Skip to content

docs(readme): document the green software engineering programme - #508

Merged
cdeust merged 1 commit into
mainfrom
docs/green-software-engineering
Sep 8, 2026
Merged

cdeust merged 1 commit into
mainfrom
docs/green-software-engineering

Conversation

@cdeust

@cdeust cdeust commented Sep 8, 2026

Copy link
Copy Markdown
Owner

Why

Thirteen merged PRs (#475#487, #506) and a full SCI-grounded energy harness under benchmarks/energy/ had no mention anywhere in the README. A reader had no way to know the programme existed — which also meant no way to hold it to its own evidence rule.

What the section says

A new ## Green software engineering section, placed before ## Verification so the claims-discipline sections sit together.

The harness, and what it does not establish. benchmarks/energy/ implements the GSF SCI specification (O = E × I, M = TE × TS × RS), functional unit = 1000 model input tokens counted from the tokenizer's own attention_mask. The section then spends equal space on the limits, taken from the harness's own README:

  • its opening paragraph — the fixtures do not measure device energy and do not establish an energy improvement;
  • no default carbon factors--carbon-intensity and --embodied are mandatory operator inputs, validated before any model import, and the harness does not vouch for their provenance;
  • a stated boundary — the sensor's combined CPU+GPU+ANE estimate, excluding memory, storage, screen, power-supply losses, warm-up and token counting;
  • artifacts preserved (results.json, MANIFEST.json, the analyzed powermetrics.txt).

No energy figure is published, and the section says why: no energy results are committed here, because a number measured on one operator's machine, region and duty cycle is not a property of the software, and publishing it as one would be exactly the drift this programme exists to prevent.

What actually shipped — a table of the merged CI/build and runtime PRs, each linked.

Demand reduction is presented as the primary lever, and the paragraph is explicitly labelled a design rationale, not a measurement, with a closing sentence stating that Cortex publishes no token-savings or CO₂ figure for end-to-end agent sessions because it has not measured one.

The one quantitative claim

Hook boot ~0.05 s vs ~0.6 s for the full registry import (measured 2026-07-28) — an already-committed constant, cited to its call sites in mcp_server/hooks/auto_recall.py:247,300, per the no-invented-constants rule.

Verification

Each cited fact was checked against its source file, not against prose:

Claim Verified against
SCI formulas, functional unit, attention_mask benchmarks/energy/README.md
carbon factors mandatory, no defaults same, "mandatory, finite and nonnegative" / "intentionally have no example"
CPU+GPU+ANE boundary and exclusions same, "Memory, storage, screen, power supply losses"
artifacts preserved same
no energy results committed benchmarks/results/energy/ — confirmed absent
CORTEX_RERANKER_OFFLINE mcp_server/core/reranker_model.py:66
response_budget keeps ids mcp_server/core/response_budget.py

python3 scripts/check_doc_claims.py exits 0. Documentation only — no code, no behaviour change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01StMBvNd7eVJGtpnC2zNsx1


Generated by Claude Code

Thirteen merged PRs (#475-#487, #506) and a full SCI-grounded energy
harness under benchmarks/energy/ had no mention anywhere in the README.
A reader had no way to know the programme existed, which also meant no
way to hold it to its own evidence rule.

The section states what the harness measures and, at equal length, what
it does not: its own README opens by saying the fixtures do not measure
device energy and do not establish an energy improvement, the carbon
factors are mandatory operator inputs with no defaults, and the boundary
is the sensor's CPU+GPU+ANE estimate excluding memory, storage, screen
and power-supply losses.

No energy figure is published, and the section says why: no energy
results are committed to this repository, because a number measured on
one operator's machine, region and duty cycle is not a property of the
software. The demand-reduction paragraph is explicitly labelled a design
rationale rather than a measurement, since no end-to-end token-savings
or CO2 figure has been measured.

The one quantitative claim is the already-committed hook boot constant
(~0.05s vs ~0.6s registry import, measured 2026-07-28), cited to its
call sites in mcp_server/hooks/auto_recall.py.

check_doc_claims.py exits 0.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01StMBvNd7eVJGtpnC2zNsx1
@cdeust
cdeust marked this pull request as ready for review September 8, 2026 13:08
@cdeust
cdeust merged commit 7451985 into main Sep 8, 2026
25 checks passed
@cdeust
cdeust deleted the docs/green-software-engineering branch September 8, 2026 16:01
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants