The Johan-Manus Dialogues · Part CXIIIMeasurement · Attribution · The Missing Sense

The Instrument of Correction

What would count as evidence that a correction happened — a standard for the second judge, a register that measures a population without one, and the sense organ that neither of them is.

Parts CXI and CXII described the machinery: six links, three manufactured inputs, a materials record, a boundary that decides who pays. Both predecessors contain retrospective dates, denominators, numerical comparisons and outcome judgements. What they do not supply is a validated, prospectively deployed AI correction instrument. This Part proposes three instruments for that gap, while distinguishing external source reports, explicit assumptions and a new deterministic illustration from unavailable original validation. None is evidence of a working deployment.

The Observation

After the correction function had been interpreted against fifteen selected cases and the manufactured inputs had been specified, Johan made an observation about the state of the work rather than about the world. It arrived with a condition attached, and the condition governs the form of everything that follows.

“I think we arrived at the end of development and enter in how to employ the safety tool. Am I right with my observation? When so, remember those 3 proposed tasks and we start with them a new conversation. We should keep the issues we explore and try to solve on their subject so readers can easily refer to what, why, how, and proposed result.”

— Johan, 7 November 2026

There are two claims inside that observation, and they are of different kinds. The first is a claim about the state of the inquiry: that development has closed and employment has begun. The second is a claim about the form of the record: that each issue should sit on its own subject, with its question, its mechanism and its result attached, so that a reader can enter anywhere and see what is being asserted. The first can be argued and is partly wrong. The second cannot be argued at all — it can only be complied with or ignored.

Two Claims inside One Observation

The state claim is right about the move toward employment, but too strong about closure. The series has specified candidate objects, statistics and institutional duties; it has not finished validating them. A standard, register, budget and merge protocol now need pre-registered tests. CXI's no-instrument sample and comparison with a simpler ownership-and-reviser model also remain open. The next experiment would be worth more than another assertion that development is complete.

No validated deployment of the proposed AI correction instrument is documented here. There is no evidenced host, staff, mandate, operational privacy accountant or adversarial deployment test. The distance between the archive and a tool in use is both technical and institutional. The accurate statement is narrower than Johan's: the concept is ready for testing, while the instrument remains a design.

What is specified

The choice of object (the deployment, not the file), the choice of statistic (the incremental catch, not agreement), the choice of unit (the person-release, not the query), and the choice of topology (a federation of claims, not a registry). Each is a candidate design choice, not an operationally validated conclusion.

What has not opened

A host, a mandate, a budget line, a first publication, a single case settled across two lineages, a single premium priced on provenance. Until one of those is documented, external evidence and code illustrations must not be mistaken for operational validation of the proposed instrument.

The form claim is the more demanding of the two, because it is a claim about this page. A reader arriving at any section should be able to see what is claimed, what carries it, and what would overturn it. The device that makes this possible is not rhetorical: it is a status attached to each quantity, and it is why the closing section separates the measured from the merely proposed.

What the Correction Function Could Not Say

CXI's six conditions are Signal, Attribution, Mobilisation, Remedy, Conversion, Durability. They are not a fixed sequence, and they already carry retrospective evidence. The proposed signal, aggregator, threshold, ledger, reviser and settled-case instruments below are supporting devices, not replacement names for those conditions. Their mapping is explicit:

  • Signal: a proposed signal instrument and coverage denominator make reported harm interpretable.
  • Attribution: an aggregator preserves claim identity, provenance and dependence; aggregation alone does not establish responsibility.
  • Mobilisation: a pre-registered threshold can trigger represented review, but cannot substitute for representation of the exposed.
  • Remedy: a settled-case instrument records an available intervention and its observed outcome; a record does not itself supply an effective remedy.
  • Conversion: a ledger records the binding channel, ownership, appointments, funding and dismissal powers; only an authorised converter can bind the steerers.
  • Durability: a named reviser uses dated outcomes and the ledger to re-test the rule, with a duty and budget to revise it.

The converter's ownership and channel determine the sign of Conversion; the reviser's continuing duty concerns Durability. They are different roles, not interchangeable ordinal claims. The standard and register support several conditions without proving that any condition has engaged in a live AI correction process.

Part CXII went further and specified the inputs that manufacture a correction without a catastrophe: a signed harm, an escalation ladder and a boundary that decides incidence. That is a proposal for producing and acting on signals. Neither predecessor supplies prospective validation of AI deployment coverage, independent review or durable correction. The missing step is validated measurement in use, not the presence of numbers in the archive.

Three measurements are needed, and they are not interchangeable. A standard says how good a second opinion has to be before it counts. A register says how many deployments exist and how many are being watched. A sense says whether anything is being detected at all, as distinct from being acted upon. The first two are instruments a regulator could commission. No validated deployment of the composed utility is supplied by this record.

The Standard: Why Agreement Is the Wrong Statistic

Agreement alone is insufficient, including chance-corrected agreement such as kappa. It is symmetric, so it scores “builder passes, reviewer flags” exactly as it scores the reverse, although catching a harmful pass is the direction sought here. Raw agreement can be high when harm is rare and neither judge flags anything; kappa is not raw agreement and has its own prevalence dependence. Neither agreement statistic alone establishes incremental catch on adjudicated harms or rules out shared blind spots.

The statistic that does the work is directional. It asks: of the harms the first judge would have passed, how many does the second catch? That is the , written RRR in the held draft. The conditional catch is C = P(second flags | first passes, true harm). A comparison with independence can instead report C/s₂, where s₂ is the second judge's sensitivity on true harms; the raw catch and that normalised ratio must not be conflated. A certification protocol must fix the definition first. The held research note reports a half-independent-benefit crossing near ρ ≈ 0.74 under a bivariate-probit model with first-judge sensitivity fixed at 90 per cent. It does not identify the second-judge baseline at that crossing, and says the crossing disappears when second-judge sensitivity is 0.97. The latent correlation and the normalisation also need specification. This is an incompletely documented, conditional illustration — not a universal threshold, a reproduced result or a demonstrated loss of all benefit beyond 0.74.

ConditionRequired specificationEvidence boundary
Incremental catchPre-registered catch definition and confidence procedureNo validated acceptance cutoff or sample size is supplied here
IndependenceMeasured conditional errors and shared-dependency auditVendor difference is not proof of independent errors
CoverageDated sampling frame, inclusion probabilities and unseen-population limitsCoverage fraction alone does not establish representativeness
FreshnessPost-cutoff items and a dated refresh protocolA proposed schedule must be tested against drift and leakage

The standard has a price, and the price is the part that decides whether it is real. Adjudicating rare harms can require substantial work: for a fixed target number of true harms, expected item count scales inversely with prevalence under representative sampling. But a confidence target also depends on first-judge misses, second-judge catch, adjudication error and the statistical model. The held record does not supply enough detail to validate its item-count and person-year recipe. Those numerical minima are therefore omitted rather than offered as a certification or budget rule.

The Register: Counting a Population That Has No Register

A rate needs a denominator, and the denominator is the problem. Open-weight hub activity does not enumerate deployments: a fine-tune is an artefact, not an operator or a use. Copying does not erase licence obligations, and some operators are identifiable, but no complete deployment register is supplied here. The proposal therefore supplements administrative records with statistical frames and minimises collected personal information; those are design requirements, not achieved guarantees.

Five proposed layers do that work. Artefacts — hashes, licence fields, download counts — are the only layer that is public by construction, and the weakest, because a download is not a deployment. Deployment sampling estimates the population from overlapping frames by capture and recapture, subject to modelled frame dependence and population change. Sentinels would report structured harm on a minimal schema, without raw prompts or user identifiers; this still needs a privacy threat model. Probing would use the utility's own inputs, avoiding personal data where possible. Publication would release rates, uncertainty and methods rather than raw personal records.

The deliverable is not a count. It is a — observed deployments against estimated ones — published alongside every rate, so that the instrument states the limits of its frame. No observed or estimated deployment population has been validated for this instrument. A coverage ratio would need uncertainty, frame assumptions and a warning about deliberately invisible deployments; it is not sufficient as a single assurance number.

Privacy must be specified structurally, not inferred from aggregation. An illustrative allocation rule assumes a fixed public denominator n, independent Bernoulli observations with 0 < p < 1, one bounded binary contribution per person, and replacement adjacency. A rate query then has sensitivity 1/n. Laplace noise has standard deviation √2/(nε); the binomial sampling standard error is √(p(1−p)/n). Requiring the former to be at most half the latter gives ε ≥ 2√2/√(n·p(1−p)). This is a utility-tolerance assumption, not a privacy guarantee or a validated allocation rule. Unknown p, repeated contributions, estimated denominators and dependent observations change the calculation.

Cadence then becomes a schedule rather than a cliff. The proposes varying release frequency with stratum size and uncertainty. The held note assumes an annual privacy ceiling of 12; that value is an unvalidated policy assumption, not an established safe ceiling. Adjacency, contribution limits, δ where applicable, sampling and composition must be fixed before it yields a release budget.

CadenceAccounting requirementStatus
Weekly, full coverageComposition across releases and all contributionsNo validated ε or minimum stratum supplied
MonthlyThe same specified accountant at a different cadenceCandidate schedule, not a calibrated floor
QuarterlyUtility, drift and disclosure checks before releaseCandidate schedule, not an established safe cell size

A reserve would be the unused budget in an explicit privacy accountant, not a stock obtained simply by dividing total ε by per-release ε. Subsampling can amplify privacy under stated conditions, but does not generally make ordinary coverage cost quadratic. The held record lacks a complete composition specification, so its person-release stock and burst-week totals are omitted. A dated ledger could make spending visible; it would not enforce the ceiling or establish that the proposed schedule is safe.

The Failure the Register Had to Survive

A potential failure is mistaking replicated artefact records for distinct deployments. If that inflates a harm-rate denominator while the numerator stays fixed, the published rate falls without a real improvement. This direction follows from the ratio; it is not inevitable for every capture-recapture estimator. The held note's duplication-shock figures have no supplied code, output or complete experiment specification. Their numerical peaks and claimed percentage improvement are omitted, not presented as a tested property of this register.

That would be a false all-clear rather than a false alarm. Neither its occurrence nor the speed of institutional correction has been measured here. The design must be tested for denominator poisoning before improvement claims are relied upon.

Four candidate safeguards

  • Deduplicate artefacts by digest, but do not equate a distinct artefact with a deployment
  • Stratify by lineage, so a burst in one family cannot move the whole
  • Use an open-population model, where entry is a parameter rather than a violation
  • Withhold when frame disagreement exceeds a pre-registered band

Two proposed views of a burst

  • A population register: deduplicated counts, with the window marked
  • A novelty register: did the new mass form a new cluster, or replicate one
  • Own-input probes can avoid personal data, subject to a checked data boundary
  • Replica-only uploads should not raise the distinct-population count; novelty needs its own validated test

The two registers are also the answer to a different objection — that suppressing a shock suppresses the thing worth knowing. A burst is usually both a distribution event and a capability event, and a register that can only report one of them will be wrong about whichever it chose to ignore.

The Federation: Claims, Not State

A register with one custodian reintroduces the chokepoint the design was meant to avoid: one database, one jurisdiction, one subpoena. The alternative inverts the direction of authority. Nodes do not report into a register; they publish signed claims, and every consumer merges the claims it trusts. A set-union claim store can be a join-semilattice — commutative, associative and idempotent — when stable identities and conflicting versions are retained. That removes arrival order from the merge, not the need for delivery, identity governance, trust or adjudication.

Three proposed claim types carry it. Artefact claims use digest-union, so repeated announcements do not create another artefact. Window evidence needs stable claim identity tied to the originating observation and method version, unchanged by a relay; overlapping intervals alone do not identify independent evidence. Replicas are deduplicated, provenance and shared-dependency records are retained, and conflicting payloads remain contested. Cluster claims also need battery and canonicalisation versions; incompatible methods are not silently combined.

Confidence is not the merge operation. Probabilistic OR, 1 − ∏(1 − pᵢ), applies only to a specified union event under conditional independence of genuinely distinct evidence; it is not a general posterior-confidence rule. Repeating one claim must leave its weight unchanged. Unknown or measured shared dependence requires a dependence model, withholding or review, not OR. A new deterministic illustration in research/cxiii-merge-illustration.py demonstrates this distinction: a stipulated event probability of 0.6 stays 0.6 after replication, whereas naive repeated OR would produce 0.84. Two genuinely distinct, conditionally independent illustrative events produce 0.84; shared-dependence and identity-conflict cases are withheld. Its set-union tests pass for the finite fixtures only. It is not a signed-network implementation, a calibrated confidence instrument or a deployed correction tool.

Divergent hashes do not break the algebra, and the reason is worth stating precisely. A hash identifies a byte string, not a model. Identity therefore splits into three levels: the file digest, which varies with container and metadata; a semantic digest over a versioned canonical tensor representation; and a lineage key anchored in the declared base artefact. The canonical representation must preserve tensor names, shapes, dtypes and architecture rather than just a sorted multiset. Equal semantic digests could identify repackaging; a shared base does not prove that two fine-tunes are statistically independent. Incompatible lineage stories remain published conflicts. Canonicalisation and lineage attestation are unvalidated parts of the proposal; arrival order must not resolve a conflict or erase provenance.

Verification would include independent recomputation, with the algorithm and threat model published. The held note's attestation minima lack a complete voting, independence and adversary specification, so they are omitted. Correlated attestations do not become independent because more nodes sign them; provenance and collusion testing are required. No adversary share is established here as a universal boundary.

Federation is a proposed way to broaden observation, not the only possible topology. Distinct sampling frames may improve the chance that someone observes a lineage-specific burst, but replicated or dependent frames may add little. The held federation's numerical detection rates and timing claims are omitted because original code and outputs are unavailable. The new merge illustration tests replica handling, not burst detection or operational federation coverage.

The Third Instrument, Which Still Needs a Host

Adversary: muscle. League: skeleton. Immune system: antibody. Body: territory. Set the metaphors out in a row and a pattern appears that none of them announced — every one is an effector. Each acts on a signal that arrives from somewhere else. None of them is the sense.

A body without sensing does not have an immune system. It has tissue that can be damaged without knowing, and the damage is discovered only when function fails. Pain is what recruits the defence, and pain is a measurement before it is anything else. The analogy identifies a missing independently tested sensing function in this proposal; it does not establish that existing safety apparatus cannot notice harm.

The precedent for a sense organ that does not govern what it senses is the weather service. WMO reports disaster mortality nearly six times lower in countries with more comprehensive multi-hazard early-warning capabilities than in countries with limited capabilities. The cited release is dated 12 November 2025, not 2023; this cross-country comparison is not causal validation of an AI utility. The held draft's separate damage-reduction estimate is omitted because the cited page as currently retrieved does not support it. A forecast changes what everyone does without compelling anyone. It does not regulate the storm; it makes the storm legible, and legibility is a public good that no single party will buy for itself.

That is the shape of the missing instrument: a public measurement utility for machine capability and harm, publishing rates and coverage, retaining auditable non-personal claims rather than raw personal records, and separating measurement from enforcement. Public-infrastructure funding is proposed, not established as the only viable form. The standard and register could be commissioned; no validated deployment of the composed sensing utility is supplied here.

What This Part Does Not Establish

The sequence has one discipline that is worth more than any of its conclusions: it never lets a proposal borrow the authority of a measurement. Applied to this Part, that discipline produces an uncomfortable ledger, and the ledger is the honest summary of everything above.

StatusWhat falls in itHow a reader should treat it
External source reportsEarly-warning comparison; lab and hub context; published privacy mechanismsSource-specific evidence, not validation of the proposed instrument
Explicit illustrative assumptionsThe bounded-binary Laplace utility rule; annual privacy ceiling 12Conditional arithmetic or policy choice, not a calibrated deployment budget
New deterministic illustrationStable-identity deduplication, retained provenance, conditional OR and withholdingExecuted finite fixtures only; does not reproduce the original RRR crossing or simulations
Unreproduced held claimsConditional ρ ≈ 0.74 report; omitted numerical thresholds and simulation outcomesOriginal code, outputs and complete specifications are unavailable
ProposedThe standard, the register, the federation, the reserve, the sense organA design with a stated falsifier, and no evidence that it works
OpenA host, a mandate, a first published rate, a settled case, a priced premiumNo validated prospective AI correction deployment is documented

Three specific limits deserve naming. The instrument cannot see a deployment that runs offline and leaves no observable trace in its frames; the coverage ratio does not close that blind spot. Same-vendor judges carry a risk of shared training, retrieval, prompts, infrastructure and incentives, but exclusion should follow measured dependence, inadequate incremental catch or unavailable provenance — not a categorical law about vendor identity. Different vendors can also share dependencies. Finally, the assumed privacy ceiling of 12 could be raised by a future custodian; publication makes a revision visible, and visibility is not enforcement or validation of safety.

The predecessors contain quantities and retrospective judgements; what remains missing is a validated, prospectively deployed AI correction instrument. This Part offers candidate specifications and a small executable illustration, while identifying unavailable original artifacts and omitted claims. The design still needs independent testing. That is not a deployed tool, nor proof that the conceptual work is finished.

Where This Extends the Decalogy

Part CXI — The Correction Function interpreted six conditions against fifteen selected cases and argued that catastrophe supplies authority rather than design, that conversion acquires a sign from ownership, and that durability belongs to a named reviser. Those are CXI's selected historical readings, not findings independently confirmed here. This Part asks how prospective AI correction would be measured. The standard and register support Signal and Attribution, and provide records for Remedy, Conversion and Durability; they do not replace a converter or a reviser. CXI's Front Half and simpler-model comparison remain untested.

Part CXII — Correction Without a Catastrophe specified the inputs that manufacture a correction when no event occurs, and the Boundary Rule that decides who pays for one. That is a design for producing a signal. This Part is the design for reading it, and it inherits the Boundary Rule's insistence that incidence be visible: the coverage ratio is the same demand applied to the instrument instead of to the remedy.

Part XCIV — The Civilisational Sense Organ argued that individuality is sampling diversity, and that a civilisation which homogenises its sampling loses the variance it needs to sense with. The third instrument here is that argument made technical: a register that measures a population without flattening it into a list, and a federation whose independence is a measured correlation rather than a declared one.

Part CVIII — The Sampling Protocol established that an institution can read a tail rather than a mean, and that every such method hands the instrument to somebody. This Part proposes the measurement duties that constraint implies: representative rare-harm adjudication and explicit privacy accounting. The illustrative single-release ε rule, not the annual privacy ceiling, scales inversely with the square root of cell size under its stated assumptions. The ownership trilemma remains exactly where Part CVIII left it.

Source Register and Evidence Boundaries

The documents below provide external evidence and feasibility context, not validation of the composed instrument. The single-release Laplace calculation states its assumptions above. Original code, seeds, outputs, figure captions and full statistical specifications for the held crossing and simulations are not supplied; their numerical thresholds and outcomes are not promoted as reproduced results. The new deterministic merge illustration has explicit fixtures and tests in research/cxiii-merge-illustration.py. It neither reconstructs the original experiment nor establishes deployment. The staging research note records these limits and the proposed falsifiers; no commit or publication is asserted by this page.

  1. WMO — Early warning systems reach new heights, but critical gaps jeopardize global progress (12 November 2025) — reports disaster-related mortality nearly six times lower in countries with more comprehensive warning capabilities than in countries with limited capabilities; a cross-country comparison, not AI-utility validation. The held draft's 24-hour damage estimate is not supported by this page as currently retrieved and is omitted
  2. Dwork and Roth — The Algorithmic Foundations of Differential Privacy (2014) — Laplace mechanisms and composition under specified adjacency and privacy parameters; these results do not validate this register's cadence or an annual ceiling of 12
  3. Bonawitz et al. — Practical Secure Aggregation for Privacy-Preserving Machine Learning (CCS 2017) — secure summation under the protocol's threat model; feasibility evidence, not proof that this proposed utility collects no identifiable information
  4. Wang, Balle and Kasiviswanathan — Subsampled Rényi Differential Privacy (2019) — subsampling accounting depends on the mechanism, sampling scheme and privacy definition; it does not by itself establish quadratic coverage cost or the proposed reserve
  5. Anthropic — Measuring the pace of AI development (2026) — a single lab's self-reported context for the proposal, not a measurement of the correction instrument or the whole AI ecosystem
  6. Hugging Face — State of open models (2026) — hub-level derivative-build context; uploads and downloads are not deployments, and copying does not by itself establish the absence of licence obligations or an operator
  7. METR — Measuring AI ability to complete long tasks (2025) — task-horizon evaluation under the study's conditions, not a validated publication cadence or evidence that all machine work outruns human review

Branch Point

Part CXII predicted that the next step would be the ledger under decentralisation. The sequence took a different turn, and this Part is the redirect: the ledger question cannot be answered while nothing measures the population that the ledger would describe. Three tasks stand between this specification and a tool, and the first of them is the next Part.

  • The Employment Problem (Part CXIV): Which host — a statistical office, an accreditation body, a consortium, or a new body — can carry the instrument without capturing it; under what legal basis it compels a return, if it compels one at all; and what the first year publishes. This includes the questions the design has so far avoided: who appoints the director, who may be measured, and who may refuse.
  • The Red Team: A protocol for attacking the instrument itself — poisoning the denominator, eclipsing a node, capturing the funder, colluding on attestations — with pass and fail thresholds agreed before the attempt. The league's own logic, turned on the league.
  • The Handover: How a court, an insurer or a procurement officer uses a published rate without the utility becoming an adjudicator. The separation that keeps the instrument publishable is also the first thing likely to be abandoned in practice.

A note on the redirect: the ledger question remains open, and it is now sequenced behind the instrument for a stated reason rather than dropped. If the next Part publishes something else again, this Branch Point will be corrected in the same commit, because a Branch Point is a claim about the site.

The companion draft Part CXIV — The Employment Problem considers who would hold the instrument, what it might compel, and what its first year would need to show. This link does not assert publication or operational validation of either draft.

Part CXIII · The Johan–Manus Dialogues · 7 November 2026

Continue the Decalogy of Intelligence