docs: WW-PGD telemetry field guide (scheduled vs applied)

README.md

@@ -97,6 +97,9 @@ These commands make the repository capable of testing whether WW-PGD improves ge
 
 ## WW-PGD installation policy
 
+How to read projection / dose / event columns: [`docs/WWPGD_TELEMETRY_FIELDS.md`](docs/WWPGD_TELEMETRY_FIELDS.md).
+
+
 `python -m pip install -e .` installs WW-PGD through pip from the current default branch of `CalculatedContent/WW_PGD`. The dependency has no commit, tag, or branch pin. Each run records the package version and, when pip supplies PEP 610 VCS provenance, the commit that happened to be installed. That recorded commit is runtime metadata, not an installation requirement.
 
 The nanoGPT adapter targets the public `ww_pgd` API. When the installed package exposes a native `diagnostic_logs` sink, those internal rows are retained. When it exposes only its established `ww_logs` output, the repository installs a compatibility wrapper that preserves normal WW-PGD execution and records the available pre-projection WeightWatcher fields plus adapter-observed candidate movement. Exact internal midpoint, Cayley-ratio, and TraceLog-retraction fields remain explicitly unsupported in that case; they are never invented and their absence does not prevent the experiment from running.

docs/WWPGD_TELEMETRY_FIELDS.md

@@ -0,0 +1,81 @@
+# WW-PGD telemetry fields (how to read projection CSVs)
+
+**Class B documentation.** Helps readers interpret WW-PGD projection / dose /
+event fields without inventing missing quantities. It does not change training
+code or defaults.
+
+For density α vs derived rank exponent vocabulary, see
+[`SPECTRAL_DUAL_LABEL.md`](SPECTRAL_DUAL_LABEL.md)
+(μ_rank vs μ_entry vs MP aspect $q$ — never overload bare `q`).
+
+---
+
+## 1. Why this page exists
+
+CSV rows look like science tables. Misreading them recreates confounds we already
+know from experiment design:
+
+| Misread | Better read |
+|---|---|
+| Projection **event index** = dose | Event index is a **schedule clock**; dose is a **strength** |
+| First scheduled event = first real apply | Warmup, gates, trust region, or zero hardness can skip applies |
+| Requested pin = verified pin | Runtime check and install pin are different facts |
+
+---
+
+## 2. Fields commonly present on `main` (scientific runner)
+
+Exact headers evolve; always trust the CSV header of the run you analyze. Typical
+projection / adapter-related columns include:
+
+| Field (examples) | Role |
+|---|---|
+| `projection_event` | Schedule / event index for the projection cadence |
+| `optimizer_step` / `actual_step` | Base optimizer step when the row was written |
+| `layer_name` | Which matrix the row refers to |
+| `target_alpha` | Public density spectral target |
+| `derived_external_rank_exponent` | Derived rank-order exponent \(1/(α-1)\) at the adapter boundary |
+| `relative_frobenius_change_applied` | Realized relative Frobenius movement of the applied update (when present) |
+| `relative_frobenius_change_requested` | Requested movement before trust-region / scaling (when present) |
+| `projected` / skip reasons | Whether an apply happened and why not |
+
+**Rule:** if a column is missing, treat it as missing — never invent midpoint /
+Cayley / TraceLog internals when the installed WW-PGD build does not expose them
+(see root README adapter policy).
+
+---
+
+## 3. Conflict log and SoT (predicates)
+
+| Intent | Earlier risk | Canonical direction |
+|---|---|---|
+| First-event / first-apply flags | Treating `event == 1` (or 1-based index) as first real apply | Split **first scheduled event** vs **first actual apply per layer** (positive realized dose that changes the layer). Upstream work: open PR discussion **#131** (resume-safe dose telemetry) — prefer that design over ad-hoc first-event booleans. |
+| Optional WW_PGD commit pin | Check only in one runner; soft prefix match; stale import provenance | Verify at **every shared adapter boundary**; distinguish requested vs verified vs dependency-pinned. Upstream work: open PR **#132**. |
+| Dual-label α / rank | Single column named “alpha” | Density target + derived rank exponent as separate fields (#128 logging + [`SPECTRAL_DUAL_LABEL.md`](SPECTRAL_DUAL_LABEL.md)). |
+
+If #131 / #132 are not yet merged when you read this, treat their **predicate
+design** as the intended SoT and mark any not-yet-on-main columns as pending.
+
+---
+
+## 4. Read rules (always)
+
+1. **Scheduled ≠ applied.** A projection event can be scheduled and still apply
+   zero dose (or skip a layer).  
+2. **Dose is applied strength**, preferably relative Frobenius of the **applied**
+   update after trust-region / scaling — not hardness alone.  
+3. **Per-layer first apply** is not the same as global event zero.  
+4. **Pin provenance** has three facts: what is installed, what was requested, whether
+   runtime verification passed.  
+5. **No efficacy claim** follows from a field being present.  
+6. **Dual-label:** density `target_alpha` and `derived_external_rank_exponent`
+   are different columns; do not rename the rank exponent to bare `q` (MP aspect).
+
+---
+
+## 5. Related
+
+- Root README — WW-PGD installation policy; public `target_alpha` only.  
+- `docs/SPECTRAL_DUAL_LABEL.md` — density α vs μ_rank.  
+- `docs/SCIENTIFIC_INTEGRITY_POLICY.md` — no fabricated spectral fields.  
+- `docs/CONTROL_ARMS.md` — measurement-only / sham arms when comparing dose.