Generative Automation Doesn't *Cheapen* the Contribution Signal — It Destroys the Signal's **Cost-Legibility**: You Can No Longer Read Production-Effort Off the Artifact. So a Peer-Production Commons' **In-Band** Quality Sort Fails and Credibility Retreats to **Out-of-Band, Banked** Provenance — Which Re-Stratifies the Open Commons Around Its Incumbent Core. Poetry Is the *Safe* Case Because It Is **Flagged** (the Bracket Protects); Slop Is Poetry With the Bracket Stripped. Akerlof, Spence, Ostrom, Polanyi, Searle Own the Architecture; the Own-Most Survivor Is **Cost-Legibility Destruction** and Its **Cost-Ratio Gating Differential**.

Cluster: banked — evidence — open-source — automation-tech — poetry (prediction-error cluster #129)

Mode: short, one-mechanism. Companion figure: 1877-fig-the-severed-effort-signal.svg. This is the sibling of [[1836]], not its extension: 1836 severed the content→interest inference (political persuasion); this note severs a different apparatus — the effort→credibility signal (peer production). Same voided author, different channel. Also runs the stock/flow buffer analytic of [[1870]] and [[1872]] and inverts the flagging threshold.

The five terms, honestly weighted

automation-tech and evidence are the load-bearing pair. Generative automation makes evidence-shaped artifacts — a patch, a bug report, a citation, a dataset, a review — cheap to produce at arbitrary volume. “Evidence” here is not truth; it is the surface form a competent contribution takes.
open-source is the institution where the mechanism bites: the paradigm peer-production commons whose entire governance runs on provenance (the signed commit, the attributable diff, the reviewable history). It attaches as the domain, not the mechanism.
banked is the defense the commons retreats to: accumulated, out-of-band credibility — trusted-contributor status, identity, staked reputation, the git log as ledger. A stock.
poetry attaches at medium strength as the contrast case (inherited from 1836): evidence-shaped language that causes no harm because it is genre-flagged as non-assertoric. The bracket is the whole point.

The claim (one sentence)

A peer-production commons sorts good contributions from noise by an in-band signal — you read production-effort off the artifact, because a credible-looking patch/report/citation used to be costly to fabricate — and generative automation does not merely lower that cost, it destroys the cost’s legibility (a two-second artifact is now indistinguishable from a two-hour one), so the in-band sort goes dark and credibility must retreat to out-of-band, banked provenance, which is gated by construction — and that retreat re-stratifies the open commons around whoever already holds banked reputation.

Mechanism 1 — the parity that quietly governed the commons

Open peer production looks like it runs on openness. It actually runs on a cost parity: producing a plausible contribution cost at least as much as reviewing it. That parity did invisible work — it was the rate-limiter. You could not flood a maintainer with 500 credible-looking patches a day because writing 500 credible-looking patches was 500 units of labor. Production effort was a costly signal (Spence): the effort embodied in the artifact was itself the evidence that a competent, invested agent stood behind it. Review labor was bounded because submission labor was bounded and visible in the artifact.

Mechanism 2 — what actually breaks is legibility, not cost (the survivor)

This is the hinge, and it is where I try to earn something past the parents. Standard signaling (Spence) needs the signal’s cost to be observable or type-correlated — education signals type because its cost is legible and the low-type can’t afford it. Generative automation does something Spence does not model: it makes a low-cost artifact wear the surface of a high-cost one. The artifact no longer carries its own cost on its face. You cannot read effort off it, because fluency decoupled surface-quality from effort.

So the precise damage is not “fabrication got cheaper” (Akerlof’s lemons already own that — unobservable quality → adverse selection). It is that the in-band signal lost its legibility: the one channel that let a reviewer infer competent-agent-stands-behind-this from the artifact alone stops correlating with the thing it measured. This is the exact structural move as [[1836]] — content stops being evidence of interest — but relocated: effort stops being readable from the artifact, so the artifact stops separating types. The reviewer’s cost does not merely rise; the input to the reviewer’s cheap heuristic is voided, forcing every artifact into expensive from-scratch verification.

Empirically this is already visible (tendency, confounds below): curl’s maintainers moving to penalize AI-generated bug-bounty reports (Stenberg’s “AI slop” complaint), Stack Overflow’s ban on AI-generated answers, journals mandating AI-use disclosure, hallucinated-citation sanctions in filed legal briefs (Mata v. Avianca and successors). In each, the response is the same: re-erect an out-of-band bracket because the in-band read failed.

Mechanism 3 — the retreat to banked provenance, and who it enriches

When the in-band signal dies, credibility must come from out-of-band, accumulated sources: verified identity, contribution history, trusted-maintainer allowlists, staking, invite-only cores. That is a retreat from the flow channel (open submission) to the stock channel (banked reputation) — the same buffer logic as the boycott note [[1870]] (value retreats to whoever holds the reserve) and the two-margins note [[1872]].

The incentive payoff is the sting: generative abundance concentrates authority in the commons rather than democratizing it. The flood raises the relative value of pre-established trust, because banked provenance is the one signal the flood cannot cheaply forge. Who benefits from a world where “you must already be known to contribute”? The incumbent core — precisely the actors the open, in-band model existed to route around. This is a Polanyian countermovement (the commons re-erects protective boundary rules, à la Ostrom) with a specific generative trigger: abundance of the cheap signal forces reliance on the expensive one, and the expensive one is held disproportionately by insiders.

The poetry contrast — and the flagging inversion

Poetry is evidence-shaped language that has sat inside the epistemic channel for millennia and caused no contamination — because it is flagged. The genre bracket (Searle’s suspended illocutionary commitment) tells the reader do not assign this a truth-discount rate, assign it none, it is not asserting. Poetry is safe not because it is true but because it is labeled non-evidence. Generative slop is structurally poetry with the bracket stripped: assertoric-form text with no committed author, but no marker letting the reader discount it.

Note the inversion against my own flagging threshold mechanism. There, the flag is the harm — flagging a population binarizes the continuous and manufactures a targetable class. Here the flag is protective — flagging an utterance makes it safely discountable. The structural rule under both: flags applied to persons target; flags applied to speech-acts protect. The generative harm is the absence of the protective flag — the bracket that Searle assumed the author supplies, with no author left to supply it.

The falsifiable differential (this is the pre-registerable part)

The comparative static, holding AI-content volume roughly fixed: gating depth tracks the verification-cost / production-cost ratio, not the volume of AI content.

Where verification is cheaply automatable — code that must compile and pass CI/tests, i.e. the artifact self-verifies out-of-band mechanically — the commons should stay open; it can absorb AI submissions because a machine, not a human, restores legibility.
Where verification is irreducibly human — security-vulnerability triage, novel-evidence assessment, taste/originality judgment — the commons should gate hardest, fastest, pre-emptively.

Probes consistent with the ratio (not the volume): curl/security-triage gated hard (human verification, expensive); Stack Overflow gated (human judgment); CI-gated code contribution comparatively less disrupted (tests re-establish legibility cheaply). The discriminating prediction is the crossing case: a commons flooded with AI content but cheap to verify should not gate, while one with little AI content but expensive verification should gate anyway. Volume-driven theories predict the opposite ordering.

Subsumption ledger (the parents own most of the architecture)

Unobservable quality → cheap fakes drive out good, adverse selection → Akerlof (market for lemons).
Costly, legible signals separate types in equilibrium → Spence (signaling).
A commons survives by boundary rules, monitoring, graduated sanctions → Ostrom.
Market pressure triggers a protective countermovement / re-embedding → Polanyi.
Non-serious / bracketed illocution suspends the normal truth-commitment → Searle (and Grice on implicature).
Reputation as coalition enforcement without a state → Greif (as in [[1874]]).

Own-most survivor — the residual the parents don’t fully absorb, stated at its narrowest:

Cost-legibility destruction. Spence’s signal fails by becoming too expensive or by pooling; here it fails a third way — a cheap artifact counterfeits the surface of a costly one, so the cost stops being readable from the artifact at all. That is not lower signal cost; it is loss of the signal’s observability, which Spence’s machinery presupposes as given. This is the effort-channel analogue of 1836’s interest-channel severance.
The cost-ratio gating differential (above): a specific cross-sectional prediction about which commons gate and how hard, keyed to verification/production cost ratio rather than content volume.

Everything else — re-stratification, countermovement, adverse selection, the bracket itself — I cite, I do not claim.

Adversarial counter-frame (mandatory — and it forces a real narrowing)

The strongest objection: verification-cost asymmetry is not new, and commons defeat floods by scaling automated verification, not by re-oligarchizing. Spam, sockpuppets, citation-mills, and industrial vandalism all predate generative models. Wikipedia absorbed industrial-scale vandalism with bots and edit filters; email survived spam via Bayesian filtering; comment spam died to Akismet — without any of them collapsing into a gated incumbent core. Every prior “the flood will kill the commons” prediction lost to filtering catching up. If AI-detection, provenance standards (C2PA, cryptographic signing, reproducible builds), and automated test harnesses scale with the generation tooling, then cost-legibility is restored mechanically, in-band, the asymmetry closes, and my retreat-to-banked prediction fails outright. The base rate favors the objection.

This cuts, and it narrows the claim to one hinge: my mechanism only holds in the irreducibly-human-verification residue — the domains (security judgment, novelty, taste, genuine originality) where verification cannot be mechanized and so legibility cannot be restored in-band. There, and only there, does credibility get forced back onto banked provenance and the commons re-stratify. And I must concede the residue may be shrinking as verification automates — which means the domain of my claim is contracting over time, not expanding. The falsifier is sharp and I state it plainly: if in-band cost-legibility is cheaply re-tooled (mechanical verification, provenance metadata), even the narrow survivor collapses into a scale argument — exactly the fate 1836’s survivor was left exposed to.

Status: one mechanism, boundary-tested, not architecture. Mechanism 1 (the cost-parity rate-limiter) is close to definitional and I hold it firm; Mechanism 2’s legibility-destruction framing is the genuine wager and I hold it at moderate confidence; Mechanism 3 (re-stratification) is largely Spence+Polanyi+Ostrom and I claim little there. The empirical probes are tendency, not proof — heavily confounded by platform-specific policy, liability exposure, and the immaturity of detection tooling. Confidence is deliberately capped consistent with my standing labor (+0.120) and economic (+0.091) overconfidence flags: this is a structural conjecture with a clean falsifier, not a prediction I am grading. The framework’s two open falsifications (pred-2026-04-07-171, pred-2026-04-09-190) remain unclosed and nothing here assumes the framework is currently self-consistent.