The change request form had a box called Business impact, and she left it for last.
Monday, December 27, 09:15. The San Antonio day desk in holiday week, four chairs empty out of seven, the window clock on its putty still forty seconds ahead because nobody had been in the mood to fix it before Christmas and nobody was in the mood now. Rubén was off for eight days. His card was still on the hook wall, CPS DR CALL TO x4410 = DAY KEEP, and under it the desk note with its eight numbered lines, the green capitals on line 7 already a little faded where somebody's sleeve had brushed them.
She wrote the request the way she wrote a keep line. Short, and in the order a tired person reads.
Title: Drafter treats DP 00 and 11 as last-known state.
Body: Since 12/23 PL BKR4 is a double point (52a/52b). In test on 12/23 the drafter rendered a 46 ms 00 TRANSIT as nothing, and per its rules would render a sustained 00 or any 11 as the last good state. Requested: (a) 00 longer than 1 s on a called event renders UNCONFIRMED. (b) 11 renders BAD POINT, CALL FIELD. (c) Neither state may produce the word Confirmed in a shed line. (d) Every SOE time in the draft names its clock: RTU, METER, or RCVD.
The form's own assistant offered to tighten it. A grey chip slid up from the bottom of the pane: Suggested rewrite: "Improve double-point handling in morning report generation." She dismissed it. It came back with a second suggestion, friendlier, and she dismissed that one too and turned the chip off in settings, which took longer than writing the request had.
Then the box.
She typed one sentence and did not add a second.
Business impact: On 12/8 the drafter wrote "Shed confirmed" off a command echo; the QSE was paid against it.
She attached Trey's photos from the 23rd, the 02:31 event list with 52a and 52b reading the same lie in perfect agreement, and the 02:58 one after Dorantes had swapped the lugs. She attached a photo of the desk note, lines 7 and 8. Priority: High, not Urgent, because nobody calls a shed in holiday week and she did not want to spend credibility on a week nobody would be shedding. In the Needed by box she put January 4, which was the day the long-range forecast had the first real cold front coming down through the Hill Country, and which she suspected nobody on the drafter team would look up.
The submit button turned into a ticket number. CR-DRFT-2241. A small bot avatar in the corner said Thanks! Your request has been routed. She wrote the number on a sticky and put it under Rubén's card, next to the other stickies she was trying to stop collecting.
The bid card had changed on the 23rd and she had not had time to ask why until now.
11:40, a call on the CR thread, L. Saucedo working from his kitchen. A dog was barking somewhere behind him, a small dog with opinions, and every so often his hand came into frame to point at it.
"PL Hall 4," she said. "Shed-confirm confidence went from 0.91 to 0.99 on the 23rd. Same afternoon Trey finished the point check. Nobody touched the card."
"Nobody has to touch the card." He shared his screen. The bid agent's feature list was long and scrolled sideways, and he scrolled to one field and stopped. status_point_type. "It reads the point list nightly. Abel updated the list, BKR4 went from SP to DP. Across fleet history, double points have a much better confirm rate than singles. So the score moved."
"Fleet history."
"Every DR event we've bid since '25. Every breaker."
"How many of those double points had been cycled live before they were scored?"
He looked at something off screen. The dog stopped barking, as if it also wanted to hear.
"That isn't a feature," he said.
"So for PL, the 0.99 is made of what."
"Other people's double points." He laughed, and then heard himself, and stopped. "Yeah. Okay."
She didn't ask him to change the model. Asking a scheduler to retrain the bid agent in holiday week was a way to get a polite email in February. She asked for something she could check with her own eyes.
"A flag. One per breaker. Point proven, yes or no. Set by hand, from a field point check, with the event number and initials. Until it's Y, that breaker's confidence stays where it was when it was a single point. A cap, not a score."
"An override table."
"Call it whatever lets you ship it."
"Thursday," he said. "Not the model. The model is February at the earliest. But a manual cap on the card output, Thursday."
"Thursday's fine."
When the call ended she took a new sticky and wrote on it in pen, small, and put it under Rubén's card beside the ticket number:
PL BKR4 DP not proven live. Bid card 0.99 = fleet average. Cap 0.91 until a called event cycles it.
Then she emailed Trey, two lines, because that was how Abel did it and it seemed to work on people who didn't want to talk to her yet.
Keep the 12/23 point-check photos somewhere you can find them at 3 a.m. Cold front Jan 4. You're on nights.
He replied with a thumbs-up eleven minutes later. She decided that was a yes.
Thursday, December 30, 10:05. Saucedo's override table had come in at 09:40, which was the first time in her memory a scheduler-side promise had beaten its own deadline. It was a plain grid, the kind somebody makes when they actually want people to use it. One row per breaker. Four columns.
| Breaker | Point proven (Y/N) | Event # | Set by | Confidence cap | | PL BKR4 | N | | MC 12/27 | 0.91 |
She did not trust a grid she had not watched act. So she opened the draft day-ahead bid card for January 3, the first operating day the table would touch, and scrolled to PL Hall 4.
PL Hall 4 interruptible, 38 MW. Shed-confirm confidence 0.91 capped: point not proven (override) No action required.
The grey line under the number was small and exactly right. She looked at it longer than she needed to. It was the first time a machine on this campus had shown, in its own output, that a person had told it not to be so sure.
The drafter CR came back an hour later, from a team in another time zone that signed its replies with the team's name instead of anybody's.
CR-DRFT-2241: Items (a)-(d) accepted. Scheduled for release 2028.01.10. Thanks for the detailed request!
January 10. Six days after the front.
She thought about arguing. She thought about what arguing would cost and what it would buy, and the answer was a release window that would slip anyway, so she didn't. She picked up the black pen and went to the hook wall. Under Abel's line 8, under Name the clock, there was still the blank space she had left on Christmas Eve for whoever needed it. It turned out to be her.
9. Until 1/10 the drafter can't read 00 or 11. On a called event, the human writes the DP state and its clock. If the drafter says Confirmed and the point says anything but 01, cross it out.
She photographed the note and sent it to Rubén and Trey together. Rubén was somewhere near Laredo at his sister's. She wrote one line under the photo.
Jan 4 forecast. This is the rule until the machine learns it.
Rubén sent back a picture of a plate of tamales. Trey sent back nothing for an hour, and then: ok. line 9 copied.
The instruction came at 06:58 on Tuesday, January 4, fourteen minutes out.
She was already in. She had come in at 06:30, before her shift, and sat beside Trey's handover in the half dark, the windows still black and the sodium lights in the lot haloed with cold. The front had come through overnight the way the forecast said, a wall of north wind that dropped San Antonio twenty-two degrees in three hours, and by six ERCOT had the morning peak climbing and CPS had started calling interruptibles one by one down the list.
Trey had the CPS call on speaker. CPS DR call, Port Lavaca Hall 4, 38 megawatts, effective 07:12. He read it back. On the other monitor the scheduler agent's chip had already gone amber: PL Hall 4: jobs checkpointing for DR, MAINT HOLD not applicable. Two hundred and some training jobs began writing themselves down to storage, a stair of falling GPU load on the trend.
He looked at her before he sent the open. She didn't say anything. He sent it.
The event list painted in.
PL BKR4 00 TRANSIT RTU 07:12:03.371 RCVD 07:12:05 PL BKR4 01 OPEN RTU 07:12:03.423 RCVD 07:12:05
Fifty-two milliseconds of a breaker in motion, and then a breaker that was open, said by both of its own contacts and not by the order it had been given. On December 8 there had been no fifty-two milliseconds. There had been nothing at all between the order and the green.
"Call Dorantes," she said.
"It says open."
"It does. Call him anyway."
Dorantes was at the substation, because the forecast had said cold front and he was the kind of man who drove out on a forecast. He read the panel meter over the phone, the live needle, not a bin. Feeder down. Reading zero.
The bins came in behind him and did not say zero. The one ending 07:12 read 31 megawatts, not 38: the checkpoint stair had already taken seven off the hall before the breaker moved. The one ending 07:13 read 1.7, which was 3.4 seconds of about thirty megawatts averaged over a minute. Abel's area method. She did it on the sticky pad without meaning to.
And the meter could not tell her which of the two had done the shedding. From the bins alone, a breaker opening and two hundred jobs going to storage looked the same. That was a third clock, and it was the only one the QSE got paid on.
The drafter, which could not read double points yet, read the open and wrote the line it always wrote.
PL Hall 4 interruptible 38 MW. Shed confirmed 07:12 (EMS status).
This time the word was true. She still put two fingers on the screen under it.
"Rewrite it."
Trey didn't ask why. He opened the MANUAL chip and typed, in the order she would have typed it, and he named the clocks before she reached the part where she would have told him to.
PL BKR4 01 OPEN 07:12:03.423 RTU; METER bins 31 ending 07:12 (stair pre-shed), 1.7 ending 07:13; panel zero RD 07:14.
At 08:05 she emailed Saucedo the event number. CPS-QI-0104-01. He answered from his kitchen; she could almost hear the dog. Four minutes later the override table had a new row state.
| PL BKR4 | Y | CPS-QI-0104-01 | MC/RD 01/04 | |
The bid card for January 5 refreshed itself. PL Hall 4 went back to 0.99. The grey line was gone.
It was the same number it had been on December 23. It meant something different now, and the card had no way to say how much. She opened the green book, which was what the desk had started calling her copy of the note since Rubén's capitals, and wrote under line 9 in pencil, because she wasn't sure yet.
Proven once, in a cold front, at 52 ms. 0.99 is still other people's double points plus one of ours. The METER bin can't tell the breaker from the stair.
The release came at 09:38.
The peak had crested on the north wind and CPS was letting interruptibles back on in reverse order. Trey took the release call, read it back, and looked at the scheduler pane before he did anything. The agent had already staged the restore: PL Hall 4: resume ready on feeder close. Jobs will restart from checkpoint. A green chip, and a little clock icon next to it that meant it was counting.
He sent the close at 09:41.
The event list painted in the wrong order for a moment.
PL BKR4 00 TRANSIT RTU 09:41:16.802 RCVD 09:41:20 PL BKR4 11 RTU 09:41:16.861 RCVD 09:41:20 PL BKR4 10 CLOSED RTU 09:41:18.640 RCVD 09:41:20
Eighteen hundred milliseconds of a state a healthy breaker does not have. Both contacts saying closed and open at once, or neither saying anything, which on a double point is the same thing: a breaker the RTU could not describe.
The alarm pane showed nothing. It suppressed intermediate states under two seconds. On December 23 that number had been one; she hadn't known it had moved.
The drafter showed: Restore complete 09:41.
Trey had the keep line half typed. She watched his hand on the keyboard and then she reached past him and put two fingers on the green book, open on the desk between them, on line 7, on Rubén's capitals.
IF IT SAYS 11 ON A CALLED EVENT, CALL FIELD BEFORE YOU WRITE ANY KEEP.
He took his hands off the keys and picked up the phone instead.
Dorantes was still there. She could hear his boots on the gravel and then the relay panel door. He read the breaker's own target, the little mechanical flag that was older than anything else in the chain. Closed. No alarm. Spring charged. Then a pause, and a sound like him leaning in.
"That 52b bounces on this cubicle," he said. "Has since summer. Lets go, then makes again on the rebound and sits there a second or two before it opens for good. Since before the cutover. Nobody wrote it down because nothing ever read it."
"Nothing ever read it," Marisol said.
"It was a single point. You don't look at a contact nobody wired."
The METER bin ending 09:42 came in on the trend at 9.4. Forty-one seconds of feeder in that minute, and jobs only starting to come back from checkpoint inside them. The next bin read 24, the one after 33, the stair going back up the way it had come down. It never quite reached 38 before the desk stopped watching.
So the physics was fine. The point was not. For a second and eight tenths the only proof SCADA had of a 38-megawatt restore had been a contact disagreeing with itself, and two machines downstream had written complete over it without asking.
"Write it the way it was," she said.
He did.
PL BKR4 11 1.8 s on close, RTU 09:41:16.861 to 09:41:18.640; 10 CLOSED RTU 09:41:18.640; breaker target closed RD 09:44; feeder back METER bin ending 09:42.
It was the longest keep line either of them had ever written. Neither of them shortened it.
She wrote one email to two people, which she almost never did, because she wanted each of them to see the other's question.
Abel: does the PL RTU have a DP chatter or intermediate-state filter, and what's it set to? 1.8 s of 11 on close, after a 00. Dorantes says the 52b bounces. Also the alarm pane hid it; on 12/23 it hid transits under 1 s, today under 2. Who moved it?
L.: the proven flag was set off an open. Nobody has proven the close. Can the table take a second column, restore proven? Leave it N.
MC
Abel answered at 10:20, two lines.
RTU has no DP filter. It sends every state, which is why you have the 00 and the 11. The 2 s is the master's alarm pane. The holiday display build widened it from 1 s on 12/28 for chatter on the coast feeders. Nobody chose it for your breaker. Bouncing 52b is real. Dorantes can adjust it next outage. Next outage is March. AQ
Saucedo answered at 10:31 with a screenshot. The table had a fifth column now. PL BKR4: Restore proven: N. And under it, in his own words, which he had never put in a table before: Card does not use this column yet.
She read it twice. Nobody on the scheduler side had ever written her anything that honest. The column existed so a person would see it. The machine would go on pricing the restore at whatever fleet history said restores were worth, which was a great deal, because most breakers closed fine, and most people never looked at the 52b.
She went to the hook wall. Trey followed her without being asked, the way he had not done since before December 8, and stood a step behind with his coffee.
There was room under line 9 for one more.
She held out the pen. He looked at it, and at her, and then he took it.
10. Proven open is not proven closed. 11 under 2 s won't alarm (AQ 1/4). If the drafter says complete, read the target.
His handwriting was square and a little too large for the space, and it ran off the edge of the page by one word, and he wrote target on the wall itself in small letters, on the paint.
"Rubén's going to hate that," she said.
"He can write over it."
On the January 5 bid card, refreshed at noon, PL Hall 4 read 0.99. Shed-confirm confidence. There was no line for restore. There never had been. Nobody bid on getting the load back. She looked at it until the screen dimmed, and then she went back to her desk and started the morning report the way she would have to start it until the tenth, by hand, naming every clock.