares what it knows across the business. The problem the framing names is fragmentation, many small systems that know nothing about each other. Ragy Thomas, one of the authors, agrees with it in a comment on Carter's LinkedIn post.

The fragmentation problem is a fair one to name, and a shared layer is a reasonable answer to it.

A shared layer also spreads whatever it holds. If that date had gone into a shared store before our check, every account plan and every briefing that drew on the store would have carried it, and each copy would seem to confirm the others. The larger the layer, the further one stale or wrong fact travels.

Security people track the deliberate version. OWASP, the open security community, lists memory or context poisoning among the risks for agentic AI: false information gets into what an agent remembers, and the agent acts on it later. An honest slip in a draft would travel the same way, which is why the check belongs before the store.

What a glass box looks like for one fact

On the book's site, the authors use the term Glass Box: every AI decision transparent, auditable and explainable. I think the approach is critical, and building it is the hard part today. The models from the mainstream AI labs are black boxes. Some open-source and open-weight models are better suited to observability, audit and explanation.

For a fact about a utility, I would make it concrete with four fields: what the fact says, where it came from (document, date, page), whether it is stated or derived, and when someone last checked it. Here is the South Bay record written that way.

"50 MGD by December 2027." Stated. USIBWC Commissioner's statement, 24 July 2025. Last checked 23 October 2026.
"50 mgd secondary processing, with a 75 mgd peaking factor by 2031." Stated. State Department FY2027 Congressional Budget Justification, 20 April 2026, IBWC section, pp. 60-61. Last checked 23 October 2026.
"December 31, 2027 is the deadline to complete all Minute 328 projects." Stated. EPA quarterly update, 21 November 2025, posted on the IBWC site. Last checked 23 October 2026.
"Two public documents give two dates for 50 MGD at the plant." Derived, by comparing the first two. Open question for USIBWC: do they describe the same scope?

Facts / Riverbend Water Authority / Plant expansionACME · workspace

Fact record: treatment plant expansion

Riverbend Water Authority · as of 24 Oct 2026 · 4 of 31 facts shown

All facts (4)Stale (1)Derived (1)Open questions (1)

FactTypeSourceLast checkedFlag

1 Expansion to 40 MGD by December 2028StatedBoard minutes · 14 Jul 2026, item 620 Oct 2026Current

2 Expansion complete by 2030StatedCapital plan FY2027-31 · p. 4120 Oct 2026Current

3 Two documents give two dates for the expansionDerivedFacts 1 and 220 Oct 2026Open question

4 Expansion financed by a $60.0M state loanStatedState loan list · 3 Mar 20268 Apr 2026Stale

StaleFact 4 was last checked 199 days ago. Reopen the source before it goes in front of a utility.

Fact 3, opened

Two documents give two dates for the expansion

SourceBoard minutes · 14 Jul 2026, item 6Capital plan FY2027-31 · p. 41TypeDerived by comparing facts 1 and 2

Last checked20 Oct 2026Open questionDo both documents describe the same scope?

The four fields from the South Bay record, on an invented utility: what the fact says, where it came from, stated or derived, and when someone last checked it. The derived fact carries its open question, and the fact nobody has reopened in 199 days is flagged. A sketch of the idea: AquaIntel does not show these fields on every fact today. Mockup on invented data: not a customer, not a real utility.

"Stated" means the document says it in words, in one place. "Derived" means someone computed it or put two sources together. A derived fact can be useful and correct. The label tells the reader where to check the join, and the join is where our review found the South Bay date.

The last-checked date does a separate job. A statement from July 2025 described the plan as it stood in July 2025. A budget request from April 2026 gives a later date, and I have found no congressional action on that request yet. A checked date tells you how old your knowledge is, so a two-year-old fact cannot pass for a new one.

Freshness is a real question in this data. In the board minutes AquaIntel had collected as of 21 September 2026, we looked at 525 utilities whose minutes files carry a meeting date. The latest minutes were a median of 76 days old. A quarter were 20 days old or less, and a quarter were more than 173 days old. For 39 utilities, nothing was newer than a year. A briefing built from that record is only as current as the last minutes the utility posted, and the last-checked date is how the reader sees that.

0 to 90 days54%

91 to 180 days (derived)23%

Over 180 days (derived)23%

How old the latest posted minutes were across 525 utilities on 21 September 2026, as a share of those utilities. The median was 76 days. A quarter were 20 days old or less, a quarter were more than 173 days old, and 39 utilities had nothing newer than a year.AquaIntel reading of public board minutes, 525 utilities whose minutes files carry a meeting date, 21 Sep 2026. The two shares over 90 days are derived.

A test for Monday

Take one briefing you will use this week: an account plan, a territory summary, or an AI-generated profile of a utility.

Pick one fact in it that would change what you do, such as a capacity number, a project date or a loan amount.
Ask for the page: the document, its date and the page number. If the briefing came from a tool, ask the tool.
Open the document and find the sentence. If the number and the date sit in the same sentence, mark the fact stated. If they come from two places, mark it derived, and check that the two places describe the same thing.
Write today's date next to it as the last-checked date.
If you cannot find the page, keep the fact out of anything that goes in front of a utility until you can.

The Monday test, one row per step. Step three is the one that caught the South Bay date in our draft: the number and the date came from two places, and the check showed the two described different things.

For utility staff, the same test works in the other direction. When a vendor brings you a figure about your own system from an AI tool, ask to see the page.

At AquaIntel we label figures "derived" in our written work, and the South Bay check is a good example of why.

Sources

USIBWC Commissioner statement on the Tijuana River sewage memorandum of understanding, 24 Jul 2025 ("at least 50 MGD ... by December 2027"): ibwc.gov
EPA first quarterly public update, 21 Nov 2025, posted on ibwc.gov ("December 31, 2027, is the deadline to complete all Minute 328 projects"; "starting to plan for an accelerated 50 mgd expansion of SBIWTP"): ibwc.gov
US Department of State, FY2027 Congressional Budget Justification, 20 Apr 2026, IBWC section pp. 60-61 (+$242.2 million, 50 mgd secondary, 75 mgd peaking factor, by 2031): state.gov
Forbes, Sandy Carter, 11 Aug 2026, column on The Enterprise Brain with comments from Ragy Thomas and Sravan Vadigepalli (agent, loop, brain framing).
The Enterprise Brain: Rewiring Your Business for the AI-Native Era (Ragy Thomas, Sravan Vadigepalli, Chandhu Nair, 2026). Book site theenterprisebrain.ai (Glass Box).
OWASP Top 10 for Agentic Applications 2026 (risk ASI06, Memory and Context Poisoning).