The Honesty Protocol — A Domain-Agnostic Method for Making Claims Under Uncertainty · Evidence tier: [S] (the method) · [A] the one external regularity it cites (publication bias / the file-drawer effect). Source: 01_EMERGENTISM/02_EPISTEMOLOGY/01_EVIDENCE_TIERS/THE_HONESTY_PROTOCOL_STANDALONE.md

What it is. A portable, theory-independent discipline anyone can adopt: one evidence-tier ladder, three registers, four mandatory fields per claim, a kill-criterion mandate, a standing negative-results register, and an irreversibility gate — then applied to itself, with its own efficacy claim held at [C] and its own kill-criterion stated. Deliberately stripped of all framework jargon; arguably the single most portable public-trust asset here.

The Honesty Protocol

A domain-agnostic method for anyone making claims under uncertainty

What this is. A small, self-contained discipline for keeping your confidence calibrated to your evidence — across a body of work, over time, in public. It is not tied to any theory, field, or worldview. A biologist, an analyst, an auditor, a forecaster, an engineer, or a journalist can adopt it tomorrow.

What it costs. One ladder, four fields per claim, three registers, two gates, and a standing rule that you publish your failures. That is the whole method.

The one sentence [S/I]. A body of work that corrects itself approaches truth; a body of work that protects itself approaches ideology. Everything below is machinery for staying on the first side of that line. This governing sentence is itself a load-bearing claim and carries its tier: the correction-vs-protection split is the protocol's organizing wager ([S]); the normative direction it names — that one approaches truth and the other approaches ideology — is an interpretive reading ([I]), not a measured result.

This document applies the protocol to itself. Every load-bearing rule is tier-marked. The protocol states its own kill criterion (Section 7). The closing section names what would make the protocol itself fail. If a method of honesty cannot survive being pointed at itself, it is not a method — it is a slogan.

Provenance (read once, then ignore). [S] This method is parent-independent in use: nothing below requires any particular theory, and the strip-test in the frontmatter is the standing check on that. It is derivative in origin: it is a distillation of an internal honesty discipline developed inside a larger body of work, stripped of every domain-specific term. Naming that debt is itself an instance of the protocol — you do not get to claim originality you did not earn. The debt does not travel with the method; the method does.

0. The failure mode it targets

Most dishonesty in serious work is not lying. It is tier drift: a claim is born as a guess, gets repeated, picks up confidence with each repetition, and is eventually cited as a fact — without anyone deciding to upgrade it. The drift is silent because nothing in ordinary prose records how sure you were allowed to be.

The protocol is designed to arrest drift by attaching, to every load-bearing claim, an explicit and non-self-upgrading record of its evidential standing — and by making the conditions for changing that standing public in advance. It does not eliminate dishonesty and it makes no such promise; it makes one specific failure mode visible and expensive.

[S] This diagnosis (tier drift as the dominant failure mode) is a structural claim about how claims decay; it is the protocol's own organizing wager, not an independently measured fact.

1. The Evidence-Tier Ladder [A / B / S / I / D / C]

Every load-bearing claim carries exactly one tier. The tier answers a single question: what kind of thing would I have to be wrong about for this claim to be wrong?

The cardinal rule: you may never silently upgrade a tier. A [C] may become [S] or [A] only by meeting its pre-stated upgrade path (Section 3), and the upgrade is recorded as an event, with a date and a reason. Downgrades follow the same discipline. Prose may be vivid; it may not move a tier. When the prose and the tier disagree, the tier governs and the prose is read as invitation.

Why six tiers and not two. "Proven vs. unproven" collapses the three distinctions that actually matter in practice: (1) yours vs. the world's evidence ([B] vs [A]), (2) conditional vs. unconditional truth ([S] vs [A]/[B]), and (3) meaning vs. mechanism ([I] vs [S]). [S] Most overclaiming, on this protocol's own wager, lives precisely in those seams — presenting a conditional result as unconditional, or an interpretation as a measurement. That seam-claim is the ladder's organizing bet, not a measured frequency.

[S] The ladder is a structural device. Its six divisions are defensible and useful, not uniquely necessary; a four-tier or eight-tier variant could serve. Adopt the granularity your domain needs — but never fewer than the three seams above.

2. The Three Registers (never blend them in one sentence)

Confidence drifts fastest when modes of speech are mixed. The protocol forces three registers apart:

The transit rule: a prover-mode claim cannot support a builder-mode decision without passing through interpreter mode — and interpreter mode must be labeled as such. "The model proves X, therefore we should do Y" is the single most common laundering of a [C]/[I] claim into an [A] decision. Make the bridge visible or it is doing dishonest work.

[I] The three-register split loosely parallels the ladder — the prover speaks in [S] necessity, the interpreter in [I] coherence, the builder in [B]/[A] once its results arrive — but the correspondence is an analogy, not an exact structural mapping (the builder's standard is empirical, so it does not sit at [C]/[D]). It is a discipline, not a discovery, and the parallel is offered as a memory aid, not a theorem.

3. The Mandatory Four Fields (the antifragility rule)

A claim is not load-bearing until a serious reader can answer four questions about it. If any field is missing, the claim is not forbidden — but it is not yet allowed to carry weight.

1. Current tier — which rung of the ladder is it on, today?
2. Upgrade path — what specific, pre-stated evidence would move it up a tier?
3. Kill / downgrade criterion — what specific, pre-stated observation would move it down or retract it?
4. Survivors — what remains standing if this claim fails entirely?

The fourth field is the load-bearing one. It severs claims that are wrongly bundled together. No claim may drag the rest of the body of work down by implication alone; when something fails, only the claims that explicitly depend on it fall with it.

The subtraction-first rule. When a claim fails, shrink it before you reinvent it. First narrow the claim to what still survives; only then add replacement structure, and only with an explicit downgrade recorded. Reflex elaboration after a failure is how bad ideas grow armor.

[S] The four-field requirement is the operational heart of the method. Its justification is structural: a claim whose failure conditions are unstated cannot be corrected, and a thing that cannot be corrected cannot be honest.

4. The Kill-Criterion Mandate (and the active/deferred split)

Every major claim must ship with a pre-registered kill criterion: a condition, stated before the result is known, whose occurrence retracts or downgrades the claim. Pre-registration is what distinguishes a kill criterion from a post-hoc excuse.

A kill criterion has four parts:

• The trigger — the specific observation that fires it (operationalized, not vibes).
• What dies — exactly which claim(s) retract or downgrade.
• What survives — exactly what is untouched (forces the dependency graph to be honest).
• Status — not-yet-tested / testing / already fired.

Active vs. deferred kill criteria. Separate the criteria you can test now from those that require future tools, data, or timescales. Mixing them inflates credibility: an untestable "kill condition" is an aspiration, not a commitment. List both, but never let a deferred criterion pose as a near-term integrity test.

Honor your own triggers. The proof that this is a method and not theater is that, when a trigger fires, the claim actually moves. A kill criterion that has fired and been ignored converts the whole protocol into decoration.

[S] Pre-registered, operationalized kill criteria are the published standard for adversarial self-testing. The active/deferred split is the protocol's own structural refinement; its value is that it prevents the file-drawer of untestable commitments from masquerading as rigor.

5. The Negative-Results Register (the file-drawer cure)

The weakest epistemic position any claimant holds is the file-drawer problem: the world sees the mappings, models, and hypotheses that worked, and never sees the ones that were tried and failed. This asymmetry makes any track record look stronger than it is.

The protocol requires a standing, published register of negative results — kept with the same status as the positive claims, not in a footnote. For each entry:

• What was tried — the candidate claim, mapping, or model.
• Why it was tried — the prior reason it seemed promising (this matters; it shows the failure was informative, not careless).
• Why it failed — the specific observation or argument that sank it.
• Verdict — rejected / downgraded / replaced, with date.

A negative result is not an embarrassment to bury; it is a load-bearing asset. It calibrates every surviving claim by showing the denominator. A body of work that publishes only successes is, by construction, miscalibrated — and the reader has no way to know by how much.

The non-negotiable commitment: if stronger evidence shows a favored claim is wrong, you publish the failure. This is the line. A claimant who hides negative results is no longer approaching truth.

[A]/[S] The file-drawer / publication-bias problem is an established, third-party, documented bias in the empirical record (Rosenthal, "The file drawer problem and tolerance for null results," Psychological Bulletin 86(3), 1979, pp. 638–641) [A]; that a standing negative-results register corrects it for an individual body of work is a structural inference, not a measured result [S]. These two halves keep their separate tiers wherever this claim reappears (cf. §7); the [A] half never absorbs the [S] half.

6. The Irreversibility Gate

Some acts cannot be undone: publishing, shipping, deploying, deleting, committing to an irreversible decision on the strength of a claim. The protocol requires a gate in front of every irreversible act, and the gate has one question:

Is the tier of the claim I am about to act on irreversibly high enough for the irreversibility of the act?

Rules of the gate:

• Reversible acts (drafts, internal notes, hypotheses) may run on any tier, including [C]. Speculate freely where the cost of being wrong is cheap.
• Irreversible acts require the supporting claims to be at [B] or [A] — verified, not merely interpreted or conjectured — unless the irreversibility itself is explicitly consented to as a bet, in writing, by an accountable person.
• When an irreversible act rests on a lower tier, it must be staged, marked, and signed off by an accountable human — never auto-approved on the strength of a model output or a conditional result.

The gate is asymmetric on purpose: it is cheap to be wrong in a draft and expensive to be wrong in public, so the burden of evidence rises with the cost of the mistake. The gate does not forbid bold action on thin evidence; it forbids unmarked bold action on thin evidence.

[S] The irreversibility gate is a structural device. It does not tell you what to do — it forces the tier of the claim and the cost of the act into the same sentence, where the mismatch (if any) becomes visible before, not after, the act.

7. The Protocol Pointed At Itself

A method of honesty must survive its own ladder, or it is a slogan. Here are the protocol's own load-bearing claims, tiered, with the protocol's own kill criterion.

The protocol's central claims and their tiers

• [S] The six-tier ladder, three registers, four fields, kill-criterion mandate, negative-results register, and irreversibility gate together constitute a coherent discipline for calibrating confidence to evidence. — Structural: it follows if you accept that uncorrectable claims cannot be honest. The premise is a choice.
• [A]/[S] The file-drawer / publication-bias problem is real and documented [A] (Rosenthal 1979, pp. 638–641 — the one genuinely third-party empirical fact the protocol leans on); that a standing negative-results register corrects it for an individual body of work [S] is the structural inference fenced in §5. (The compound tier is carried verbatim from §5; the [A] half is not allowed to swallow the [S] half here.)
• [C] Adopting this protocol measurably reduces overclaiming and improves calibration in a real body of work over time. — This is a conjecture, not a demonstrated result. The protocol has not been run as a controlled trial against a matched body of work that did not use it. It earns its place by being testable, not by being proven.

The protocol's own kill criterion (pre-registered, active)

• Trigger: A body of work adopts the protocol in full — tiers, four fields, kill criteria, and a published negative-results register — and over a sustained period (say, two years or one major revision cycle) shows no improvement in calibration versus a comparable body of work that did not adopt it; OR the protocol is shown to systematically suppress true claims (chilling correct boldness) at a rate that outweighs the overclaiming it catches.
• What dies: the [C] efficacy claim above, and the [S] claim that this particular arrangement of devices is the right one.
• What survives: the [A] fact (the file-drawer problem is still real), and the weaker [S] claim that some explicit tiering discipline is better than none — the kill criterion above tests this design, not the existence of any tiering at all.
• Status: not-yet-tested. No such controlled comparison has been run. Marking it tested would itself be a tier-drift violation.

The deferred kill criterion (honestly labeled as deferred, not a near-term integrity test)

• A long-run study across many independent bodies of work could, in principle, show that disciplines like this one entrench whatever tiering a community agrees on — including bad tierings — faster than they correct them. That would downgrade the whole approach. The tools to run it at scale do not currently exist; this is an aspiration, not a commitment, and it is listed here precisely so it cannot masquerade as one.

This document's own negative-results register (so §5's rule is not a success-only register that is, by §5's own argument, miscalibrated)

§5 demands a standing, published register of what was tried and failed. A protocol document that shipped only its polished claims would be in violation of its own §5. So, kept here with the same status as the positive claims:

• What was tried: an earlier draft of this document described the method as one that "solves" the file-drawer problem and tier drift. Why it seemed promising: the machinery genuinely does target both. Why it failed: "solves" is an [A]-shaped word for an [S]/[C] result — the method makes one failure mode visible and expensive, it does not eliminate dishonesty, and §0 now says so explicitly. Verdict: downgraded "solves" → "targets / makes visible," 2026-05-31, with the over-claim recorded rather than quietly deleted.
• What was tried: naming the top rung "[A] Asserted / measured." Why it seemed promising: alliteration with the gloss. Why it failed: "asserted" is the lowest-warrant speech act and inverts the very ladder it heads — naming the strongest tier after bare assertion licenses exactly the laundering the protocol forbids. Verdict: renamed "[A] Attested / measured," 2026-05-31.
• What was tried: stating the cost as "six fields per claim." Why it failed: the protocol mandates four fields (§3); the count drifted between the abstract and §3 — precisely the silent-bookkeeping-drift the method exists to catch, caught in the method's own front matter. Verdict: corrected to "four fields," 2026-05-31.

The register is not an embarrassment to bury; it is the only thing that lets a reader calibrate the surviving claims above. Its non-emptiness is itself the evidence that §5 was applied to this page and not merely printed on it.

8. Worked Example (carried at its honest tier)

The example below is deliberately mundane and from no special field, to demonstrate the strip-test: the machinery works without a single domain term. Every claim in the example carries its own tier, and the example as a whole is offered at [D] — it demonstrates the shape of the method; it is not itself evidence that the method works.

A forecasting analyst writes: "Our model shows that customers who open three or more support tickets churn at 4× the base rate; therefore we should auto-escalate any account that hits three tickets."

Run it through the protocol:

What the example does and does not show. [D] It shows the shape of the method on a realistic claim. It does not show that the method produces better forecasts — that remains the [C] claim of Section 7. The example is an illustration, not evidence; treating it as evidence would be precisely the tier-drift the protocol exists to catch.

The whole method, in one breath

Tier every load-bearing claim on [A/B/S/I/D/C] and never upgrade silently. Keep prover, interpreter, and builder apart in any one sentence. Give every load-bearing claim its four fields. Pre-register a kill criterion and honor it when it fires. Keep a published register of what failed. Gate every irreversible act on a tier high enough for its cost. And point all of the above at your own method first — including this document, which carries its efficacy claim at [C] and its own kill criterion in Section 7.

A body of work that corrects itself approaches truth; a body of work that protects itself approaches ideology. This document tries to stay on the first side of that line by submitting to its own rules. Whether it succeeds is, by its own admission, not yet shown.

• ⊙ ○