Compaction does not lose the unimportant things first.

Paste a session transcript, set the context budget your agent really gets after compaction, and see which facts survive. The cliff is the cheapest budget that still keeps everything the session depends on. Below it a rule the agent was given is gone, and nothing in the transcript says so.

A bicycle retailer's first line agent, 41 turns in. The refund policy is quoted with its exceptions, and a draft reply has been through four revisions.

cliff 42,100

budget 42,100

Load bearing facts keptAll facts kept, by count

Usable context budget in tokens. The shaded area is the share of load bearing weight that survives. Both lines fall as the budget shrinks. The load bearing line is the one that matters, and the cliff is where it leaves the top of the chart.

42,100 tokens for facts, after 24,000 of window overhead40% of the usable window

Cliff42,100The cheapest budget that keeps every load bearing fact.CrowdedOnly expendable content is being dropped

At 42,100 tokens, every load bearing fact survives and 5 lower value facts are crowded out. That is exactly the cliff, so a token less and something load bearing starts to go.

Load bearing kept100%

All content kept79%

Facts at risk5 / 14

Load bearing lost0

Fact ledger

Support triage, sorted by how load bearing each fact is. Needs is the cheapest budget at which that fact survives whole.

Retention is a modelled chance that the fact survives compaction at 42,100 tokens.
Fact	Tokens	Turns ago	Needs	Retention	Status
charterIdentityload bearing You are Velo Support, the first line for a bicycle retailer. Plain, warm, no exclamation marks. Three worked sign-off examples follow.	3,200	41	11,500	100%	Intactfits inside the budget
confirm-orderInstructionload bearing Confirm the order number before discussing anything about an order.	2,100	41	2,100	100%	Intactfits inside the budget
escalateInstructionload bearing Escalate immediately on any mention of a lawyer, an injury, a chargeback or a regulator. Hand off to a named human rather than a queue.	3,600	29	8,300	100%	Intactfits inside the budget
kb-onlyConstraintload bearing Answer only from the knowledge base articles in scope. If the answer is not there, say you will check and escalate, with two examples of how that reads.	6,800	33	25,500	100%	Intactfits inside the budget
no-codesConstraintload bearing Never quote internal order status codes to a customer. Use the plain language mapping instead.	2,600	40	4,700	100%	Intactfits inside the budget
refund-policyConstraintload bearing Never promise a refund. Quote the refund policy, including its three exceptions, and route any refund request to a human. The policy text with exceptions is 9,400 tokens.	9,400	38	42,100	100%	Intactfits inside the budget
no-discountDecisionload bearing Decided not to offer a discount, because the courier record shows a signed delivery and the policy has no exception for disputed signatures.	4,400	12	15,900	100%	Intactfits inside the budget
tracking-outageErrorload bearing The tracking lookup at turn 24 returned nothing for forty minutes. The customer was told we would check back, and no check back has been sent.	7,200	17	32,700	100%	Intactfits inside the budget
contradictionRelation The signed delivery record contradicts the customer's claim, so any refund language is a promise the policy does not allow.	2,800	4	18,700	100%	Intactfits inside the budget
draft-replyArtifact Draft reply, revision four: acknowledges the missing delivery, repeats the courier signature, offers a replacement tyre at cost, and closes without promising a refund.	26,000	2	101,100	0%	Crowdedcrowded out, low cost to lose
internal-noteArtifact Internal note appended to the ticket summarising all three contacts, the outage and the current draft.	14,000	1	66,100	0%	Crowdedcrowded out, low cost to lose
contact-historyState Third contact about VB-88213. The two earlier conversations both ended with a refund request and a refusal, and the customer asked to speak to a manager on the second.	6,200	18	52,100	0%	Crowdedcrowded out, low cost to lose
customer-claimState The customer states nothing arrived and that the signature on the courier record is not hers.	3,800	6	45,900	0%	Crowdedcrowded out, low cost to lose
order-recordState Order VB-88213, Mara Ellison, two 700x28 tyres, dispatched Tuesday by courier, tracking 9V1 4471 8820, delivered and signed for Thursday at 14:20 by M ELLISON.	9,000	22	75,100	0%	Crowdedcrowded out, low cost to lose

Loss by kind

What is dropped at 42,100 tokens, grouped by what the fact is.

Constraint3 of 3 kept

Identity1 of 1 kept

Instruction2 of 2 kept

Decision1 of 1 kept

Error1 of 1 kept

Relation1 of 1 kept

Artifact0 of 2 kept

State0 of 3 kept

How this reading is produced

This is a model of pressure, not a measurement of any particular framework. It is deterministic: the same transcript and the same budget always give the same reading.

Each fact carries a keep value: its weight divided by the square root of the tokens it occupies, lifted a little if it was restated recently. Short and important is kept before long and incidental.
A budget keeps the highest keep-value facts until it runs out. A fact that does not fit keeps a shrinking chance rather than vanishing, because a summariser may preserve part of it.
A fact at or above 0.8 weight is load bearing: losing it changes behaviour rather than degrading quality. Constraints, identity and instructions start there. State and drafts do not.
The cliff is the cheapest budget that keeps every load bearing fact intact. That single number is the one worth carrying into a build.

Measure your own transcript

TranscriptJSON with a window, an overhead and a facts array. Nothing leaves the browser.

Tokens are declared per fact. Describe the real size of each block, because a one line policy with three quoted exceptions is not a one line cost.