Token economics

What a coding session consumes with Majordomus and without it, from matched runs judged by the same hidden tests. Every number is generated from recorded evidence and labelled by how it was obtained.

Majordomus does not publish a token-saving figure it has not measured. Everything below is rendered from docs/generated/economics.json, which one calculator derives from the recorded benchmark runs; nothing on this page is typed by hand, and majordomus economics check refuses a savings number that is.

Two measurements answer different questions. The total-token reduction comes from matched live sessions — the same task, model and hidden acceptance tests, once without Majordomus and once with it — and is the only kind of number that could support a claim about what Majordomus saves. Context selection counts what the context compiler selects out of what it judged relevant; no model is involved, and it is not a saving.

Every token figure below is a reduction, 1 - treatment / control: positive means Majordomus used fewer tokens, and a negative reduction means Majordomus used more tokens. It is shown as it is, with the sign the calculator gave it.

The methodology, the control and treatment definitions and the reproduction commands are in the economics document.

Verdict

no verified claim
No verified total-token-savings claim is available: no valid matched pair of live runs is recorded under methodology 1.

The publication rule is not met:

0 valid pair(s); the rule asks for 30
0 categor(ies) with at least 5 valid pairs; the rule asks for 4
0 repetition(s) of csv-credit-notes (baseline) in suite pilot, the least repeated; the rule asks for 3 of every task and variant
no interval: too few valid pairs, or too few tasks among them

Metrics

Class says how a number came to be: observed (reported by the provider), counted (a pinned tokenizer over bytes on disk), derived (arithmetic over those), counterfactual (modelled, never a saving).

Metric	Value	Status	Class	n	Interval	What it is not
effective_token_reduction	—	not measured	derived from observed	0	—	not context reduction: this is everything the provider reported for the whole task, not what one mechanism selected
cost_reduction	—	not measured	derived from observed	0	—	not a bill: the harness prices usage from its own table at run time, and historical runs are never re-priced
tool_call_reduction	—	not measured	derived from observed	0	—	not a token metric
continuation_reduction	—	not measured	derived from observed	0	—	only the session that continued the work, not the whole task
first_request_overhead	—	not measured	derived from observed	0	—	not the whole overhead: what the treatment reads later because its instructions say to is in the total, not here
completion_rate.baseline	—	not measured	derived from observed	0	—	not a quality score: a run passes the gates or it does not
tokens_per_completed_task.baseline	—	not measured	observed	0	—
transcript_resume_avoided.baseline	—	not measured	counterfactual from observed	0	—	not a saving of Majordomus: both arms start the next session fresh; this is the modelled cost of resuming a transcript instead
completion_rate.majordomus	—	not measured	derived from observed	0	—	not a quality score: a run passes the gates or it does not
tokens_per_completed_task.majordomus	—	not measured	observed	0	—
transcript_resume_avoided.majordomus	—	not measured	counterfactual from observed	0	—	not a saving of Majordomus: both arms start the next session fresh; this is the modelled cost of resuming a transcript instead
context_reduction_ratio	+97.9%	measured	derived from counted	315	95%: +97.9% to +98.0%	not total token savings: it says what the compiler selected from what it found relevant, not what a session consumed, and a session remains free to read anything
context_cost_model_error	-1.6%	measured	derived from counted	315	—	not a saving: it says how far the budget's unit is from counted tokens

Evidence

Suite	Kind	Evidence	Runs	Valid pairs	Failed (control / treatment / both)	Revisions	Models
context	context	current	9	—	—	1cc12af61de4 487cedf6f5a3 78f09f8af8e3 79bb4e3c18a9 8183164ade2d 8d4486c1a9df 94c515480804 a588b8eb5fac d446499470fe
pilot	live	no evidence	0	0 / 0	0 / 0 / 0

Every pair

Valid or not. Tokens are everything the provider reported for every session of the run; the reduction is 1 - treatment / control, and a negative reduction means Majordomus used more tokens. An excluded pair shows its reduction so that the outlier rule's effect is visible; only valid pairs enter the metrics.

Task	Category	Rep	Status	Control tokens	Treatment tokens	Token reduction
csv-credit-notes	regression	1	missing	—	—	—
csv-credit-notes	regression	2	missing	—	—	—
currency-czk	small-feature	1	missing	—	—	—
currency-czk	small-feature	2	missing	—	—	—
invoice-discounts	cross-module	1	missing	—	—	—
invoice-discounts	cross-module	2	missing	—	—	—
recurring-invoices	multi-session	1	missing	—	—	—
recurring-invoices	multi-session	2	missing	—	—	—
storage-v2	schema-change	1	missing	—	—	—
storage-v2	schema-change	2	missing	—	—	—
test-repair	test-repair	1	missing	—	—	—
test-repair	test-repair	2	missing	—	—	—
vat-rounding	bug-fix	1	missing	—	—	—
vat-rounding	bug-fix	2	missing	—	—	—

Context selection

At revision 79bb4e3c18a9, 315 issues of the plan were compiled by the context compiler; counted with o200k_base (tiktoken-rs 0.12.0), it selected 7857785 of 396047691 candidate tokens across all of them. This is selection, not a session's saving, and it counts the corpus at that revision, not at the current one.

The same ratio at each of the 9 recorded runs, oldest first. A point is a recorded run and nothing between two points was measured.

Recorded at	Reduction	n	Revision
2026-09-26T12:29:55Z	+96.2%	216	1cc12af61de4
2026-10-02T02:22:32Z	+97.1%	238	a588b8eb5fac
2026-10-02T09:55:23Z	+97.2%	239	8183164ade2d
2026-10-02T11:26:34Z	+97.2%	243	94c515480804
2026-10-02T12:48:07Z	+97.3%	239	78f09f8af8e3
2026-10-03T05:11:05Z	+97.3%	243	8d4486c1a9df
2026-10-04T02:41:01Z	+97.4%	244	487cedf6f5a3
2026-10-06T19:36:53Z	+97.5%	257	d446499470fe
2026-10-11T18:48:45Z	+97.9%	315	79bb4e3c18a9

Hypotheses

Stated before any evidence existed, so that the evidence can refute them. They are not results.

long-running-work (untested): Long-running agent development work consumes 30 to 70 percent fewer tokens with Majordomus, centred near half.
short-tasks (untested): Short isolated tasks gain less and may cost more, because the bootstrap and briefing are paid on every session.
continuity (untested): Work that spans sessions gains the most, because a handover replaces rediscovery.

Generated from docs/generated/economics.json; the human-readable report is docs/generated/economics.md, and the raw runs are under .ai/repo/benchmarks/economics/runs.