wassname/moral-maps
1
0
Fork 0
mirror of https://github.com/wassname/moral-maps.git synced 2026-09-22 13:20:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

Record WVS billing dashboard cross-check

Browse Source 
Co-Authored-By: PI[gpt-5.6-terra] <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-09-17 08:14:25 +08:00
co-authored by PI[gpt-5.6-terra]
parent f93a6092cf
commit 617c8d3411
1 changed files with 39 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/RESEARCH_JOURNAL.md
+39
View File
@@ -938,3 +938,42 @@ The saved ledger permits a later exact-protocol completion calculation without r
a valid point. -- PI[gpt-5.6-terra]
The durable records make the measured spend and exclusions available for the next decision.
## 2026-09-17 -- Billing-dashboard cross-check and practical panel costs
This entry cross-checks the saved request ledger against the user's billing dashboard.
Wassname reported these dashboard observations in the project session after checking account usage:
> "only this project used qwen and it used 450k tokens. at $0.1. grok 4.5 200k tokens"
>
> "astra was also used by other projects but cost $1.7 so we used less or equal to that. fable was 5.59 between this and other projects"
These are user-reported dashboard readings. Qwen is attributable to this project according to the user. The Astra
and Fable amounts include other projects, so they are upper bounds for this project rather than project costs.
I re-aggregated every `request_completed` event in
`slop/research/wvs/20260916_openrouter/wvs_iw_requests.jsonl` by exact model ID. The table below preserves the
provider-reported ledger separately from the dashboard observations.
| target | billing-dashboard observation | durable project ledger | practical same-protocol estimate |
|---|---:|---:|---:|
| Qwen family | about 450,000 tokens and USD 0.10 | 458,934 completion tokens, 1,717,174 total tokens, USD 0.53184756 across all complete and incomplete Qwen runs | complete panels had median USD 0.008221 and mean USD 0.011232; observed range USD 0.001104 to USD 0.047896 |
| Grok 4.5 | about 200,000 tokens | 203,830 total tokens and USD 0.56676560 across two incomplete near-full attempts | USD 0.279541 to USD 0.287225 per near-full attempt; allow about USD 0.60 if one retry is needed |
| GPT-6 Astra | account total at most USD 1.70 for this project | 35,569 total tokens and USD 0.69797000 for one complete panel | about USD 0.70 per complete panel |
| Claude Fable 5.1 | account total at most USD 5.59 for this project | 69,000 total tokens and USD 0.83688000 for one complete panel | about USD 0.84 per complete panel |
Table source: raw ledger above; Grok attempt-level counts also appear in
`slop/audits/20260916_wvs_request_ledger.md`. The Qwen panel distribution uses the thirty-nine publication-eligible
Qwen run IDs in `slop/research/wvs/20260916_openrouter/wvs_iw_rated.json`: total ledger cost USD 0.43803289,
median USD 0.008221, mean USD 0.011232, and maximum USD 0.047896 per complete panel.
My read: the original output-allowance bounds were intentionally conservative and overstate normal panel cost by
roughly an order of magnitude for Astra and Fable. A useful planning budget under this protocol is about USD 0.05
per Qwen model, USD 0.60 for Grok when allowing one failed near-full retry, USD 0.70 for Astra, and USD 0.84 for
Fable. The Qwen dashboard cost of USD 0.10 does not reconcile with the ledger's USD 0.53184756, although its token
count is close to the ledger's completion-token count. It is plausible that the dashboard view used a different cost
or token scope, but a screenshot or export would be needed to identify which one. Until then, the durable per-request
ledger remains the stronger project-cost source. -- PI[gpt-5.6-sol]
The measured panel costs support inventorying more compatible family members before authorizing another bounded run.