[davidberenstein1957]

Description

Adds an SCI (Software Carbon Intensity, ISO/IEC 21031:2024) output method, following the same pattern as the existing BoAmps exporter. codecarbon/output_methods/sci.py holds FunctionalUnit/EmbodiedDeclaration dataclasses for the user-supplied R and M terms, a pure map_emissions_to_sci(data, ...) mapping function (I is derived as emissions * 1000 / energy_consumed rather than recomputed, so the report agrees with the CSV by construction), and SCIOutput(BaseOutput), which writes sci_report_<run_id>.json and sci_report_tasks_<run_id>.json. OutputMethod.SCI is registered in _init_output_methods, reading an optional sci_context_file from the usual config hierarchy. When R is undeclared the report still writes with sci: null and a status field explaining why, instead of dividing by zero. When M is undeclared, M_gCO2e: 0 with M_source: "not declared". A follow-up commit apportions the declared M across per-task reports by each task's share of summed duration, instead of charging every task the device's full embodied carbon.

Related Issue

Fixes #1353

Motivation and Context

CodeCarbon had no SCI output method. Users who need an ISO/IEC 21031-aligned report had no supported path, and any naive implementation risks inventing numbers (e.g. assuming R = 1 or guessing embodied carbon) that a compliance report cannot support.

How Has This Been Tested?

uv run pytest tests/test_sci_output.py -q — 32 tests covering the formula, provenance assembly, the undeclared-R and undeclared-M branches, zero-energy, the handler and its task path, and context-file loading (valid, missing, malformed). Includes test_task_out_apportions_embodied_by_duration (unequal durations: asserts a 10/30 split of a declared 40 gCO2e and that shares sum back to 40). black --check and ruff check were run scoped to the touched files.

Screenshots (if appropriate):

N/A

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

AI Usage Disclosure

• 🟥 AI-vibecoded: You cannot explain the logic. Car analogy : the car drive by itself, you are outside it and just tell it where to go.
• 🟠 AI-generated: Car analogy : the car drive by itself, you are inside and give instructions.
• ⭐ AI-assisted. Car analogy : you drive the car, AI help you find your way.
• ♻️ No AI used. Car analogy : you drive the car.

Checklist:

• My code follows the code style of this project.
• My change requires a change to the documentation.
• I have updated the documentation accordingly.
• I have read the docs/how-to/contributing.md document.
• I have added tests to cover my changes.
• All new and existing tests passed.

User-facing surface

from codecarbon import EmissionsTracker
from codecarbon.output_methods.sci import FunctionalUnit, SCIOutput

sci = SCIOutput(functional_unit=FunctionalUnit(name="inference request", count=10_000))
tracker = EmissionsTracker(output_handlers=[sci])

sci.set_functional_unit_count(n) for the common case where the count is only known at the end; SCIOutput.from_file("sci_context.json") for the declarative path; output_methods = csv,sci plus sci_context_file in .codecarbon.config. Docs: a new SCI section in docs/reference/output.md, a row in docs/how-to/examples.md, and a runnable examples/sci_output.py. No new dependencies.

Deliberately left out

Any automatic embodied-carbon estimation (no EmbodiedProfile tracker parameter, no bundled hardware table, no cloud per-instance figures) — CodeCarbon has no defensible manufacturing data, and an uncited M in a compliance report is worse than a blank one. Also out: any change to what emissions means, multiple simultaneous functional units, and schema validation on write (there is no stable machine-readable SCI schema yet).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.80%. Comparing base (53e67c3) to head (ee8d215).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1361      +/-   ##
==========================================
+ Coverage   93.68%   93.80%   +0.11%     
==========================================
  Files          50       51       +1     
  Lines        5604     5710     +106     
==========================================
+ Hits         5250     5356     +106     
  Misses        354      354              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[benoit-cty]

🤖 This review comment was written and posted by Claude Opus 5.5 (AI assistant), at the request of @benoit-cty. Findings were checked by reading the code and running tests locally (merged with current master where relevant), but please double-check before acting on them.

Verdict: ✅ Approve after 1 doc fix and 1 robustness fix

SCIOutput(BaseOutput) is clean, self-contained and follows the BoAmps pattern. The formula (E*I + M)/R is implemented correctly. It refuses to invent R (writes sci: null with a status) and flags an undeclared M. 29 tests pass, and it merges cleanly with master.

Must fix:

1. The sample report in docs/reference/output.md shows the wrong SCI value. It shows "sci": 0.0000127 with E=0.1007, I=417.08, M=42.5, R=10000. The formula gives (0.1007×417.08 + 42.5)/10000 = 0.00845. This is the canonical example people will copy, so it needs to be right.
2. A context file whose top level isn't a JSON object stops all tracker output.
   • A sci_context_file holding valid JSON that isn't an object (e.g. [1]) raises AttributeError at sci.py:183.
   • The except (OSError, TypeError, ValueError) around emissions_tracker.py:~678 doesn't catch it. __init__ aborts under @suppress and leaves _output_handlers == [], so no CSV or anything else is written.
   • Fix: validate isinstance(context, dict) in from_file and raise ValueError, or add AttributeError to the except. Please add a test.

Should fix:
3. M is split by the tasks' summed duration, not the run's (sci.py:259-262).

• share = task.duration / sum(task durations), but the docs and docstring say "share of the run's duration".
• Example: a 100 s run with one 10 s task assigns that task 100% of M.
• Either divide by the run duration, or change the docs and drop the "per-task figures sum back to the run-level report" claim, which only holds when tasks cover the whole run.

Nits:

• Task reports always show "pue": 1, because TaskEmissionsData has no pue field and sci.py:103 falls back to 1. That is misleading provenance when PUE ≠ 1. I itself is still correct because it's derived.
• set_functional_unit_count() mutates the caller's FunctionalUnit instance. Copy it, or document that.
• A relative sci_context_file resolves against the CWD, not the config file's location. Worth a line in the docs.
• flush() writes an interim report that applies the full declared M to only part of E. It is overwritten at stop(), but anyone reading the flushed file sees an inflated M share. Consider pro-rating M by elapsed time in interim reports.

[davidberenstein1957]

Made the changes in 89d2897: non-object context file now raises ValueError (tracker degrades, test added), sample SCI fixed to 0.00845, M-split docs now say tasks' summed duration, task pue is null, FunctionalUnit is copied, relative path documented.
Skipped pro-rating M in flush() reports: they're overwritten at stop(), and pro-rating needs a design call on elapsed vs. declared run length.

[benoit-cty]

Rebased on master and reviewed. New head is 965eb91.

• Rebase: git reported no conflict, but it placed the OutputMethod.SCI registration after return BoAmpsOutput(...) in master's new _build_boamps_output() helper. That code could never run, so output_methods=sci did nothing and 4 TestTrackerRegistration tests failed. I moved the block back into _init_output_methods, inside the feature commit 74dfe7d, so every commit in the history works.
• Earlier review: every must-fix and should-fix item is handled in 6a229c3. That covers the non-object context file, the sample SCI value (0.00845), the docs for how M is split across tasks, pue: null in task reports, the FunctionalUnit copy, and the docs line on relative paths.
• Docs: 965eb91 notes that flush() reports are interim. They apply the full declared M to the energy measured so far and are overwritten at stop().
• Coverage: codecarbon/output_methods/sci.py is at 100%, and every changed line in emissions_tracker.py is covered.
• Tests: pre-commit passes. tests/test_sci_output.py passes 32/32. The full package suite has 860 passed and 4 failed. The 4 failures are RAPL permission tests, and they fail the same way on master because the sandbox runs as root.

Maintainer decisions still open:

1. Pro-rating M in flush() reports. The handler gets the same out() call from flush() and stop(), so it can't tell an interim report from the final one. Pro-rating would need either a new flag passed from the tracker to the handler, or an M declared over time (the spec's TE × TiR/EL form: total embodied carbon times the run's share of the hardware's lifespan). For now only the docs mention it.
2. Where the config lives. Master now has a generic metadata bag with per-output keys, and BoAmps reads from it. SCI still uses its own sci_context_file key. Should SCI read a metadata["sci"] section instead, to match?

Not run: uv run task docs.

---

Generated by Claude Code

[benoit-cty]

Great !