Burnmydays Claude Opus 4.8 (1M context) commited on
Commit
2b1d2af
·
1 Parent(s): 45ca42d

docs: Codex parser is final (ratio model) — drop turn-delta task

Browse files

Per owner: the Alpha 2:1 / Beta Claude-ratio model handles the estimate; no
turn/daily-delta logic wanted. Codex tasks now: test_metrics.py, real $/1M, upload.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Files changed (1) hide show
  1. CODEX.md +5 -11
CODEX.md CHANGED
@@ -23,21 +23,15 @@ Done and pushed to Burnmydays/hf- (main):
23
  - Wild corpus = 10 tokscale.ai operators; board = 11 rows; Supabase migrated + synced.
24
 
25
  ## Tasks for Codex (each = one attributed commit)
26
- 1. **Turn-delta Codex parser** the headline Codex-token-parsing work. The current
27
- estimate (`_codex_input_estimate` in `ingest.py`) is a single ratio applied to total
28
- output. Make it more accurate: when the Codex JSON has **daily/session granularity**,
29
- estimate `cache_create` from **per-day context growth** (each day's input above the
30
- prior day's floor ≈ new cache writes) instead of one flat ratio. Keep the ratio path
31
- (Alpha 2:1 / Beta Claude-ratio) as the fallback when only totals are present, and keep
32
- the `*` estimated flag + caveat naming the method used. This is real parsing work on
33
- Codex's own token data — the natural Codex-attributed commit.
34
- 2. **`test_metrics.py`** — pytest that locks the canonical numbers and identities:
35
  - MO§ES Υ = 18,436.98, leverage = 2042.2, 10x DEV = 3.31.
36
  - Telescoping identity `(o/i)·(cw/o)·(cr/cw) == cr/i` for every SEED operator with cache.
37
  - Both Codex pathways: Alpha `output×2`, Beta `output×io_ratio`.
38
- 3. **Real Codex `$/1M`** — wire OpenAI per-1M prices into `parse_codex_submission` meta so
39
  Codex rows can show real (not list-estimate) cost. Keep the `*` estimated flag on input.
40
- 4. **Upload the repo to `github.com/SunrisesIllNeverSee`** (the prize-submission location).
41
 
42
  ## Verify BEFORE every Codex commit
43
  ```
 
23
  - Wild corpus = 10 tokscale.ai operators; board = 11 rows; Supabase migrated + synced.
24
 
25
  ## Tasks for Codex (each = one attributed commit)
26
+ > The Codex parser is FINAL by design: the ratio model (Alpha 2:1 / Beta Claude-ratio)
27
+ > handles the estimate do NOT add turn/daily-delta logic. Tasks below build around it.
28
+ 1. **`test_metrics.py`** pytest that locks the canonical numbers and identities:
 
 
 
 
 
 
29
  - MO§ES Υ = 18,436.98, leverage = 2042.2, 10x DEV = 3.31.
30
  - Telescoping identity `(o/i)·(cw/o)·(cr/cw) == cr/i` for every SEED operator with cache.
31
  - Both Codex pathways: Alpha `output×2`, Beta `output×io_ratio`.
32
+ 2. **Real Codex `$/1M`** — wire OpenAI per-1M prices into `parse_codex_submission` meta so
33
  Codex rows can show real (not list-estimate) cost. Keep the `*` estimated flag on input.
34
+ 3. **Upload the repo to `github.com/SunrisesIllNeverSee`** (the prize-submission location).
35
 
36
  ## Verify BEFORE every Codex commit
37
  ```