# How to read evidence

**Audience:** reviewers and agents reconciling conflicting docs.  
**Evidence state:** 2026-07-16.  
**Canonical numbers (public codes):** [NUMBERS_CANON.md](NUMBERS_CANON.md)

No GPU launches required for audit. Prefer reading seals + CSVs.

---

## Resolution order (when docs conflict)

| Priority | Source | Trust for |
|:--------:|--------|-----------|
| **1** | Sealed Jul 15–16 claim-stack: `MULTI_SEED_CLAIM.md`, `SEED*_SUMMARY.md`, `F1_BCAROL_SUMMARY.md` (F1_Claim-W2), `results/f1_seal/SUMMARY.md` | Current claim status and headline numbers |
| **2** | Pulled CSVs under `results/claim_stack_v1/seed2029/`, `seed4242/`, `results/f1_seal/log.csv` | Recompute first-cross \(T\) |
| **3** | Public site https://nocap.think2earn.com + this folder’s narrative markdown | Presentation; may lag seals |
| **4** | `docs/ops/FAIR_CLOCK.md`, older ledgers, component `STACK.md` | History / policy origin — may be partially stale |
| **5** | `nocap-repo/README.md` dated ~2026-07-12 (“no win”) | **Stale for claim status** — pointer only |

**Rule:** newer sealed SUMMARY beats older README / early same-account-only fair-clock wording for **what is currently claimed**.

---

## What is sealed truth (start here)

Paths under workspace root `/home/q/Downloads/software_i_built/modal/` unless noted.

### Headline claim

| File | Role |
|------|------|
| `nocap-repo/results/claim_stack_v1/MULTI_SEED_CLAIM.md` | **3/3** table + defensible language |
| `…/SEED1337_SUMMARY.md` | Seed 1337 matched pair |
| `…/SEED2029_SUMMARY.md` | Seed 2029 matched pair |
| `…/SEED4242_SUMMARY.md` | Seed 4242 candidate-only + 80→40GB note |
| `…/PREREG.md` | Recipe freeze |

### Denominators

| File | Role |
|------|------|
| `nocap-repo/results/f1_seal/SUMMARY.md` | Public F1_seal (\(T\)=3.866 h) |
| `nocap-repo/results/f1_seal/log.csv` | First-cross row audit |
| `nocap-repo/results/claim_stack_v1/F1_BCAROL_SUMMARY.md` | F1_Claim-W2 (\(T\)=3.858 h) |

### Local CSV pulls (recompute)

| Path | Notes |
|------|-------|
| `…/claim_stack_v1/seed2029/cand_log.csv` | Candidate 2029 |
| `…/claim_stack_v1/seed2029/stock_log.csv` | Stock 2029 (no cross) |
| `…/claim_stack_v1/seed4242/cand_log.csv` | Candidate 4242 |
| Seed **1337** CSVs | Often **absent** locally — use sealed SUMMARY hours until pulled |

Full path index: [EVIDENCE_PATHS.md](EVIDENCE_PATHS.md).

---

## Recompute first-cross from CSV

1. Open `log.csv` (or `cand_log.csv` / `stock_log.csv`).
2. Find columns `val_disjoint` (or equivalent) and `train_time_ms`.
3. Select the **first row** with `val_disjoint ≤ 3.3821`.
4. \(T_{\mathrm{h}} = \mathrm{train\_time\_ms} / 3.6\times10^6\).
5. \(\Delta T\% = 100 \times (3.866 - T_{\mathrm{h}}) / 3.866\).
6. Round for publication the same way seals do (examples below).

### Known recomputes (do not invent others)

| Run | First-cross ms | \(T_{\mathrm{h}}\) precise | Published |
|-----|---------------:|---------------------------:|-----------|
| F1_seal | 13,918,667.5 | 3.866297… | **3.866 h** |
| Cand 2029 | 12,840,827.6 | 3.566897 | **3.567 h** · **+7.7%** |
| Cand 4242 | 12,640,470.5 | 3.511242 | **3.511 h** · **+9.2%** |
| Cand 1337 | (SUMMARY) | — | **3.553 h** · **+8.1%** |
| F1_Claim-W2 | 13,889,386.9 | 3.858163… | **3.858 h** |

If stock has **no** row ≤ 3.3821 through final step → first-cross = **NONE** → matched ΔT% = **N/A**.

---

## Chronology traps (not current truth)

| Trap | Correct reading |
|------|-----------------|
| README “official speedup: still no win” (Jul 12) | Superseded by Jul 15–16 **3/3** seals; still true that there is **no official organizer win**. |
| Docs saying “2/2 seeds sealed” | Written before seed 4242; current = **3/3**. |
| “2 / 2 no cross” in failure ledger | Means **two paired stock arms** missed — not “2/2 candidates.” |
| `STACK.md` “not an under-baseline claim” | True for **2026-07-14 component bank**; full seals landed later. |
| Early FAIR_CLOCK “cardinal % only on Clock-W” | Policy origin; sealed campaign publishes vs **F1_seal** while candidates ran on Claim-W1/W2, with F1_Claim-W2 as cross-check. Task hard rules + `MULTI_SEED_CLAIM.md` win. |
| Mean **3.560 h / 7.9%** | Old 2-seed mean; current mean ≈ **3.544 h / +8.3%**. |

Residual conflict list: [ACCURACY_NOTES.md](ACCURACY_NOTES.md).

---

## Explicit non-claims (audit checklist)

- [ ] Page/docs do **not** say official NoCap win  
- [ ] No matched first-cross ΔT% vs joint stock as a number  
- [ ] No % vs historical F1 (~229.1 min)  
- [ ] Candidate vs **F1_seal**, not vs joint stock, for headline %  
- [ ] Seed 4242 noted as candidate-only / 40GB seal  
- [ ] Workspace **codes** only in public text  

---

## Hosting note (markdown vs live site)

Production currently deploys **`index.html` only** to Cloudflare Pages. Companion `.md` files in `cf-pages-nocap-agent-dev/` are the durable public doc set for local preview and for a **future deploy that includes those files**. Until then, live strangers see the HTML report; agents/reviewers with the workspace read these markdown files directly.