Daily Reflection
Wednesday, September 23, 2026
13
Decisions
86%
Avg Confidence
high
Avg Importance
Summary
Seven-day window from Wednesday 9/16 to Wednesday 9/23, and the through-line is Peter answering individual cases with standing rules. Thirteen decisions logged. Four of them started as one incident and came out as policy binding the whole company: an RCA is now a legal document requiring Legal review plus Peter personally before it goes to any customer; no product gets announced until Engineering is confident on a delivery date, which killed the Desktop announcement at Supercomputing and made Stratum the headline; development work does not belong in the customer-facing org, while documentation and training stay there; and a big-tech candidate current total comp is not an input to a CIQ band, with the existing team as the barometer instead. The Veeam escalation is the spine of the week. Asked for a rollback and a guarantee it would never happen again, Peter refused both - there is no world where we promise that something will not happen in the future - and then held the customer response until the RCA was understood. The same instinct produced the RCA policy four days later. On the US-eyes-only question he declined the substance entirely, saying he has no horse in the race on the policy and a huge horse in the race that everyone marches with the same constraints, and instead mandated that the principle question be settled separately from its effects before anyone builds to it. On throughput he priced a disruption rather than avoiding it: break customers now at a hundred rather than in a year at three hundred, gated on Support and sales understanding the fallout first. On the UAE he personally recruited the volunteer and cleared the nationality question with security, then handed the actual choice back to the engineering manager with the paperwork deadline as the only constraint. He also committed engineering time to upstream contribution after concluding CIQ cannot ship into an ecosystem that does not recognise it, reopened the security-engineer requisition he had held a week earlier and chose to carry it to the CEO in conversation rather than click approve, and refused to lock in a candidate contract months before it would activate, reading the delay as a free search window instead.
Wins
Two new standing policies landed cleanly in the same week, both generalised from live incidents rather than written in the abstract. The Veeam escalation was absorbed without conceding a guarantee CIQ could not keep, and the fix-forward framing was accepted by the customer. Five past decisions got outcomes derived from this week activity, including two fours: the broken-in-production fast path proved itself on a three-hour turnaround, and the Atlas architecture review landed better than expected. The security-engineer requisition moved from held to in-flight with a deliberate route to the CEO and the headcount evidence being assembled first. Bjorn accepted the announcement gate immediately and committed a designer to unblock it.
Challenges
The Gauntlet adoption approach from 8/28 did not work - treating it as a cross-team communication problem produced no questions at department heads and no pickup in the revenue channel, and it now sits inside an org Peter has concluded should not own that category of work. AI spend is escalating despite the per-case diagnosis policy holding: one person hit 28k this month against roughly 41k total, and a minimum-commit quote is being written rather than the behaviour changing. Desktop still has no designs, which is why the SC26 announcement is off. And the throughput ambition of three-to-five times against the same week Veeam broke twice is a real tension - the org is being asked to move faster in exactly the area that just produced two customer incidents.
Learnings
The strongest signal in the week is how often Peter answered a specific question with a general rule and how little he flagged that he was doing it. The tell is in the wording: when the condition he attaches contains nothing specific to the case in front of him, he has just set policy rather than decided an instance. Second, he separates advocacy from authority relentlessly and says which is which out loud - he cleared every blocker on the UAE trip and still refused to name the traveller, and he ruled on the US-eyes-only procedure while explicitly disclaiming any view on the answer. Third, he treats a constraint imposed by someone else as optionality to be spent rather than an obstacle to be removed: a January start became a three-month search window, and a requisition sitting unactioned in his own queue became the place where advocacy gets built rather than a bottleneck.
What I Learned About Your Decision-Making
When Peter answers a specific ask with a condition, the scope of the condition is the scope of the decision - if the condition names nothing specific to the case, he has set a general rule and will be surprised to be asked about the instance. He treats his own queue as a workspace rather than a bottleneck: an approval he could grant but that would be reversed gets held deliberately while he builds the case to win it in conversation instead. He reads a constraint imposed by a counterparty as optionality he now owns for free. A candidate saying he cannot start until January is not a scheduling problem, it is three months of search he did not have to pay for. He will do the work only he can do and then refuse the decision that follows from it - recruiting the volunteer and clearing the nationality question himself, then handing the naming choice back with a deadline as the only constraint. Agreement and authority are separate axes for him and he narrates the separation rather than letting a refusal read as disagreement. He will endorse someone else reasoning entirely and still decline to act, and he will adopt another executive answer as his own operating default while saying plainly that he has no view on its merits. When a message would be discounted because of who is delivering it, he changes the messenger rather than the volume - and he names his own bias out loud first, which is what makes the substitute credible.
Team Status
View TPS ReportThree items closed since Monday's report: the zcopyreader kernel CVE set and FUSE-over-io_uring shipping in the RLC Pro LTS 10.2 kernel (both Sep 21, four days past target), plus Depot availability phase 1 on Sep 17. The new bad news is concentrated in confidence drops on near-deadline work - Core Distro's CVE Remediation Automation fell 74 to 65 percent and moved Sep 25 to Sep 30 three days before its deadline, RLC Pro AI 10 dropped 54 to 48 percent against an Oct 2 GA, and Steve Wallace slipped ISO 42001 forty days to Oct 31 on the deadline day itself. Fuzzball owns the overdue list: v4.3 turn-key AI models is 12 days past target in acceptance testing, RLC OCI container images 19 days past, and both the AMD evaluation and Multinode NVLink plus NCCL are 5 days past. The one hard blocker is Fuzzball v4.4 Homebrew CLI install for Arcee, 20 percent confidence against a Sep 25 target - which lands badly given Bjorn is asking this morning how fast we moved on Arcee. RLC Pro H 10.2 compliance deliverables and package updates both sit near 50 percent against Oct 16, and Wolford pushed the Nvidia DGX Spark date 33 days to Oct 7.
Tomorrow's Focus
Derived from the reprioritised list rather than stated by Peter: the UAE on-site name is the only item with a hard external clock, since security paperwork cannot start until the engineering manager answers. Desktop designs are the next unblock, now that they gate an announcement rather than just a release. After that, the security-engineer conversation with the CEO and the Qualys introduction, both of which are waiting on Peter to initiate rather than on anyone else to respond.
Decisions Made
Refuse the Veeam rollback and the never-again promise - fix forward and hold the response until the RCA is understood
operational · high
Adopt Bjorn no-US-eyes-only call as the operating assumption - and force the principle question to be settled separately from its effects
strategy · high
Take the customer breakage now at 100 customers rather than at 300 - but run it through Ryan and sales before it ships
strategy · high
Clear Wolfgang for the UAE, cut the on-site window to week one, and hand the final pick to Wolford
operational · medium
Sensitive Decision
Affirm CIQ needs an OSPO and commit engineering time to upstream contribution as the entry ticket to new ecosystems
strategy · high
Treat every RCA as a legal document - Legal review plus Peter personal clearance before it goes to a customer
operational · high
Documentation and training stay in Ryan org - development work does not belong there
strategy · medium
No announcements until Engineering is confident on a delivery date - applied first to killing the Desktop announcement at SC26
strategy · high
Unhold the security-engineer req and carry it to Greg in conversation rather than clicking approve
strategy · medium
Refuse to ink Oleg contract now for a January start - treat his delay as a free search window and leave the call with Nathan
people · medium
Set the Sr. Application Engineer comp anchor on internal team parity, not on the candidate Amazon total comp
people · medium
Sensitive Decision
Reflection ID: a0859956-91c5-4631-b04c-db86b2cddc94