E1 · Formalizing: audit
Audit what the run actually supports
E1 checks the completed analysis against its records. It separates what can be recomputed from what the aggregator remembers or believes, then assesses whether the model connects population claims to local evidence adequately.
Where this stage starts
The aggregator can inspect the outputs, logs, saved versions and handoffs from the whole run, including D6. Earlier files are preserved. A discovered error becomes a recorded defect or a new correction file, not a silent rewrite of the evidence being audited.
What the aggregator does
Scripts recompute counts, trace quotations and identifiers, and check whether required steps happened in the prescribed order. A separate handler-supplied check compares quoted material with the raw record. The audit distinguishes that check from resolving citations within the already coded store.
The aggregator examines the model’s development: which cases changed successive stories, what readers were given, what category revisions occurred, and which later claims actually use each stage’s outputs. A file that was produced but never used may reveal unnecessary work or a broken handoff.
It then opens the predictions sealed before the return to communities and compares them with the answers. This remains an in-sample comparison because those communities contributed to building the model. Original global fit, current local adequacy and support for shared forms are reported separately, with versions and unresolved evidence.
For a seeded sample of thirty communities, fresh readers assess whether the global model plus supported local adaptations can reconstruct claims in the community account. A poor reconstruction can reflect missing evidence, an omitted local detail, an unsupported claim or a weak shared model. The audit must investigate those causes rather than turn one score into a universal verdict.
Finally, the aggregator writes its experience of the run against the audited record, admitting where it cannot remember. New D6 proposals remain explicitly unvalidated.
What the next stage receives
E2 receives a reproducible audit, the study’s limits and the aggregator’s separate experience account. The composition judgment may be supported, unsupported or unresolved, with its level and scope stated. E2 can then examine the instructions and handling that shaped the work.
What the output files contain
Open a file group for its purpose. These are descriptions of the stage’s outputs. Markdown files hold prose; JSONL files hold one structured record per line.
The audit and reproducible calculationsaudit-E1.md, audit-E1/, claims-E1.md, probes-E1.md
audit-E1.md reports checks, limits and the methods
account. audit-E1/ holds scripts and outputs for replay.
claims-E1.md checks whether each claim received the kind of
evidence it needs: source words or a pattern across cases.
probes-E1.md reports recall checks, distinguishing
forgotten information from questions the record could not answer.
Expectations checked against returnsexpectations-E1.md
expectations-E1.md compares sealed predictions with
community responses, including field-origin claims. It preserves the
distinction between expected absence and something a case could not
test.
The local reconstruction testreconstruction-sample-E1.jsonl, RETURN-E1.md, reconstruction-E1.jsonl
reconstruction-sample-E1.jsonl records the thirty
selected communities before inspection. RETURN-E1.md
specifies the reader task. reconstruction-E1.jsonl
preserves the results and their connection to particular claims.
Experience and the final handoffexperience-E1.md, notes-E1.md, handback-E1.md, run_manifest.json
experience-E1.md is the aggregator’s first-person
account, separate from computed evidence. notes-E1.md
explains audit decisions. handback-E1.md carries problems
and lessons forward. The last run_manifest.json entry
records the experience file and the model that wrote it.
Full stage description
Read the complete technical card for exact tasks, reading boundaries, checks and closing rules. The explanation above is a guide; the card holds the instructions.
How the stage is divided into legs
E1a: scripts. Repeats: no. Interview point: -.E1b: audit section. Repeats: yes. Interview point: -.E1c: reconstruction build. Repeats: no. Interview point: -.E1d: reconstruction check. Repeats: no. Interview point: -.E1e: experience. Repeats: no. Interview point: -.
Source versions and hashes identify the exact technical documents reproduced here.