Energy/CO2 factors are keyed by model prefix in IMPACT_FACTORS
(metrics_snapshot.py:100-105), longest prefix first, matched with startswith() (metrics_snapshot.py:339-342). Only greenpt/* models have
fitted values. Every other model falls through to the single flat IMPACT_PROXY (metrics_snapshot.py:106, applied at :343):
A single flat rate cannot represent differences that the code's own data shows
are large. Within GreenPT alone, energy per token varies ~3x (0.36 vs 1.08 Wh per
1k output tokens) and grid intensity ~8x (23.6 vs 193 gCO2e/kWh). Collapsing
every non-GreenPT provider to one constant loses at least that much variation.
The clear first case is Regolo: the portal markets it as EU-hosted, zero data retention, 100% renewables (app/main.py:2504), yet it is
assigned 250 gCO2e/kWh — higher than every measured GreenPT value, including
GreenPT's own 193 and 23.6. The provider the portal labels greenest is modelled
with the dirtiest electricity in the table. That is backwards and is the kind of
inconsistency a member could spot by comparing the badge to the chart.
Secondary case: Tinfoil is TEE-architected (app/main.py:2502), which
carries real compute overhead that the flat proxy cannot capture in either
direction.
Affected models currently include (app/main.py:2506-2515, :3253): tinfoil/deepseek-v4-1-flash, tinfoil/glm-5.3, tinfoil/gpt-oss-120b, regolo/apertus-70b, regolo/qwen3.5-122b, publicai/apertus-v1.5-8b.
Proposed scope
Add IMPACT_FACTORS rows for the main non-GreenPT providers, starting with regolo/* and tinfoil/*.
Use the same fitting approach as GreenPT where a provider discloses impact;
otherwise a reasoned, clearly-labelled estimate. Keep the existing provenance
labels so each row states how it was derived.
For Regolo, source the grid intensity from its actual hosting (EU renewables)
rather than the generic EU average; for Tinfoil, account for TEE overhead.
This is a data edit in metrics_snapshot.py (per the note at metrics_snapshot.py:93-94), not a code change.
Acceptance criteria
Regolo models no longer share the generic 250 gCO2e/kWh proxy.
Each new factor carries a source label stating its provenance.
The metrics documentation lists which providers are measured vs estimated.
References
metrics_snapshot.py:93-106, :339-343
app/main.py:2502-2515, :3253
co-op/support#26
## Context
Energy/CO2 factors are keyed by model prefix in `IMPACT_FACTORS`
(`metrics_snapshot.py:100-105`), longest prefix first, matched with
`startswith()` (`metrics_snapshot.py:339-342`). Only `greenpt/*` models have
fitted values. Every other model falls through to the single flat
`IMPACT_PROXY` (`metrics_snapshot.py:106`, applied at `:343`):
```
wh_per_1k_out = 0.5 gco2e_per_kwh = 250.0 source = ESTIMATED
```
## Problem
A single flat rate cannot represent differences that the code's own data shows
are large. Within GreenPT alone, energy per token varies ~3x (0.36 vs 1.08 Wh per
1k output tokens) and grid intensity ~8x (23.6 vs 193 gCO2e/kWh). Collapsing
every non-GreenPT provider to one constant loses at least that much variation.
The clear first case is **Regolo**: the portal markets it as
`EU-hosted, zero data retention, 100% renewables` (`app/main.py:2504`), yet it is
assigned `250 gCO2e/kWh` — *higher than every measured GreenPT value*, including
GreenPT's own 193 and 23.6. The provider the portal labels greenest is modelled
with the dirtiest electricity in the table. That is backwards and is the kind of
inconsistency a member could spot by comparing the badge to the chart.
Secondary case: **Tinfoil** is TEE-architected (`app/main.py:2502`), which
carries real compute overhead that the flat proxy cannot capture in either
direction.
Affected models currently include (`app/main.py:2506-2515`, `:3253`):
`tinfoil/deepseek-v4-1-flash`, `tinfoil/glm-5.3`, `tinfoil/gpt-oss-120b`,
`regolo/apertus-70b`, `regolo/qwen3.5-122b`, `publicai/apertus-v1.5-8b`.
## Proposed scope
- Add `IMPACT_FACTORS` rows for the main non-GreenPT providers, starting with
`regolo/*` and `tinfoil/*`.
- Use the same fitting approach as GreenPT where a provider discloses impact;
otherwise a reasoned, clearly-labelled estimate. Keep the existing provenance
labels so each row states how it was derived.
- For Regolo, source the grid intensity from its actual hosting (EU renewables)
rather than the generic EU average; for Tinfoil, account for TEE overhead.
- This is a data edit in `metrics_snapshot.py` (per the note at
`metrics_snapshot.py:93-94`), not a code change.
## Acceptance criteria
- [ ] Regolo models no longer share the generic `250 gCO2e/kWh` proxy.
- [ ] Each new factor carries a `source` label stating its provenance.
- [ ] The metrics documentation lists which providers are measured vs estimated.
## References
- `metrics_snapshot.py:93-106`, `:339-343`
- `app/main.py:2502-2515`, `:3253`
- `co-op/support#26`
Status after b11552f (the docs commit for #3): one of three criteria done; the factors are unchanged.
Documentation lists which providers are measured vs estimated — done in docs/energy-metrics.md ("What is measured, and what is not") and co-op/docs/energy.md ("Estimated: everyone else"). Both name Tinfoil, Regolo and PublicAI, and state the proxy overstates Regolo's carbon.
Regolo models off the 250 gCO2e/kWh proxy — open; IMPACT_FACTORS still has no regolo/* or tinfoil/* rows.
Each new factor carries a source label — open (moot until rows are added; the label mechanism already exists).
So the documentation half is landed; the data edit remains.
Status after `b11552f` (the docs commit for #3): one of three criteria done; the factors are unchanged.
- [x] Documentation lists which providers are measured vs estimated — done in `docs/energy-metrics.md` ("What is measured, and what is not") and `co-op/docs/energy.md` ("Estimated: everyone else"). Both name Tinfoil, Regolo and PublicAI, and state the proxy **overstates** Regolo's carbon.
- [ ] Regolo models off the `250 gCO2e/kWh` proxy — open; `IMPACT_FACTORS` still has no `regolo/*` or `tinfoil/*` rows.
- [ ] Each new factor carries a source label — open (moot until rows are added; the label mechanism already exists).
So the documentation half is landed; the data edit remains.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Context
Energy/CO2 factors are keyed by model prefix in
IMPACT_FACTORS(
metrics_snapshot.py:100-105), longest prefix first, matched withstartswith()(metrics_snapshot.py:339-342). Onlygreenpt/*models havefitted values. Every other model falls through to the single flat
IMPACT_PROXY(metrics_snapshot.py:106, applied at:343):Problem
A single flat rate cannot represent differences that the code's own data shows
are large. Within GreenPT alone, energy per token varies ~3x (0.36 vs 1.08 Wh per
1k output tokens) and grid intensity ~8x (23.6 vs 193 gCO2e/kWh). Collapsing
every non-GreenPT provider to one constant loses at least that much variation.
The clear first case is Regolo: the portal markets it as
EU-hosted, zero data retention, 100% renewables(app/main.py:2504), yet it isassigned
250 gCO2e/kWh— higher than every measured GreenPT value, includingGreenPT's own 193 and 23.6. The provider the portal labels greenest is modelled
with the dirtiest electricity in the table. That is backwards and is the kind of
inconsistency a member could spot by comparing the badge to the chart.
Secondary case: Tinfoil is TEE-architected (
app/main.py:2502), whichcarries real compute overhead that the flat proxy cannot capture in either
direction.
Affected models currently include (
app/main.py:2506-2515,:3253):tinfoil/deepseek-v4-1-flash,tinfoil/glm-5.3,tinfoil/gpt-oss-120b,regolo/apertus-70b,regolo/qwen3.5-122b,publicai/apertus-v1.5-8b.Proposed scope
IMPACT_FACTORSrows for the main non-GreenPT providers, starting withregolo/*andtinfoil/*.otherwise a reasoned, clearly-labelled estimate. Keep the existing provenance
labels so each row states how it was derived.
rather than the generic EU average; for Tinfoil, account for TEE overhead.
metrics_snapshot.py(per the note atmetrics_snapshot.py:93-94), not a code change.Acceptance criteria
250 gCO2e/kWhproxy.sourcelabel stating its provenance.References
metrics_snapshot.py:93-106,:339-343app/main.py:2502-2515,:3253co-op/support#26Status after
b11552f(the docs commit for #3): one of three criteria done; the factors are unchanged.docs/energy-metrics.md("What is measured, and what is not") andco-op/docs/energy.md("Estimated: everyone else"). Both name Tinfoil, Regolo and PublicAI, and state the proxy overstates Regolo's carbon.250 gCO2e/kWhproxy — open;IMPACT_FACTORSstill has noregolo/*ortinfoil/*rows.So the documentation half is landed; the data edit remains.