EmmanuelRTM Evidence tagging
Every factual sentence you write into paper/ or final/ carries at least
one inline evidence tag binding it to a run artifact:
| Tag | Grounds | Example |
|---|---|---|
[EV:score:<path.json>#<dotted.key>] |
a number | Accuracy reaches 0.9852 [EV:score:best/eval.json#score]. |
[EV:log:<path>:L<n>[-L<m>]] |
an event/observation | Training converged without errors [EV:log:best/solve.log:L2]. |
[EV:cite:<key>] |
prior work | kNN error is bounded by twice the Bayes error [EV:cite:cover1967nearest]. |
[EV:artifact:<path>] |
existence/availability | The full implementation is included [EV:artifact:best/solution.py]. |
[EV:config:<path.json>#<key>] |
a setup fact | We ran 5 parallel branches [EV:config:run-config.json#branches]. |
Rules:
- Paths are run-directory-relative. The tag must resolve —
ground_check.pyopens the artifact and compares. - Restate numbers exactly as they appear in the artifact (rounding within 0.5% and %-vs-fraction conversion are tolerated).
- One tag per fact; a sentence with two facts gets two tags.
- Cannot find an artifact for a statement? Then it is an assumption: move it
under a
## Assumptionsheader, or delete it. Never leave an untagged factual sentence in the body — GROUND flags it UNSUPPORTED. - Citation keys must already exist in
bibliography.jsonl; a hook blocks unknown keys. You may not add bibliography entries yourself. - Failed experiments are evidence too — tag them with their branch's eval.json/audit.md and report them honestly.