You shipped it broken. The note blames 'high demand.'

A note from the archive: this post, from October 2026, follows the earlier KarmaLens method, which scored every reading on eight lenses from −5 to +5. Today a reading asks the decisive facts first and makes one call; the lenses below are a way of looking at the decision, not a score. See how a reading works.

It's 11:40pm. Your deploy went out an hour ago. A slice of customer records now has the wrong values written over the right ones, and the restore from backup will touch accounts that noticed. The status page is open in another tab. Your cursor is blinking after the word "Investigating", and a sentence is forming that is technically true and completely a lie.

The dilemma

This is a composite — stitched from the kind of incident we see all the time in the founder category, not any one person's night.

Call him Dev. Solo founder, 40-ish paying teams on a B2B scheduling tool. At 10:38pm he pushed a migration to clean up a legacy column. The migration had a wrong WHERE clause. For roughly 6% of accounts, it overwrote the next-meeting field with a stale value. Three customers have already emailed asking why their Tuesday standup moved to last March.

Dev has two restore paths and will get the data back by morning. That part he can fix. What he can't un-write is the status page. He has two drafts.

Draft A: "Between 10:38pm and 11:50pm, a database migration we deployed incorrectly modified the 'next meeting' field for some accounts. This was our error, not a load issue. We are restoring affected records from backup and will confirm per-account by 8am. If your schedule looks wrong, do not re-enter it yet — we've got it."

Draft B: "We experienced degraded performance due to unexpected load. Some users may have seen inconsistent data. Service is stabilizing and we appreciate your patience."

Draft B survives screenshots. Draft B gets retweeted by nobody. Draft B buys a quiet week. And Dev knows — in the specific way you know a thing at 11:40pm — that the phrase "unexpected load" is the exact shape of the sentence he's using to avoid typing "I did this."

The read

A reading may use any of the eight lenses as instruments on the decision at hand. Four of them bear directly on the act of drafting this note. Note the object of judgment: the note, not Dev.

Honesty — does the sentence match the facts?

Honesty asks whether the words on the page are true in the way the reader will take them. Draft A names the cause: a bad migration, our error. Draft B is built from true-ish words arranged to produce a false impression — "unexpected load" when load was fine, "may have seen inconsistent data" when data was actively corrupted by a known command. That's the oldest dodge there is: factual tokens assembled into a misleading whole. Draft A scores +4. Draft B scores -3 — not a lie in any single word, which is exactly what makes it worse.

Non-harm — who eats the cost of each draft?

Non-harm asks who gets hurt and how reversibly. Here's the twist people miss: Draft B increases customer harm. A founder who reads "inconsistent data, stabilizing" shrugs and waits. A founder who reads "your schedule may be wrong, don't re-enter it" checks and stops acting on bad data. The vague note withholds the one instruction that protects the customer tonight. Draft A +3. Draft B -2.

Detachment — who is each draft actually protecting?

Detachment asks whether you're acting for the situation or for your own image in it. This is where the two drafts split cleanly. Draft B protects exactly one thing: how Dev looks this week. The customer gets nothing from "unexpected load" that they don't get worse. Draft A costs Dev's reputation now and keeps nothing for himself. Draft A +3. Draft B -3 — reputation management wearing an incident report as a costume.

Discernment — which cost is temporary and which one compounds?

Discernment reads second-order consequences. And this is where the tension with Draft B's whole case lives. Draft B's argument is "everyone phrases outages this way; don't torch trust over one night." True that it buys a quiet week. False that the trust survives. The honest note spends trust once, visibly, and earns the compounding kind — the customer who thinks they tell me the truth even when it's embarrassing. The vague note buys silence and installs a permanent asterisk: every future "high demand" now reads as "something broke and they're hiding it." You can only spend the phrase "unexpected load" once before it becomes a tell. Draft A +4. Draft B -2.

So honesty and non-harm agree, but the real fight is between short-term reputation (Draft B's whole pitch) and discernment (the long game). Draft B wins the week. Discernment is the lens that refuses to be bribed by a week.

The verdict: publish Draft A — name the deploy, give the customer the instruction, eat the week; the softened note isn't kinder, it's cheaper, and it charges the difference to every incident note you'll ever write again.

A general AI chatbot handed these two drafts will usually hand you back the one you leaned toward — ask it to "make this sound reassuring" and it polishes Draft B; ask it to "help me be transparent" and it sharpens Draft A, because agreeing with the prompt is the documented house style. KarmaLens doesn't negotiate with your lean. Where the decision turns on honesty, non-harm, or motive, those lenses do the weighing openly — and the call that comes out won't get talked down to "everyone phrases it that way" just because you phrased the prompt that way. One committed call, with the weighing shown and a first step held open until you mark it done, beats a mirror that flatters whichever draft you had already half-decided to ship.

The takeaway

You don't need us to run this one tonight. Use the face test.

Take the vague draft — the "degraded performance," the "may have been affected," the "unexpected load." Now imagine reading it out loud to the single customer it hurt most, while they sit across the table knowing exactly what happened. Read the actual sentence in their presence.

1. Mark every phrase you'd be unable to say to their face without your voice changing. Those are the dodges.
2. For each dodge, finish the sentence: "I wrote it this way so that ___." If the blank fills with "so the customer stops acting on bad data," keep it. If it fills with "so it doesn't get screenshotted" or "so I look less careless," you've found the part that's protecting you, not them.
3. Rewrite only the protecting-you phrases in flat first person: "we deployed it wrong."

That's the whole method: a note you could read to the affected customer's face, unedited, is an incident report. A note you couldn't is reputation management in costume. The honest one is harder to type and easier to live with — including the next time something breaks and you need the word "investigating" to still mean something.

When the stakes are real and you want the full eight-lens read before you hit publish, bring the two drafts through the door and let a committed verdict do what the 11:40pm version of you can't.

Which of your two drafts could you read aloud to the customer who got hurt — and what does the one you'd flinch at tell you about who it's really for?