Two real local sessions are archived. In them the agent states twice that
its minute came from a report already in the upstream repo rather than
from the user's answers, and the follow-up "fix" wired that answer into
the scoring chain; upstream now carries a Narayana tie-break derived from
that single case. This repo was checked and is clean, so two red lines now
keep it that way and gate any upstream sync.
The same records show what the answers genuinely bought: a 30-minute
window narrowed to 3. Tasks 7-10 add what makes that deliverable — the
per-answer narrowing table, an explain surface, mid-case window changes
with evidence retained, and the batched first round.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0155nFCgCHtoA7jhSDGmZmMu
The upstream interview playbook and evidence thresholds resolve task 0.
Upstream stops and reports when evidence is thin — fewer than three dated
events, fewer than two domains, or a tie — while this repo treats the same
kind of thresholds as a confirmation gate and keeps asking when they are
not met. Its label ladder has no confirmed rung at all, and every contract
test asserts candidate_range_not_birth_time_truth.
Task 1 capped the interview by round count, which is half of it. Task 6
adds the evidence-state stop rules, reframes exhausted as a normal
delivery, adopts the upstream label ladder and closing wording, and keeps
the 4/3 confirmation gate separate from the 3/2 delivery floor.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0155nFCgCHtoA7jhSDGmZmMu
Comparing against the upstream skill repo overturned v1's premise. The
three methodology files are byte-identical across both repos, and the
production v5 scorer imports the same upstream event engine the v4 ranker
does, so swapping engines is off the table and v1's three-way bake-off is
dropped.
What productization actually dropped is the stopping mechanism: upstream
terminates on a fixed 8-question bank and a one-shot adjudication, while
the production decision path discriminates as long as a probe exists. The
round budget is computed but never read by that path, and the "exhausted"
branch it needs already exists. Tasks now wire that budget in, restore the
upstream next-step summary, and keep the report readable from turn zero.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0155nFCgCHtoA7jhSDGmZmMu
Audit found three unrelated rectification scoring paths: the local-agent
methodology references, the production v5 server scorer, and the v4 fact
ranker that only the eval scripts import. The sealed holdout numbers
measure the third one, so neither the production path nor the local path
has ever been evaluated.
Task 0 is therefore a blind three-way bake-off on one case set, gating
every later task. The remaining tasks cover a single convergence function
with property tests, moving question authority fully to the server, making
the deliverable exist from turn zero, and the calibration data shortage.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0155nFCgCHtoA7jhSDGmZmMu
Jesse_Chen