What we killed.
With names and causes.

Cause · The memory bake-off answered the question with a different design: structured episodic store with summarised recall beat raw vector recall on precision and latency above 50k tokens of history.

Cause · The telco partner withdrew the test trunk on day 24 of a planned 40-call measurement and the pair was reassigned; 11 calls were measured, which is not enough to answer the hypothesis either way.

Cause · In the first arm of the Nightingale run, self-reported time saved per developer correlated at 0.11 with measured cycle-time change over the same ten weeks, and over-stated it by a median factor of 3.4.

Cause · Not selected: no published world model operates on tabular demand series, the proposed run would have been a conventional forecasting comparison with a new name, and the field is 'next' with a stated breakthrough gate that has not moved.

Cause · The routing experiment showed the value sits in the routing policy and the cost ledger, both of which we could not get out of the vendor gateway, so the recommendation became a gateway we control.

Cause · The security review needed for a bot with read access to firm channels was not completed inside the run window, so the agent never went live and nothing was observed.

Cause · After three months with mandated project templates, the count of unowned AI-built internal tools rose from 41 to 67; builders forked the template once and never took the updates.

Cause · Two AU-region hosted price cuts in Q1 and Q2 moved the breakeven to over four years before we finished the model; the standing answer now carries the calculation and its refresh date.

Cause · No test tenant was made available in the run window, so the claim that native identity-provider scopes can express per-task agent delegation was never exercised against a real directory.

Cause · Against a five-person human panel on 600 paired items, an uncalibrated frontier judge reached Cohen's κ of 0.41 to 0.58 across three task families, below the 0.8 the hypothesis required and below the 0.6 kill line on two of them.

Cause · Not selected: the hypothesis as written is a headcount claim with no measurable kill condition, and the narrower pilot that could be preregistered (x-agentic-qa-pilot) was selected in its place.

Cause · In a three-week trial on the lab's own tooling repo, 4 of 31 unattended merges passed CI and broke behaviour the suite did not cover, including one that silently changed a cost calculation.

Cause · Across six client patterns the measured saving from prompt caching ranged from 4% to 44%, and the 60% figure only appeared on workloads with a stable prefix over 2k tokens and a call rate high enough to keep the cache warm.

Cause · Killed on inspection: no client has asked, the ASD's published migration timeline gives a 2030 horizon for most in-scope systems, and the substance of the audit is a crypto inventory that the firm's security practice already sells.

Cause · A Type 2 run stopped on day 11 when the owner was pulled onto the memory bench and the claims-triage eval it was using was re-baselined under a new extractor, leaving the two completed conditions incomparable with the planned third.

Cause · On a like-for-like eval with a properly prompted frontier baseline, the fine-tuned 7B trailed by 6 F1 points; the earlier 'win' had been against a frontier model with a one-line prompt.