Checkpoint taken: 08/09/2026--18:50:47:557 AEST
Flagged from a saved BDO/tamsulosin/meth harm-reduction transcript where Claude asked repeated check-in questions with no actual passive time-tracking or monitoring behind them — Claude has no way to sense elapsed real time or detect, unprompted, that it's repeating a prior suggestion across a long session.
Grounding case: the dev-vm-network-slowdown saga, where months of software-layer checks (disk, CPU, filesystem, swap, LVM) were exhausted before the real cause — WAN RTT/packet loss, a physical-layer issue — was found. Used as the reference example of a diagnostic loop with no built-in signal for when to abandon one explanatory domain and jump to another.
One row per troubleshooting reply:
[stamp] | Elapsed: Xm Ys | Problem: <short> | Action: <fix or debugging request>Every single troubleshooting reply: a full sweep of the entire log so far (not just the newest row against the list), checking both columns for duplicates. No time or count threshold — always on, during troubleshooting.
Borrowed from the R15 branch-table convention (return code = byte-offset into a 4-byte-wide jump table):
| RC | Meaning |
|---|---|
| 0 | All clean |
| 4 | Info / not applicable — not troubleshooting, or check skipped |
| 8 | Error — a duplicate Problem or Action detected |
| 16 | Catastrophic — a previously-resolved Problem has recurred |
Mandatory behavior on RC 8 or RC 16: do not generate another same-domain guess. Stop, surface the log plainly, and hand the decision to Andrew — including the option of deferring to a different AI model for a fresh perspective, since a different model's different training priors are more likely to break the same blind spot.
Applies only during genuine troubleshooting exchanges (fault → symptom → fix-iteration), judged turn-by-turn rather than toggled by a fixed switch. Claude states plainly when it judges logging has started or ended, and corrects on request if the classification is wrong.
p-ALT — "pseudo Artificial Lateral Thinking." Deliberately named pseudo: this is a mechanical trip-wire that forces a frame-change regardless of whether any actual insight occurs, not a claim that real lateral thinking is happening. A prosthesis for the behavior, not the cognitive event itself.
Whether there should be a grace period before logging starts (Andrew described a pattern of "no checks for the first 20–60 minutes, then it kicks in") versus logging from row one, always. Current default: logging from row one, no grace period — Andrew's 20–60-minute point is tentatively read as an observation about past session behavior rather than a spec requirement, but this has not been explicitly confirmed either way.
Run the protocol against a real past troubleshooting case Andrew will revisit (not yet selected), rather than a constructed scenario — designing a synthetic case where the true cause is hardware-layer but only software-layer symptoms are queryable is itself hard to build well. All relevant conversation transcripts are saved, so the eventual test case is a real artifact, not a reconstruction from memory.