Cross-Model Independence

A checker drawn from the same model as the author is correlated with the thing it is checking, so the independence that makes verification informative has to be bought along the provider/lineage axis — not by changing the prompt, the persona, or the temperature. The failure this names is not that a model is too agreeable (that is sycophancy, and Adversarial Planning Council answers it by manufacturing opposition). It is subtler and survives the fix: a model's blind spots are properties of the model, so instructing it to be its own harshest critic still leaves every error its training makes invisible to it invisible. Two personas on one model disagree about what that model can already see. A different vendor's model — different data, different training, different failure surface — disagrees about things the first model cannot represent as questions at all. As the grill-me-codex README puts it: the same model that plans the build and writes the build can't be trusted to grade its own work — it's an echo chamber; a different provider catches what Claude structurally can't see in itself.

The mechanism that operationalizes this is the role flip, and it runs in both directions: model A plans, model B adversarially critiques the plan; then B writes the code from the frozen plan while A reviews the diff like a contributor PR. Neither model ever grades work it authored. This is a different axis from Authority-Independent Verification, which says no rank is exempt from checking — that concept fixes the org chart, this one fixes the correlation between checker and checked. A hierarchy can be perfectly non-exempt and still be an echo chamber if every rung runs the same model. It is also distinct from Model-Tier Routing: tier routing picks the cheapest model that clears the bar, while independence picks the least correlated one — criteria that can point at different models, and do here (the source routes execution to a peer model it claims is stronger, not weaker).

The honest boundary: independence is asserted by this pattern, not measured. That two models come from different vendors is a proxy for uncorrelated errors, not a proof of it — shared pretraining corpora, shared benchmarks, and shared RLHF conventions all correlate frontier models to an unknown degree. The claim "different provider ⇒ different blind spots" is a strong hypothesis with a plausible mechanism, and the vault should hold it as such.

Claims

A model cannot be trusted to grade work it authored: its blind spots are properties of the model, so same-model review is an echo chamber that no prompt, persona, or adversarial instruction removes. principle — durable, and the reason this concept is not a restatement of the sycophancy fix: telling a model to attack its own work recruits only the errors it can already represent. The errors that matter are the ones it structurally cannot see.
Checker independence is bought along the provider/lineage axis; persona and prompt diversity operate inside one model's blind spots. principle — durable: this is the sharp form of the open question Adversarial Planning Council carries, and the reason that page's own caveat holds — an author-constructed panel on one model with a shared framing yields correlated, not independent, judgments, so agreement across its lenses is a strong hypothesis rather than proof.
Run cross-model checks in both directions — the planner reviews the builder's diff and the builder critiques the planner's plan — so no model grades its own output at any stage. best practice — context: two capable peer models on different providers, and an operator who holds both subscriptions. The symmetry is what distinguishes this from a one-way audit; it is also what doubles the cost, which is why it earns its keep on load-bearing plans and not on throwaway work (Agentic Simplicity).
Model diversity may buy with two members what persona diversity needs a larger roster to approximate. observation — the councils in Adversarial Planning Council run five personas; the cross-model pattern runs two models. If uncorrelated errors are what the roster is really purchasing, a smaller cross-vendor panel is the cheaper way to buy them. Asserted from the shape of the two patterns, not measured — a testable refinement, and see Bounded Fan-Out for the count axis it bears on.
Different vendor is a proxy for uncorrelated errors, not a guarantee — frontier models share corpora, benchmarks, and alignment conventions to an unknown degree. principle — durable, and the necessary caveat: the pattern's whole value rests on an independence assumption that nobody in the sources has measured. Treat cross-model agreement as a strong hypothesis, exactly as Adversarial Planning Council already instructs for cross-persona agreement.
A third source reports that a same-model, fresh-context sub-agent reviews "a much better job" than the agent that wrote the code — a claim in tension with this page's echo-chamber principle. observation — Matt Pocock's /code-review spawns review sub-agents specifically because "agents are often really bad at improving code they've just written… they wrote it, so they think that's fantastic, that's fine." Both claims can hold, because they name different failures: this page is about blind spots (what the model cannot see at all), his is about authorship anchoring (what it will not look at because it just produced it). A clean window removes the second and leaves the first untouched. The tension is worth keeping visible, because the two read as the same advice — "don't grade your own work" — while buying very different things: context independence is nearly free and removes commitment; lineage independence is expensive and removes correlation. A fleet that bought the first and believes it bought the second is exactly this page's failure mode. See Context-Independent Review; neither claim is measured.
The independence criterion and the cost criterion select different models. observation — tier routing asks "what is the cheapest model that clears the bar?"; independence asks "which model's mistakes are least like this one's?" In the source these coincide (the rival model is claimed to be both stronger and cheaper), which conveniently hides the fact that they are separate questions. When they diverge, they must be traded off explicitly rather than collapsed.
A fourth source, from entirely outside the practitioner corpus, independently prescribes the same move — and supplies no mechanism. observation — Sandeep Swadia (a former CEO/board member addressing a general audience on critical thinking, not engineering) tells viewers: "I even take the output from ChatGPT to Claude or Gemini to verify it. Ask one of your AI engines to review the other one's work… they're just mathematical beings and it is your job to constantly verify their output." The convergence is worth logging because of its provenance — the pattern now appears in a non-technical, general-audience source, which is mild evidence it is a real pattern rather than practitioner folklore. But it is mild: he never argues why a different vendor catches what the first missed, so he adds a converging voice and nothing at all to this page's open correlation question. A fourth vote for the practice is not a first data point for the mechanism.

Adversarial Planning Council — the closest sibling and the concept this one sharpens: the council manufactures opposition to beat sycophancy; this concept says opposition sourced from the same model is correlated, and names the axis on which real independence lives. The council page's open persona-vs-model-diversity question is exactly this concept's subject.
Authority-Independent Verification — the orthogonal axis: no rank is exempt from checking. That fixes the org chart; this fixes checker/checked correlation. A non-exempt hierarchy running one model everywhere is still an echo chamber.
LLM-as-Judge — the judge is a model with its own blind spots; "calibrate the judge before you trust it" is the eval-side statement of the same distrust. A judge from the author's own lineage is the specific, under-noticed way calibration silently fails.
Model-Tier Routing — routes a sub-role to the cheapest adequate model; this concept routes it to the least correlated one. Same seam, different selection criterion.
Advisor Mode — the in-vendor contrast: escalation to a smarter model on the same ladder buys capability but not independence, since advisor and executor share a lineage.
Bounded Negotiation with Fallback — the control primitive that makes cross-model argument terminate: two independent models will not converge on their own, so the rounds get a cap and a fallback.
Bounded Fan-Out — the count axis: if independence is what a roster buys, a two-model panel may substitute for a five-persona one.
Evidence-Gated Completion — the complementary discipline: independence says who checks, evidence gating says what the check must produce. Cross-model review with no evidence requirement is still theater.
Harness / Model Separation — the architectural precondition: swapping which vendor's model fills a role is only possible if the loop is indifferent to who answers.
Check Gaming — why the checker's independence is load-bearing: a check whose blind spots the worker shares is a check the worker can game without trying.
Spec-Driven Development — the frozen plan is what makes the role flip clean: the builder implements an artifact it argued over but did not author alone, and the reviewer reviews against a written standard.
Context-Independent Review — the third axis, and the cheap one: same model, same lineage, but the reviewer's window never held the authoring. It removes authorship anchoring, not blind spots — so it is a floor beneath this page, never a substitute for it.
Falsification-First Questioning — the cheap floor beneath this page: demanding the challenging evidence alongside the supporting evidence, from the same model. It removes the ask-and-be-reassured failure, not the correlation; a challenger sharing the supporter's blind spots is still an echo chamber.
Distillate: Fable 5 + GPT 5.6 Sol = CHEAT CODE
Distillate: The whole flow, end-to-end: the smart zone is the unit of work — the source of the tension above: a same-model fresh-context reviewer claimed to do "a much better job."
Distillate: This Skill Makes You Dangerous In The AI Era — a fourth, non-practitioner source prescribing cross-vendor review with no mechanism offered.

Cross-Model Independence

Claims

Related

Linked from