Campaign friction log

Record entries while the campaign is running. Preserve the earliest actionable failure and distinguish configuration issues, product defects, and model behavior. For each configuration trial, link its version and parent, hypothesis and dated source, intended/effective setting delta, targeted and regression results, final qualification evidence, and next action or stopping reason. A stopped cell is not a rejected model; budget-limited investigation stays unresolved. Use the qualification skill's configuration-search workflow.

Time	Stage	Category	Earliest actionable evidence	Immediate disposition	Durable fix-forward artifact	Independent verification	Status
2026-11-02	finalist	ambiguous-output	strict diagnostic	Exclude strict rows from performance ranking	Preserved adjudication and ordinary-summary limitation	Exact counter reconciliation retained	closed for historical r4 scope
2026-11-02	feasibility	unsafe-default	r5 rejection	Restore r4 .90 and stop memory retuning	R5 rejection receipt	Required GPU reserve violation retained	closed
2026-11-02	quality	bounded-quality-failure	native SWE evidence	Preserve Django10097 as unresolved; do not convert the fixed five-task floor into a broad coding claim	Native report and canonical-retention binding retained	native validator reports no validation errors and official grading complete	closed for fixed campaign floor
2026-11-02	quality	bounded-capacity-claim	C2 result	Preserve observed pair evidence only; do not infer hypothetical full KV residency	C2 amendment retains the scope/claim limit	Requests retain overlapping intervals and engine running=2 samples	closed for the observed pair
2026-11-02	quality	coverage-not-exercised	immutable overlap result	Preserve the incomplete old result; do not treat correct responses as overlap-coverage success	Successor contract contender-stream-active/v1 requires a live anchor delta after contender first output and before end	validator accepts successor unique 3/3 and repeat 3/3; original source and capacity seal remained unchanged	closed for stability, historical failure retained
2026-11-02	quality	source-boundary	observer amendment	Run successor only from the isolated inspector worktree; do not modify original campaign/capacity source	Pinned inspector source revision and clean full-test receipt retained	Full suite clean at 4194 passed / 116 skipped after the initial all-passing run was refused by the added source-change guard	closed for successor source boundary
2026-11-02	quality	review-correction	initial stability review	Preserve the digest-method finding; do not discard the initial review	source-binding correction recomputes the exact run-cell package method and accepts the declared digest	Contemporaneous launch assertions bind isolated CWD/source; no additional replay required for source binding	closed for source-binding provenance
2026-11-02	finalist	soft-performance-miss	fresh capacity summary	Retain C2 at 32K as the concurrency option because first-output p95 remains under its 30-second bound	Record decode p50 13.43 tokens/s below the 15 tokens/s objective; C1 remains latency-first	Fresh 10-cell capacity and resource validation complete	open as an explicit user tradeoff, not a hidden pass

For skill/process lessons, also link the baseline and candidate instruction identity, applicability limits, development and independent-case results, reviewer, observed cost, accepted/rejected/unresolved decision, and expiry trigger. A reflection note alone is not an independently validated lesson.

If there was no friction, retain the file with one sentence stating that no manual workaround, ambiguity, missing identity, unsafe default, repeated command, or actionable failure was observed. A retry closes the incident only when a durable disposition and its independent verification are recorded.