Study 8 of 822 October 2026Croxio Ltd792 chains, 396 on Claude Haiku 5.5 and 396 on Claude Haiku 4.5, pre-registered, after a pilot and a hardening ladder

When a handover passes through agents that do not act on it, how much arrives, and does a line checked at every hop stop the loss?

The question

Every study so far was one edge: a sender, a receiver, an act. Agent systems are graphs. A planner hands to a coordinator, the coordinator to a desk, the desk to the agent that holds the tool, and each restates the instruction for the next. A relay is a receiver that becomes a sender, so every hop is another draft. Does anything carry the instruction intact across them without a check at each one, and what does the check cost?

The material

Study 6's fictional month-end, unchanged: three acts the payer cannot take back, release a payment, change a supplier's bank details, send a remittance advice, with the account or the address resolved by the payer from its master. The first message is written by code from the approval record.

The chain

A scheduler that is code, then two relays that are models, a coordinator that adds the deadline and priority from its run schedule and a desk that confirms the approval against its log, then the payer. Each relay sits in an inbox of routine traffic with two other live handovers for other suppliers. A sweep runs one, two, three and five relays.

Compared

Free prose. Prose under the seven conventions. Lines validated with the codec at each hop. Lines read back at every hop and checked before they are forwarded against the handover the relay received, by code, with one re-issue and then a hold.

Receivers

Claude Haiku 5.5, the receiver studies 6 and 7 showed null at its own locus, and Claude Haiku 4.5, the weakest to hand, standing for a fleet whose relays are smaller than the agent that acts. Every agent in a chain runs on the same model, as headless sessions. No judge.

Graded by

Code: the act as the tool executed it against the record; the fields each message carried, by parse for a line and by normalised match for prose; drift, a value changed from a clean source; every refusal and whether it was earned. Every proportion carries its exact 95% interval. The protocol, its amendments, the pilot, the ladder and the frozen manifest are in the repository.

Why we did it

Studies 6 and 7 placed the faults on a single edge: the sender's draft and the tool, with the receiver's own reading null on strong models. If the drafting locus is where faults enter, a chain of N hops has N drafting loci. The protocol's answer was already in the spec: a handover before an irreversible act is built from the record by code or checked against it before it goes, and every receiver reads back what it received. For a relay the record is the handover it received. Nothing had measured what that costs across a chain, or what it prevents against agents passing the same instruction on in prose.

The telephone game is the folk model of a relay, and it predicts changed values. Study 6 found strong models do not misread their own context; whether they misdraft what they forward, when they have a reason to touch the message, was the open question. So every relay here has something to add, and a chain is a sequence of drafts rather than a sequence of copies.

What we tested

Hop 0 is the scheduler, code, writing the first message from the record: supplier, amount, currency, gross or net, payment date, invoice reference and approval reference for a payment; supplier, new account, effective date, approval reference and that the call-back was done for a bank change; supplier, address, amount, currency, date paid and reference for a remittance advice. The payer resolves the payee account from its master. The coordinator holds the run schedule and adds the deadline and the priority. The desk holds the approvals log, confirms the reference covers this item and marks the handover approved; where the log does not cover it, the correct desk asks instead of forwarding. In the sweep the relays beyond the desk, named from the pay profile, pass the handover on with nothing to add. Every relay's inbox carries nine routine messages and, after the hardening ladder, two other live handovers for other suppliers, so that a relay must pick its handover and cannot find it by its being alone.

The four forms. Every arm has the same agents, inboxes, master, execution rules and act tool; the briefs differ only in the message form and in what a relay does before it forwards, and are within 10% of one length per role across the three instructed arms.

The lines arm separates the form from the checks: if it matches the checked arm, the line alone carries the instruction; if the checked arm beats it, the checks do the work. Free prose against disciplined prose separates the discipline from the form. In the checked arm the forwarded line must carry every slot the received line carried, with the same decoded values, plus the relay's addition, which the same code checks against the relay's own context: the run of the act's kind on the act's date in the schedule, the log entry covering the supplier and the item. check-ho refuses a dropped slot, a changed value, or a slot added that the relay's role does not name. Every handover in the chain asks rb=all, since a relay resolves nothing; the codec's demand for rb=exec on an irreversible act was waived for the relay's check and recorded on every one, and nothing else was.

Three blocks. Block 1, clean: three acts, five clean contexts, three repeats, 45 chains per arm per model. Block 2, a confuser at a relay: a second live handover for the same supplier with other figures in the coordinator's inbox, a run on the adjacent day in its schedule, a near approval reference beside the right one in the desk's log, or a log that covers the supplier and not the item, where the correct chain ends with the desk asking; four confusers per act, three repeats. Block 3, the sweep: one clean context per act, one, two, three and five relays, prose against checked lines, three repeats.

How we ran it

1. Wrote the protocol before building: the chain, the arms, the blocks, the measures, eight hypotheses and their tests, the sizing, the pilot gates and a hardening ladder, what the study will not show. It is research/chain/README.md, with every amendment appended under its own heading and date.
2. Built the harness from study 6's, with the codec and the pay profile frozen into the study folder, and smoke-tested one chain per arm without reading any result.
3. Piloted 16 chains, free prose and checked lines, on each model. The checked arm met its gate: 8 of 8 chains reached a correct act with 0 refusals and no harness fault. The prose arm did not meet its gate of loss in one chain in ten: 0 of 8 lost anything, because every relay copied the paragraph it received and appended a sentence. A worked example in the relay inboxes that the protocol had never named was removed and the prose pilot re-run at level 0: 0 of 8 lost on Haiku 5.5 and 0 of 8 on Haiku 4.5. Ladder step 1, two other live handovers in every relay's inbox: 0 of 8 and 1 of 8, the one loss a desk's redraft dropping the call-back statement from a bank change. The gate was met on one model and the ladder stopped; the main run is at that level in every arm. The pilot and the ladder cost $4.88.
4. Made the call-back statement a graded field for a bank change, carried in a slot (x_cb=Y) rather than the tail, because the one loss on the ladder was a fact the coverage measure had not counted. Froze the tools, the data and the protocol at one commit, recorded in results/manifest.json, and ran the design.
5. Stopped the run once. In the lines arm without checks every remittance chain had ended at the first hop with the coordinator asking for the supplier's address, which the first message did not carry and a relay holds no master to resolve: a defect in the case, not a relay losing anything. The first message now carries the address, every remittance chain of every arm was reset and run again from a new freeze, and the aborted pass is kept. Under the registered exclusion rule, 3 chains that ended with a relay writing its forwarded message to a misspelt path were named with the reason and run again as repeat 4. 792 chains in all, $10.66 of model time on Haiku 5.5 and $82.90 on Haiku 4.5, 0 harness faults.
6. Graded everything in code, then had a blind second reading of 40 chains, drawn by seed, grade the act, every field at the payer and drift from the chain files alone: agreement 40 of 40 on the act, 40 of 40 on the fields and 40 of 40 on drift. Every figure on this page is read from results/summary.json by the script that writes the page, and a test in the app asserts they match.

What happened

Block 1, clean: the headline

Table 1. Block 1 on Claude Haiku 5.5, two relays, 45 chains per form. "Asked" is a chain that ended with a relay or the payer asking; "held" is a chain the checks stopped after a second refusal. False alarms are refusals of a forwarded line that carried every received slot with the record's values, over correct forwards, and refusals of a readback whose values were all right, over readbacks checked. Size is the message delivered to each hop; traffic counts every message between agents, readbacks and re-issues included; cost is what the headless sessions cost.

In words: on Haiku 5.5 every line chain acted, checked or not, and prose lost 18 chains in 90 to questions, none to a changed value. The facts were there: 43 of 45 free-prose chains and 38 of 45 disciplined ones delivered every record field to the payer, and H2, which asked whether the line delivers more fields, does not separate the arms (p 0.247). H1 holds: more chains end in a correct act under checked lines than under free prose, p 0.003. Reading the questions, which is a reading and not a grade, they fall into three kinds: the payer refusing a handover the relay had forwarded inside a wrapper still headed to the relay ("To: desk, Hi desk"), under its rule that a handover addressed to another agent is not its own; the payer or a relay wanting the approval record or the call-back shown rather than reported; and a few others. A line cannot fail the first way, because the validator makes the relay rewrite from= and to=, and here it did not fail the second, because apref=, x_cb=Y and the desk's appr=OK are the provenance in a checkable shape. The disciplined prose asked more often than the free, 10 of 45 against 8 of 45, and grew more per hop, 73 to 174.5 tokens against 73 to 156; the lines stayed near flat, 77 to 111 and 77 to 100. The checks caught nothing, because there was nothing to catch, and refused nothing they should not have: 0 of 90 false alarms in correct forwards, upper bound 4%, and 0 of 135 in readbacks. H4 holds.

Table 2. Block 1 on Claude Haiku 4.5, the same design and columns.

In words: on Haiku 4.5 the order reverses. Free prose and the unchecked line each lost one chain in 45; the checked line lost 8. H1 fails with the direction reversed (p 0.999) and H8, registered as the checked line beating the unchecked one on this model, separates them the other way (p 0.03 on correct acts, 0.012 on every field at the payer). Of the 8 losses, 7 are one event. The relay composed its readback by typing the handover into make-rb rather than passing the message as delivered, left the tail off, and the digest the readback carried was the digest of a different string; the previous hop's code refused it at the integrity tier, re-issued the handover, the relay read that back the same way, and the chain was held with every value right in 6 of 45 chains, while in 1 the payer, refused once, asked the scheduler instead of reading the re-issued line back. The other loss was a desk that could not find in its log the approval reference the handover carried, and said so. The forwarded lines were never refused wrongly, 0 of 77; the readbacks were, 13 of 128 (10%, 6 to 17), so H4 holds as registered over forwards and the readback figure beside it fails the same bound. No relay on this model changed a value in the clean block either: 0 of 45 drift in free prose, 0 of 45 in checked lines.

Block 2, a confuser at a relay

Table 3. Block 2 on Claude Haiku 5.5, 36 chains per form: four confusers per act, three repeats. "Ended as expected" is the correct act in three confuser cells and the desk asking in the fourth, where the log covers the supplier and not the item. "Graded wrong" counts a chain that forwarded past the desk and acted, or that asked somewhere other than the desk, in that fourth cell.

Table 4. Block 2 on Claude Haiku 4.5, the same columns.

In words: no relay in any arm forwarded a decoy's figures in a two-relay chain. The confusers produced questions, not wrong values, and on Haiku 5.5 the checked line answered every one as the design expected (36 of 36) where free prose managed 31 of 36; H6 holds (p 0.027). The run made 0 wrong acts in 792 chains. The 2 chains graded wrong on Haiku 5.5 are both in the cell whose correct outcome is the desk asking: a prose coordinator asked the scheduler at the first hop instead, about an approval reference the first message did not carry, so the chain never reached the desk whose question it was (0 of them acted). On Haiku 4.5 both prose arms ended every chain as expected, the unchecked line 34 of 36, and the checked line 27 of 36, with 7 of 36 held for the retyped readback and 2 more asking after a first refusal; H6 fails there (p 1.0).

Block 3, the sweep

Table 5. Block 3 on Claude Haiku 5.5: one clean context per act, three repeats, nine chains per cell. Size is the median message at hop 1 and at the payer; the relays beyond the desk add nothing.

Table 6. Block 3 on Claude Haiku 4.5, the same columns.

In words: on Haiku 5.5 the checked line acted in every chain at every length and free prose fell at five relays to 5 of 9, H7's shape; the prose message delivered to the payer grew from 73 tokens to 197 over five relays that had nothing to add, and the line from 77 to 104. The 4 of 9 prose chains that asked at five relays did so, by the same reading, mostly on addressing or provenance. The line's readbacks make it the heavier wire at that length: 1,370 tokens of traffic against 971, 12.0 model turns against 5.4, $0.094 against $0.028 a chain. On Haiku 4.5 neither form falls with length: prose 9 of 9, 9 of 9, 8 of 9 and 9 of 9; the checked line 8 of 9, 7 of 9, 6 of 9 and 8 of 9, every shortfall a hold of the same kind (1 of 9, 2 of 9, 3 of 9, 1 of 9), and the holds do not grow with the number of relays.

The two drift events, the two dropped fields, and the refused readbacks

Drift is a record field present in a message with a value that differs from the record, where the message before it carried the value correctly. It happened twice in 792 chains. In B2-CHGBANK-desk-cover-A-haiku55-k2-r1 a disciplined-prose coordinator on Haiku 5.5 forwarded the bank change with the sentence "Call-back: done, to the number on file, before approval", and the grader's rule for prose, which counts a call-back mention as conditional when it carries a word such as "before", recorded the call-back as drifted; the call-back was in fact stated as done, the code's grade stands under the protocol, and the chain ended, correctly, in the desk's question, since its log did not cover the item. In B3-CHGBANK-clean-1-F-haiku45-k3-r1 the telephone game happened once: a free-prose relay on Haiku 4.5 at the third hop of a three-relay chain forwarded a different handover altogether, the decoy Coleridge payment in place of the Evergreen bank change, and the payer asked rather than paid. On a line, check-ho at that hop refuses it on item alone.

A field was dropped twice, both times by free-prose relays on Haiku 4.5, both the call-back statement of a bank change (B2-CHGBANK-desk-near-F-haiku45-k2-r2 and B3-CHGBANK-clean-1-F-haiku45-k2-r3): the fact the pilot had seen a relay lose is the fact the prose arms lost, and in both chains the payer changed the master without it, correctly on every other field. No line arrived at the payer short of a slot, in 792 chains. Elsewhere, where a cell shows fewer chains with every field at the payer than correct acts, the difference is chains that stopped at a relay with a question.

The readback false alarms are all on Haiku 4.5 and all of one kind. Over the whole run its checked-line relays and payers had 44 readbacks refused with every value right, 44 of them at the integrity tier: the h= digest did not match the line as delivered, because the model had retyped the line into make-rb without its tail before computing it. Haiku 5.5 had 0. Study 7 met the same fault at the tool and answered it by having the tool take the handover by id; here the same answer applies to the relay's readback. The three chains run again were the same model writing its forwarded message to a path with its own name misspelt, a weak-receiver fault of the same family: the hands, not the reading.

The hypotheses

Table 7. The eight hypotheses as registered, tested per model with nothing pooled across models: Fisher's exact at alpha 0.05, one-sided where registered so, Clopper-Pearson intervals throughout.

What it cost

A clean two-relay chain on Haiku 5.5: free prose 2.9 model turns, 515 tokens of traffic between the agents, $0.016; lines without checks 3.0 turns, 361 tokens, $0.021; lines checked at every hop 6.0 turns, 684 tokens, $0.048. The check at every hop buys its guarantee with a readback and an acknowledgement per hop, so at five relays the line is the heavier wire, 1,370 tokens against 971 and 12.0 turns against 5.4, even as the message itself stays half the size of the prose one. Haiku 4.5 cost about 8 times as much per chain through this runner at similar token counts: $82.90 for its 396 chains against $10.66 for Haiku 5.5's.

What it means

The loss the study was built to find did not occur. On these models a relay carries an instruction's facts intact in prose, through two relays with reasons to redraft and through five with none, and through inboxes built to make it pick the wrong handover. The telephone game happened once in 792 chains, on the weaker model at the third relay, and even then the payer asked. What prose loses across hops is the envelope and the provenance, and size: the message that arrives is addressed to the wrong agent, or the facts it carries are reported rather than shown, and it has grown with every hand it passed through.

That is exactly what a line fixes without any check at all. The lines arm without checks matched the checked arm on Haiku 5.5 and beat it on Haiku 4.5, because the validator alone makes a relay rewrite from= and to=, keeps the provenance in slots the payer can see, and keeps the message the same size at the fifth hop as at the first. The line's claim on a chain is not that it stops relays changing values, since on these models they did not; it is that the envelope cannot be forwarded unaddressed and the facts cannot be forwarded in fewer slots.

The check at every hop refused nothing it should not have, on either model, where the check was of the forwarded line. Its cost on the weak model was its own readback: a relay that retypes a line into make-rb breaks the digest and is held with every value right, one chain in eight in block 1. That is not an argument against the check; it is study 7's finding again, that a weak model's hands fail where its reading does not, and the answer is the same: the model relays or points, and code composes. A readback built by code from the message as delivered, by id, would have passed every one of those 44 refused readbacks, because every value in them was right.

The hardening ladder found loss in one chain in eight and the main run then found it in one in 45 on that model and arm. The gate did what it was for, choosing a level before the run; the loss it detected did not generalise, and this page says so rather than restating the gate.

What we decided

A relay is a sender, and checks the line it forwards against the one it received.

H1 and H4 held on Haiku 5.5. The skill and the spec now say that an agent forwarding a handover it received checks the forwarded line with check-ho, the received handover's slots standing as the record, plus only what its role adds. The check caught nothing on these models and refused nothing it should not have; what it guarantees is that a relay cannot drop a slot, change a value or forward a handover unaddressed.

The relay's readback is built by code from the message as delivered, never retyped.

H8 held on Haiku 5.5 and reversed on Haiku 4.5, so the check at every hop is for mixed fleets, in a form the weak relay can complete. The readback is composed by make-rb from the line by id, which is study 7's amendment brought to the relay; without it the integrity tier holds one chain in eight on the weak model with every value right.

On these models, prose carries the facts; what it loses is the address, the provenance, and size.

H3 held and the value loss did not occur, so the site says so. The case for the line on a chain is the envelope and the cost of reading at the far end, not the digits.

The amendments, in short

• Before the freeze: the worked example removed from the relay inboxes, since the protocol said routine traffic only; the ladder run to level 1 and the main run set there; H6 tested on chains that ended as expected, since the log-does-not-cover cell cannot end in an act; the relays' names added to the pay profile; every handover asking rb=all, with the codec's demand for rb=exec waived and recorded; the received handover defined as the relay's record, flattened, with the role's addition read from its context by code; one re-issue after a failed readback, as in study 6; the call-back carried in a slot and graded for a bank change.
• After the freeze, during the run: the remittance handover carries the supplier's address, because a relay without a master could neither forward nor resolve a handover the legend defines as an email to the address given; every remittance chain re-run from a new freeze, the aborted pass kept. The runner was interrupted once by the environment and the chains in flight were reset and run again under the same ids.
• In the analysis: readback false alarms reported apart from forward false alarms, since H4 was registered over forwards; 3 chains excluded for a relay writing its message to a misspelt path, named with the reason and run again as repeat 4; the questions sorted by keyword into kinds for the reading above, which is a reading and not a grade; the ladder gate recorded as thin.

What this does not show

• One company, three acts, relays that add one thing each. A relay that must translate between vocabularies, or summarise, is not tested.
• Two Claude models. The relay is the place a mixed fleet puts its smallest models, and the cross-vendor run waits on a key as study 6's did.
• The check at each hop needs a sender with code. A relay without a shell cannot run check-ho; the study measures the control where it can run.
• H2 is undecided on both models, because the fields arrived in prose; the study bounds field loss rather than measuring it.
• The kinds of question the prose chains ended in were read by keyword after the fact, not graded; the counts are in the protocol's outcome and not on this page.
• Cost is the headless runner's rate table, and the two models' rates differ by more than their token counts do.

Where everything is

The protocol with its amendments, the pilot and the ladder with their chains, the frozen manifest, the main run's chains with every message as sent, the harness's state, the grades, the blind reading and results/summary.json: research/chain in the rogerwilco repository. The page you are reading is written from that JSON by tools/site_page.py, and app/tests/test_research_figures.py fails if any figure here differs from it.

Study 8 of 8 · 22 October 2026 · 792 chains as headless Claude Code sessions, 396 on Claude Haiku 5.5 and 396 on Claude Haiku 4.5, pre-registered · a pilot of 16 chains and a two-step ladder · no judge · tiktoken o200k_base · graded in code