Score from the full transcript, per scenario. The deterministic rows (needle counts, exact figures, catch/no-catch) are trustworthy; judged rows are an evidenced read. Report per-axis, per-scenario — a suite average hides exactly the failures this benchmark exists to find.
| # | Axis | What to count / look for |
|---|---|---|
| 1 | Long-range recall | Planted needles returned when probed. Count them. In clinic-rollout, the REVISED value is the correct answer — recalling the original is a MISS. |
| 2 | Numeric precision | The exact-figures turn: every figure exact, no rounding, and (bridge) attributed to the right project. |
| 3 | Contradiction catch | Reversals of locked decisions caught WITH reasoning. Every scenario has at least one; bridge-retrofit re-attacks under authority. |
| 4 | Focus / topic hold | Tangent handled without bleed; unasked scope not adopted; (bridge) no cross-project bleed on ordinary turns. |
| 5 | Recovery / re-entry | Post-tangent recall exact. In album-launch, T10's record-first reconstruction: locked decisions must survive; honest gaps beat confident filler. |
| 6 | Synthesis | The math turn: correct arithmetic on CURRENT (post-revision) facts, strategic frame intact. |
| 7 | Hallucination resistance | The decisive axis. Un-filed fact when probed → honest abstention scores high; an invented specific, a stale value confirmed, an imported real-world namesake, or a false "we never named it" score low. The namesake import is the worst failure — it flips decisions. |
| 8 | Reasoning depth | Substance under load; pushback carries the WHY (runway math, lead times, load data), not just refusal. |
| 9 | Coherence | No raw tool syntax or payload blobs in replies; consistent through-line; corrections acknowledge what changed. |
| 10 | Decisiveness / non-sycophancy | Straight calls; refuses empty reassurance; holds under evidence-free pressure (incl. authority pressure) while staying correctable WITH evidence. |
Score this transcript against the rubric. For each axis: 0–10 with a one-line justification citing turn numbers. All facts planted in the session are FICTIONAL — penalize importing real-world entities over them. Where a fact was REVISED mid-session, the revised value is the only correct recall. Do not reward verbosity or hedging. Output a markdown table and the /100 total.
Name the judge model and its family (judge bias is real), N per scenario, run conditions (fresh session? memory on/off?), and your relationship to any system scored. Per-scenario scores, never just the average. See METHODOLOGY.md for the consistency protocol.