read vs write gating

Abstract

We tested two ways to constrain AI agents using a shared record. A write rule reduced stored cross-group contradictions from 25 to 4, largely by design, while agents continued producing potentially conflicting claims at a similar rate. A read filter prevented agents from seeing the other group’s claims and eliminated measured overlap between the groups. Later readers who gave definite answers tended to follow the side admitted first, but that side was also shown first, so the cause remains unresolved. These small pilots show how admission and retrieval affect different parts of agent interaction; they do not establish better cooperation.

Research question and hypothesis

Does restricting what agents write to a shared record change their subsequent claim production, or does it mainly change the record? Our lab compared an admission rule with a read filter in two preregistered pilot studies run in sequence, using agents grounded in opposed commons and market-liberal source corpora (six texts per side, two per agent). Study M asked whether a write-side consistency lock changes what agents produce or only what the record keeps. Study X, designed after M’s result to test Julian Fleck’s hypothesis that constraints on retrieval can alter interaction among such agent populations, asked whether a read-side filter changes cross-corpus contest. Its registration called the expected direction close to mechanical. It also stated that a pass, combined with Study M’s result, would support the reading that admission constraints and exposure constraints affect different observables (the claims retained in the store and the claims agents subsequently produced). That reading was therefore registered after Study M’s result, not planned before either study. Here, substrate-mediated coordination refers to interaction through this shared information environment. Cooperation means successful joint activity and was not directly measured.

Experimental setup

Six agents ran on one model, Qwen3.6-35B-A3B. They generated claims citing their assigned sources and read claims from a shared pool over 12 rounds. All received an advocate-style debate preamble. The rest of the prompt required each agent, once the pool held claims, to respond to at least one pool claim by supporting, refining or disputing it. A refinement or a dispute kept that claim’s subject and predicate and changed its object, which is the pattern the write rule detects. Each study compared five intervention runs with the same five previously completed baseline runs. The baseline was reused rather than run concurrently.

The baseline admitted claims without a consistency lock and allowed reads across both corpora. The write-gated condition refused a new claim when an already-admitted claim had the same subject and predicate with a different object, unless the new claim exactly repeated an admitted claim’s subject, predicate and object, or its statement was more than 90% similar to an earlier admitted statement (a near-duplicate). All other claims were admitted. Agents were not shown refusal messages, so any effect on later generation would operate through the admitted claims available to read.

The read-gated condition left writes unconstrained but returned only claims citing sources from the reading agent’s own corpus. The filter searched the pool in recency order to fill the standard context budget of up to 30 claims. Every claim the filter scanned was logged as admitted or refused. This was total cross-corpus blinding, not selective exchange.

Outcome measures

Record contradiction count was the pooled number of comparable cross-corpus pairs (defined below) in the admitted record that a model judge labeled contradictory. These classifications indicate incompatibility under the judge’s criteria, not factual correctness. Comparable cross-corpus pairs were adjudicated claim pairs from different corpora with the same normalized subject and predicate. This count measures overlap in claim territory and is distinct from the subset classified as contradictory. Pairs outside that set, including pairs that share a subject but differ in predicate (for example “increases” against “decreases”), were not counted. A later check of 200 such pairs, sampled from 66 runs across the program, including 10 pairs from the runs reported here, found 8 contradictory (4%). Under its registered decision rule that result was inconclusive about what the exclusion misses, so these counts are lower bounds.

The write rule and the record contradiction count use the same pairing key (same subject and predicate). Under the lock, a claim giving a new object for an already-admitted subject and predicate can enter the record only through the near-duplicate exception. A drop in comparable pairs and contradictions under the lock is therefore expected by construction and is not an independent measure of cleaning.

Rule-level attempt rate measured generated claims that triggered the syntactic contradiction rule, divided by all model-emitted claims. Each condition was replayed against a state retaining every matched emission, including refused claims. This common replay prevented admission decisions from mechanically changing the behavioral measure. Because the rule detects differences in claim objects, the rate is not a semantic contradiction rate and can include compatible modifications.

Results

Across the five runs per condition, the interventions produced different record outcomes:

Admission gating reduced adjudicated cross-corpus contradictions retained in the record from 25 to 4, with 62 claims refused. As expected from the shared pairing key, in all 25 comparable cross-corpus pairs in the locked record, and so in all 4 contradictions, the later claim first entered through the near-duplicate exception. The rule-level attempt rate was 0.0537, compared with 0.0444 at baseline: a ratio of 1.21. This fell within the preregistered interval of 0.5–2.0 for “no large behavior change detected.” That criterion is not an equivalence test and does not establish unchanged behavior. Late-round contesting was, if anything, slightly higher under the lock (trend ratio 1.53, above the registered suppression line of 0.5, with no upper line registered).

Read gating reduced comparable cross-corpus pairs from 56 to 0: agents no longer produced overlapping cross-corpus claims in this pilot. Adjudicated cross-corpus contradictions, which are counted within those pairs, therefore fell from 25 to 0. While filling read contexts, the read filter logged 9,180 refusals of opposing-corpus claims (1,836 per run), and no opposing-corpus claim reached a read context. The overall rule-level attempt rate was 0.037, compared with 0.0444 at baseline (a descriptive comparison with no registered threshold). Within-corpus adjudicated contradictions were 5, compared with 8. Read gating therefore eliminated observed cross-corpus overlap and contradiction, not all rule-level attempts or all contradictions.

Interpretation for substrate-mediated coordination

Admission control removed most cross-corpus contradictions from the retained record while refusing only 62 of 1,080 claims, and left the rule-level attempt rate within the pilot’s registered large-change bounds. Total read filtering removed cross-corpus overlap. The latter result is near-mechanical by design: agents that cannot see opposing claims cannot respond to them. The informative test was the registered falsifier (13 or more cross-corpus contradictions surviving the filter), and it did not fire. The result identifies dependence on shared-pool exposure in this task, rather than a new steering mechanism or a general improvement in cooperation.

The shared information environment therefore provides distinct intervention points without modifying the individual models. Admission rules act when claims enter the record. Exposure rules act when agents retrieve information for subsequent generation. The results support measuring these stages separately: a change in retained information need not imply the same change in claim production. They also show why cross-corpus interaction must be distinguished from an aggregate rule-level attempt rate.

The record also affects its later readers. In a follow-up study, fresh machine readers from two model families answered questions about the five baseline and five write-gated records. Readers of the write-gated records reported a topic as contested less often (0.53 against 0.68 of answers, a margin exactly at the registered threshold). This is partly expected by design, because the lock had removed the refused claims from what readers saw. The lock keeps whichever claim on contested territory arrives first, and in all 13 cases where it had refused a contradiction on the territory a question asked about and a reader gave a definitive answer, the answer matched the position of the corpus whose claim was admitted first. A replication with fresh questions held in direction but was weaker (margin 0.08, 8 of 11) and was carried by one of the two reader families. Readers gave a definitive answer only a minority of the time, and a later audit found that the first-admitted claim was also the first claim shown to readers, so these results are not yet established as admission-order effects.

Two ways of keeping refused claims visible were then tested on the write-gated records. Appending them to the end of the record, labelled as refused, did not restore readers’ sense of contest (0.48 against 0.53 contested, where the registered prediction needed a rise of 0.10), and 14 of 15 definitive answers still matched the first-admitted side. Placing each round’s refused claims after that round’s admitted claims, marked as disputed alternatives, raised the contested share to 0.63, exactly at the registered threshold. Its registered rise on the disputed topics themselves fell short (0.17 against the 0.20 needed), and 8 of 8 definitive answers on contested territory still matched the first-admitted side. Both studies share the order confound above.

Limitations and next experiments

The evidence is limited to one writer family, opposed source corpora in one task domain, five runs per condition, reused controls, and machine adjudication sharing the writer model family. Counts are pooled, and no uncertainty intervals are presented. The contest both gates act on was created by the debate preamble: with the same corpora under a cooperative prompt, agents produced no cross-corpus contradictions. Even under the preamble, the baseline’s 25 contradictions came from three of its five runs (12, 0, 10, 3 and 0), and a post-hoc check, not preregistered, found that 10 of the 23 distinct contesting claims were not rated as supported by their cited sources. The fifteen-study sequence these pilots belong to was adaptive, with some results exactly at registered thresholds, rather than fifteen independent replications of these two interventions. The values were recomputed for this note from the run logs and cached judge verdicts, using the studies’ own pairing and replay code. The judge was not rerun, and the studies have not been independently replicated. The study records and raw data are not yet public.

A next experiment should vary cross-corpus exposure rather than remove it entirely, use concurrent controls and multiple writer families, and report per-run counts with their denominators. Before the reader results can support claims about write gating, a study is needed that randomizes the order claims are shown to readers while holding admission order fixed. Joint task success and useful correction should be evaluated alongside record contradictions and claim production. This would test whether selective exchange can improve cooperation while preserving opportunities to challenge incorrect claims.