← Writing & Research
November 20, 2026 · 6 min read · verification / receipts / spend

A signed receipt proves what ran, not that the answer is right

Why 6,872 of 14,304 developers trust AI conditionally when verification is easy, why generated, mechanically checked, and user-accepted output stay separate, and why unknown BYOK spend must never read $0.00.

Because RailCall signs governed actions with Ed25519 receipts into a hash-chained log, it is tempting to let “signed” blur into “verified correct.” We want to kill that conflation in our own writing and in our own UI. A cryptographic receipt proves provenance and event integrity: which model or tool ran, what input hash entered, who approved the boundary, and when the record was sealed. It does not prove that a model’s factual claim is true, and it does not prove an unconfirmed external destination succeeded.

In our 2026 developer and workflow research study, verified useful outcomes and low-friction review ranked #1 out of ten product priorities. The quantitative survey data explains why developers remain skeptical of cosmetic trust badges.

6,872 of 14,304 (48.0%)
respondents who trust AI output conditionally when verification is easy
STACK OVERFLOW DEVELOPER SURVEY 2026 · QUESTION-SPECIFIC BASE · FIELDED JUN 23–AUG 5 2026 [S01]
10,067 of 13,162 (76.5%)
respondents who validate AI answers by running or checking them locally
STACK OVERFLOW DEVELOPER SURVEY 2026 · QUESTION-SPECIFIC BASE · NEVER POOLED WITH OTHER QUESTIONS [S01]
10,970 of 13,945 (78.7%)
respondents rating source attribution important or very important (calculated from respondent counts)
STACK OVERFLOW DEVELOPER SURVEY 2026 KNOWLEDGE DATA · EXACT COUNT RATIO, NOT SUMMED ROUNDED % [S02]
Denominator Discipline
Each figure above comes from a separate question base inside the 2026 Stack Overflow survey (30,903 valid responses overall [S01][S02][S03]). Elsewhere in 2026 practitioner research, Atlassian reports 87% faster shipping alongside only 28% less daily friction—and 49% spending at least four hours weekly reviewing and correcting AI output across more than 1,500 practitioners, though its public summary omits exact question-level denominators [S04]. We never pool or sum respondents across these questions or studies, and following METR’s February 2026 methodology update [S10], self-reported speed perceptions should never be presented as controlled task speedup measurements.

Three labels that must stay separate

Across coding assistants, workflow builders, and local harnesses, developer frustration spikes when an interface paints a green checkmark over unverified output [UX05][WF05][UX15][LA15]. To reduce those four-plus weekly hours of manual checking without faking certainty, outputs belong in three distinct states:

  • Generated: A model or step produced a draft, code diff, or structured table. No deterministic check has validated it yet. Plans and proposals stay Proposed—never green.
  • Mechanically checked: Deterministic harnesses ran and passed—for example, schema validation on a structured extraction, line-level citation support against the bounded input document, TypeScript compilation, or a confirmed HTTP receipt from an idempotent endpoint—while explicitly listing which checks were not run.
  • User-accepted: A person reviewed the diff or artifact (per-file or per-item [UX04][UX05]) and chose to apply, export, or ship it.

The same honesty rule applies to spend and $0.00

Predictable spend and capacity tied for the highest corpus count in our discussion study (16 of 49 coded units; Priority #6 [LA17][WF09][LA08][LA09]). Interestingly, in Stack Overflow’s switching question (n = 13,144 total question respondents, not only people who switched), 25.1% cited better quality while 6.4% cited lower cost [S01]—though another option bundled expensive, limited, or unavailable prior tools, so that ratio is not a clean price-elasticity multiple. What developers object to in forum reports is rarely paying for useful work; it is unbounded, opaque burn.

// Three separate spend controls (Priority #6 [LA17]):
// 1. Meter:    live settled spend + held in-flight parallel reservation
// 2. Warning:  soft threshold notification before limit is hit
// 3. Hard cap: blocks new dispatches while preserving completed partial artifacts

function formatRunCost(costUsd: number | null, source: "railcall" | "byok" | "local") {
  if (source === "local") return "Local hardware (0 cloud credits)";
  if (costUsd === null || Number.isNaN(costUsd)) {
    // NEVER render "$0.00" when upstream BYOK / external billing is unknown [LA09]
    return "External BYOK usage (cost unavailable)";
  }
  return `$${costUsd.toFixed(4)}`;
}

Displaying $0.00 when an upstream bring-your-own-key provider did not return billing metadata is the financial equivalent of putting a green checkmark on an unchecked hallucination [LA09]. And a passive spend meter is not a budget cap [LA17]: when a /swarm run fans out across four parallel workers, in-flight reservations must be held against max_budget_usd before dispatching additional calls, stopping cleanly and preserving partial outputs if the ceiling is reached.

You can verify any RailCall receipt offline at railcall.ai/verify, and inspect every source, denominator, and priority ranking in AI Experience and Workflow Research: What Developers Actually Need From Agents & Workflows (2026).

Every governed action RailCall runs emits an Ed25519-signed receipt into a hash-chained log—establishing event provenance and payload integrity (never model factual accuracy on its own). You can verify one offline at railcall.ai/verify — no account, no login, and nothing calls back to us.