# Sample-level ground truth by counterfactual replay

One file per latent corpus and grain,
`<instance>/latent/seed=<k>/seqtruth-forced-ok-counterfactual-<request|session>.json`:
25 corpora × 2 grains = 50 files. `index.json` lists every file with its size, its sha256 and its counts.

The files are derived from the latent corpora under `instances/` (tool version 0.3.0). They sit outside the corpus
folders, so every corpus still verifies against its own `manifest.json`. Licence: CC BY 4.0, as the corpora.

## What a file holds

The truth edges of the sample-level graph of every sampled test trace of one corpus at one grain. A truth edge joins two
failed attempts of one request. It is decided by counterfactual replay: the simulator re-runs the request with the earlier
attempt set to ok and every random draw unchanged. The pair is a *but-for* edge if the outcome of the later attempt
changes. It is a *contingent* edge if it is not a but-for edge and the outcome changes once the other failed callees of
the same caller are set to ok in both runs. At the session grain the client's error record of a request is also a target
of that request's BFF attempt. Pairs whose target succeeded, pairs across requests, and pairs between two client attempts
of one journey step are not scored.

The traces are the test sample of the reference evaluation (`read` names the run): the test split in completion order
(`ordering: end`), at most `max_len` tokens per trace, pairs at most `max_lag` positions apart.

## Format

Header:

- `corpus_id`, `grain`, `split`, `read`, `ordering`, `c`, `max_len`, `max_lag`: the corpus and the sample.
- `counts`: `sequences` (sampled traces), `sequences_with_candidates` (traces with a scored pair), `sequences_with_truth`
  (traces with a truth edge), `candidates` (scored pairs), `truth_directed` (truth edges). The `*_landed` counts refer
  to a truth read from the recorded parent of each span; they are kept for comparison and are not the truth of this file.
- `directed_pairs`: the truth edges by kind (`single` = but-for, `contingent_only` = contingent; by source; by whether
  the call is critical; by whether the two event types are an edge of the population-level summary graph).
- `dropped_candidates`: pairs left out of the scored pairs (`ok_target`, `client_retry`, `unmapped`).
- `mapping`, `reruns`, `rerun_pairs_not_scored`, `old_against_new`: bookkeeping of the build. `oracle_check` gives, per
  re-run shard, the number of client records, hops and sessions that were compared with the corpus's oracle tables.

`sequences`, one entry per sampled trace:

- `seq`, `trace_id`: the trace's index in the sample and its identifier in the corpus's views.
- `length`: positions of the stored sequence: a start marker, the trace's tokens (at most `max_len`), an end marker.
- `healthy`: true when the trace has no failed attempt.
- `req`: one entry per window index `t = 0 … length − c − 1`, the request identifier (as in the corpus's oracle tables)
  of the token at that index; `null` for the end marker and for a position that no request could be assigned to.
  Window index `t` is token number `c + t − 1` of the trace, counting from 0.
- `pairs`: lists in the order of `pair_fields`:
  `j`, `q` (window indices of the source and the target, `j < q`);
  `column` (`directed`: a truth edge; `ancestor`: another failed attempt of the request that changes under the same
  replay, not a truth edge);
  `single` (true: but-for); `contingent` (true: the target changes with the caller's other failed callees set to ok);
  `source` (`final`: the callee's last attempt; `attempt0`: a first attempt that was retried; `edge`: the BFF attempt,
  with the client's error record as target);
  `critical` (whether the call is critical; `null` for the client pair);
  `token_j`, `token_q` (the two tokens, so that the indexing can be checked).