Examples

Four scenarios, each showing the same failure of a conventional toolchain and what Broks Forge does instead.

1 · The overnight failure

Monday, 09:04. An evaluation failed overnight. Nobody knows why.

The usual path. Open the job. Read the error. Open the runs list. Scroll for a failed one. Copy the error string into a log search. Guess at the retention window. Ask in Slack whether anyone changed anything. Wait.

With Broks Forge. Open the evaluation, click Investigate. The Root Cause Explorer has already assembled:

• Immediate cause — "The agent's credentials are being rejected." 6 of 8 sampled failed runs

returned HTTP 401. derived, near-certain.

• Contributing — every failing run reached the same provider. inferred, likely.
• Historical — "This has happened before — Checkout Quality #1 failed 19 days ago." The

recorded cause was identical, which escalates the verdict from attention to risk.

• Related change — "Payments Provider credentials were rotated 3 hours before this ran."

inferred. Proximity, not proof — but it is the first thing to check.

• Timeline — the rotation, the run, the break, all on one axis.

The answer to "what do I do" was available before the first log search would have finished.

2 · The prompt that quietly regressed

A support prompt is on v8. Customers say the tone got worse. Nobody knows which version did it.

The usual path. Diff eight versions in Git. Read eight text diffs with no explanation attached to any of them. Form a theory. Hope.

With Broks Forge. Open the prompt's Evolution tab and read the AI Git timeline. Each revision carries its recorded rationale:

   ● v8  Current production   promoted 6 days ago
   │     "Trimmed the preamble for latency."
   ● v7  Superseded
   │     "Added explicit refusal instruction."
   ● v6  Superseded
   │     "Softer tone after complaints."      ◄── the tone work

v6 added the tone work; v8 trimmed the preamble "for latency" — and took the tone instruction with it. Someone removed a decision without knowing it was one.

Ask Brok "What changed between these revisions?" for the field-by-field diff, then "Show me the evidence" to see whether any evaluation actually covered v8. Often the answer is no — which is itself the finding.

This is the failure mode Engineering Memory exists to prevent.

3 · The cost rise nobody can explain

Spend is up 40% this month. The finance question arrives on a Thursday.

The usual path. A cost dashboard shows the rise. It cannot say what changed, because it does not model changes.

With Broks Forge. Ask Brok "Why did cost increase?". It reads the real evaluation telemetry across the period and reports the trend with its per-run breakdown — and, critically, states the epistemic footing. If the record cannot attribute the rise, it says unknown rather than inventing a cause.

Then follow the thread: "What changed overnight?" for the period summary, and the artifact's Evolution tab for what was promoted in that window. If a model was swapped in a revision, the rationale is right there on the timeline.

The platform will not pretend to a causal claim it cannot support. What it does instead is put the spend, the changes and the reasons on the same page so a human can make the call.

4 · The promotion nobody can defend

Design review. Someone asks: "why is the agent configured this way, and what proves it works?"

The usual path. Silence, then archaeology.

With Broks Forge. Ask "What engineering decisions remain unsupported?"

"2 decisions have no evidence behind them." "'Refund Prompt v3 promoted' has no evaluation standing behind it."

An unsupported promotion cannot be defended in review and cannot be safely reversed, because nobody knows what it was worth. Each one comes with a next action — open the decision, or measure the artifact.

Run the evaluation, and the decision acquires evidence. The claim becomes Knowledge. Next review, the answer is a link.

The pattern across all four

See also: The Engineering Workflow · Best Practices