Context & reliability engineering

You probably need this if

• It forgets an order number or a dollar amount it was told twenty turns ago.
• It starts describing what code like this usually does instead of what your code actually did.
• Anything in the middle of a long document doesn't make it into the summary.
• It escalates when someone sounds annoyed, and pushes on when nobody wrote a policy for the situation.
• The final report makes claims you can no longer trace back to anything.

What you get

Facts that survive the retelling

Amounts, dates, reference numbers, anything the customer actually said — pulled out and carried forward as facts, sitting outside whatever the running summary does to them. Summaries keep the story and lose the specifics. This is the fix for that.

Trimming at the source

A lookup returns forty fields and the workflow uses five. Cut it at the boundary, before the other thirty-five pile up and crowd out the thing that mattered.

Putting things where they'll be read

Long inputs get read unevenly: the start and the end land, the middle blurs. So conclusions go up front, detail goes under clear headings, and nothing load-bearing gets buried at the halfway mark.

Escalation with real triggers

Somebody asks for a person, they get a person, immediately. No policy covers the situation? Escalate — that's the actual signal, not how complicated the case looks. Two records match the search? Ask which one, rather than picking.

Claims you can trace

Every finding carries where it came from and when, all the way through to the final write-up. Sources that disagree get shown disagreeing, with attribution, instead of being quietly resolved in favour of whichever was read last.

Work that survives a crash

Findings written down as they're made, state saved somewhere it can be picked up again, and a resume path that doesn't mean redoing the entire investigation.

The judgment behind it

These are the calls that decide whether a system survives contact with production or just demos well. They are not obvious, and getting them wrong is rarely obvious either — until it is.

Annoyance is not difficulty.

Escalating on tone sends the wrong cases to humans and misses the quiet ones where the rules simply have nothing to say. Asking the model how confident it feels fails for the same reason. What actually works: they asked for a person, the policy doesn't cover it, or it's stuck.

Summaries lose in a predictable direction.

They keep the narrative and drop the precision. Anything numeric, dated, or directly quoted has to be carried separately, or it won't survive the trip.

Two sources disagreeing is often one being older.

A good share of what looks like contradiction is really a gap in dates. Carry the dates through and the contradiction explains itself.

Aim review attention, don't spread it.

Reviewer time is the scarce thing. Spreading it evenly across everything wastes most of it. Point it at what's genuinely uncertain — and keep spot-checking the confident pile, because that's where new failure modes show up first.

Engagement shape

Usually 2–5 weeks: a review of what your current runs are actually losing, the facts layer and trimming built, escalation criteria written with worked examples, and traceability enforced at your agent boundaries.