The Brain · 03.03
Feynman Repair: A Grounded Self-Explanation, or Nothing at All
The real Feynman technique made concrete: force a compact explanation of a failure, validate it against real evidence, and commit nothing at all if it isn't actually grounded.
Named for a real technique, not a metaphor
The Feynman technique is simple to state and hard to fake: explain something in plain language, and if the explanation doesn't hold up, you didn't understand it — you were pattern-matching. Agentis makes that concrete rather than aspirational. When a workflow node fails in a way the cheap layers couldn't resolve, the platform forces a compact, grounded explanation of what actually went wrong, checks that explanation against real evidence from the run, and only commits a reusable lesson if it survives that check. A weak or ungrounded explanation isn't patched over or stored anyway — it produces nothing, which is the correct, intended outcome, not a failure of the system.
Four tiers of failure-understanding, cheapest first
| Tier | What it is |
|---|---|
| L0 | Deterministic, model-free failure analysis — pattern-matches the error, no reasoning involved. |
| L1 | The engine's own self-heal retry — re-dispatch the node with the error added as context. |
| L2 | A canned, templated agent-memory lesson from the failure reflection service. |
| L3 — Feynman | A queued, model-graded, evidence-checked repair lesson that lands in durable memory (agent- or workspace-scoped) so future dispatches can retrieve it. |
Feynman repair is the last and most expensive tier, reached only when the cheaper three didn't resolve things — it's the layer that actually tries to understand, not just react.
Two gates a weak explanation can't pass
A candidate explanation has a real shape — what failed, why it failed, what assumption was wrong, what to verify next time, and the distilled lesson itself, typed as procedural or conceptual and scoped to the agent or the workspace. Before any of that is committed, two thresholds have to clear: the explanation must overlap real evidence from the run by at least 0.18 (a grounding check — an explanation that doesn't actually reference what happened is rejected, no matter how plausible it reads), and the model's own confidence in it must be at least 0.5. Fail either gate, and nothing is written — there is no fallback that stores a plausible-sounding-but-unverified lesson anyway.
A queue job, not a per-run tax
Feynman repair never runs on every failure — that would make every ordinary error expensive. It's triggered by exactly three conditions: the cheaper self-heal tiers have been exhausted, the same node has now failed repeatedly across separate runs (three occurrences trips it), or a genuine contradiction was detected. Outside those triggers, it simply doesn't run — an occasional, targeted queue job, not a tax paid on every dispatch.
Continue
Writing memory is reconciliation, not append: deterministic extraction, a two-phase Formation Judge (ADD/UPDATE/NOOP), PACER routing, and Feynman repair.
Hybrid TF-IDF plus embedding recall with freshness decay, trust weighting, MMR diversification, and citation-or-abstain grounded answers.