01 · Freeze
You supply one sanitized workload. Configuration and inputs are digest-recorded before anything runs, so the scoring target cannot move once results are visible.
Context-control evidence assessmentFixed fee · frozen scope · reproducible
Measure the context pipeline before you commit to it.
One workload, frozen and digest-recorded. Three paths compared on tokens, cost, evidence retention, citations, latency and failure behaviour. A written recommendation to proceed, revise, or stop.
Context compression is easy to adopt on a demo and expensive to unwind in production. This is a bounded engagement that answers whether it is worth doing on your workload, at a fixed fee, before anything is installed.
02 · Compare
Three paths are run against that frozen workload: your current baseline, your gateway-native compression, and Maha.
03 · Decide
You receive sanitized per-workload findings and a written recommendation to proceed, revise, or stop. Stop is a real outcome.
Read the method before you pay for it
The whole method is public.
A five-figure fee for measurement work is only honest if the prospect can read the method first. These four documents are published in full, they include their own limitations, and none of them requires a conversation to obtain.
Sample assessment
A complete worked deliverable in the format you would receive, so the output is known before the engagement starts.
Open ↗Security boundary
What is handled, what is retained, and what never leaves your side. Read this before preparing a workload.
Open ↗Live gateway evaluation
A full run on a synthetic 20-workload corpus: per-workload rows, aggregates, cost assumptions, and its own stated limitations.
Open ↗MCRB-1 dense baseline
The retrieval baseline measured against the frozen cohort — including where it scores higher than Maha does.
Open ↗Every figure in those artifacts is bounded by the run that produced it. The gateway evaluation is a single execution against a synthetic corpus and does not establish performance on any customer workload — the artifact says so itself, and that sentence is the reason this assessment exists.
Fees
A fixed fee for a bounded decision.
The fee is agreed before the workload is frozen and does not move with the result. An assessment that recommends stopping costs the same as one that recommends proceeding.
Standard Context-Control Evidence Assessment
$12,500
One customer-supplied workload, three paths, a written recommendation.
Extended Assessment
$25,000
Multiple workloads or a second gateway configuration, with per-workload findings for each.
Founding design partner · $2,500
Available to the first two signed customers, agreeing in advance to act as a named reference and to permit an anonymized integration note.
This is not a general or negotiable discount. It is a fixed exchange for reference participation, and it closes after two customers.
Scope and limits
What is measured, and what is refused.
What it produces
- • A customer-supplied, sanitized document or RAG workload. No production credentials and no personal data.
- • Configuration and workload frozen and digest-recorded before anything runs, so the scoring target cannot move after results are seen.
- • Three paths compared: your baseline, your gateway-native compression, and Maha.
- • Token and cost measurement, evidence retention, citations and provenance, latency, and failure-path behaviour.
- • Sanitized per-workload findings and a written proceed, revise, or stop recommendation.
Explicit limits
- • No production deployment. The assessment measures; it does not install.
- • No performance or savings guarantee. Nothing is promised before measurement.
- • No certification or compliance opinion of any kind.
- • No open-ended discovery, data migration, or custom implementation work.
No production credentials and no personal data are accepted. The workload you supply must be sanitized before it is sent, and the security boundary document states exactly what is handled and retained.
What to judge Maha on
Determinism and evidence, not a headline number.
- • Deterministic selection: the same inputs produce the same pack, with no model in the path.
- • Hard budgets: the declared token budget is enforced rather than advised.
- • Source-linked provenance: every retained passage carries its source and passage identifier.
- • Stable hashes: input and output commitments a reviewer can recompute.
- • Reproducible evidence: per-workload rows and a one-command check, not a headline.
No retention-superiority claim is made here. The public evidence package includes a dense baseline that scores higher on evidence retention than Maha's production scorer on the frozen MCRB-1 cohort. It is linked above rather than omitted, because a positioning line contradicted by your own published artifact is worse than no positioning line.
How to start
One email, then a frozen scope.
Describe the workload and the decision you are trying to make. If the assessment is not the right instrument, that is said before a fee is quoted — the scoping conversation is not itself billable.