The ORC Value Standard: a unit of measure and a business model for subscription pricing

Summary

A subscription's price can be read against one public number: what it would cost to rebuild the service lawfully. This standard turns that number into a unit, gives it an error range, prices professional finesse into it, lets the owner's real code inform it under seal, and issues a Value Score only to plans that stay cheaper than replication three years ahead.

The example in three lines. COTBase Premium costs about a tenth of its replication price today. On public evidence that advantage holds for 27 months at 90% confidence. No Value Score is issued yet: the plan is still cheaper at month 36 in 68% of outcomes, and the test requires 90%.

What would earn the score. A price at or below about $9 a month, or sealed proof of a component whose quality takes time to verify, or a service that a replica must keep paying for.

Status. This is version 0.2, for comment. The Index and the instrument do not exist yet, and the example uses declared ranges.

What version 0.2 adds

Version 0.2 turns the rule of version 0.1 into a measuring standard: one unit, an error range on every figure, a priced hurdle for professional finesse, a way to measure the owner's real code without publishing it, and a score that looks three years ahead.

Everything else stands. The standard still prices the lawful route only, advises in one direction only, tests three replicators, separates six layers, reports the floor, and hard-codes no cost.

The unit: price-to-replication

The unit is P/R: what a buyer pays, divided by what it would cost to replicate the same thing lawfully. At 1 the two are equal. Below 1, subscribing is cheaper.

ORC is the price to reverse engineer, as version 0.1 defines it, now stated at a fidelity F and a quality Q. S is the outlay of k users at price p over the horizon. P/R is the inverse of the version 0.1 replication multiple.

Why a unit. Markets compare well when they share one denominator. US lenders have had to disclose the annual percentage rate since the Truth in Lending Act of 1968, so that borrowers can compare credit terms. EU shops must show a unit price beside the selling price to make prices comparable. Investors read a share price against earnings, today and forward. P/R reads a subscription against its replication price in the same way.

Three readings. One formula is read at three dates.

Standard conditions. Two P/R figures can be compared only if they were measured the same way.

Every pricing model converts. Only the outlay changes.

A plan at P/R 0.10 costs a tenth of its replication price. A plan at 1.3 carries a 30% premium over replication, and the label says what the premium buys.

The finesse hurdle

Professional finesse is what a working copy lacks: the gap between doing the same things and doing them as well. Version 0.2 prices that gap.

What economics says. The gap is what has always protected creators. Teece showed that who profits from an innovation depends on how tacit its knowledge is and on who owns the complementary assets. Von Hippel called information that is costly to acquire, transfer and use elsewhere 'sticky'. Lippman and Rumelt showed that when the causes of a firm's performance are ambiguous, imitation stays uncertain.

What AI rebuilds show. The gap is visible in public results. The compiler of version 0.1 cost under $20,000 and still emits slower code than GCC with optimisation off. On GSO, 102 optimisation tasks from 10 codebases, leading agents matched the expert's result in fewer than 5% of attempts when it was published in 2025. On PaperBench, the best agent scored 21% at replicating 20 research papers from scratch. Function arrives first. Quality arrives later, and at a cost.

Quality parity. A replica reaches quality parity when it matches the original on a declared public measure within a stated tolerance: accuracy against public ground truth, speed, error rate, forecast skill. Every ORC figure now names its quality target as well as its fidelity level.

Decision points. Finesse is sized the way function is. Function is counted in function points. Finesse is counted in decision points: choices in the quality-critical logic that move the quality measure and cannot be read from public sources. Constants, thresholds, weights, rule-table rows, ordering rules and special cases all count.

The cost. Four terms are added to the build B of version 0.1.

Proof time. The lag is a statistical fact about the quality measure, not an opinion.

Δ is the time between independent public observations of quality, σ their noise and ε the tolerance. With weekly observations and noise three times the tolerance, proof at 5% significance and 80% power needs 56 weeks, about 13 months. During the lag a replicator pays for a seat on the original and for the attention to compare.

Finesse grades. Until the Index measures finesse, a public assessment assigns a grade and a declared factor.

The rule that follows. Discovery gets cheaper with AI. Proof does not. Finesse an outsider can observe is rediscovered ever more cheaply, so it lengthens the runway by months. Finesse that can only be proven over time puts a floor under the price.

Hidden finesse counts only against a public quality measure. A claim of excellence that nobody can check adds nothing to the score.

Open method, sealed evidence

The method, the instrument and the Index are public. The owner's code never is. A sealed measurement lets the real code inform the score while only inferences leave the owner's environment.

At level A4 the run takes place in front of an accredited assessor, or inside an attested enclave.

Why a seal. Without it the public tier has to guess hidden work and finesse, and the range stays wide. With it the owner trades evidence for a narrower range. Other fields already work this way: the evidence stays private and the conclusion is published.

Four assurance levels. Each label states how much evidence stands behind it.

The instrument. It is open source, runs read-only inside the owner's environment, and does seven things.

1. Hidden work. It counts function points in the code and compares them with the public manifest.
2. Decision points. It lists constants, thresholds, weights, rule rows and special cases in the quality-critical code.
3. Materiality. It perturbs each one and re-runs the quality harness. Only those that move the public quality measure beyond tolerance count.
4. Identifiability. It checks whether the perturbation changes what an outsider can observe. If so the decision is discoverable. If not it is hidden.
5. Interdependence. It samples pairs to see whether one decision's best value depends on another, which makes a search harder.
6. Proof time. It computes the verification lag from the noise of the public quality measure.
7. Binding. It confirms that the measured code reproduces the live service's public behaviour on a random probe set.

Steps 3 and 4 are mutation testing pointed at a new question. They need no judgement, so two assessors get the same counts from the same code.

What leaves, and what stays.

Rules that keep the seal honest.

• No judgement under seal. The instrument is deterministic: same code, same version, same counts.
• A public anchor. Sealed finesse counts only against a quality measure that outsiders can check.
• Consistency. A sealed median outside the public range goes to a second, attested run before it is published.
• Bands, not counts. Public outputs are too coarse to serve as a build plan.
• Commitment. The hash fixes the record on the day of assessment. A dispute opens that record, not a new one.
• Sanction. A false binding or a tampered harness withdraws every score of that owner, publicly.

The error range

Every figure carries an 80% range, and the range says where its uncertainty comes from.

How it is computed. The checker draws each uncertain input from a declared distribution and recomputes the result 100,000 times with a fixed seed. It reports the median with the 10th and 90th percentiles. Anyone who runs the same inputs gets the same numbers.

Where it comes from. The worked example later in this document splits its range as follows.

Shares of the variance of the logarithm of replication price, each input varied alone. They are rounded, so a column may not sum to 100.

Today most of the range is the missing Index. Three years out, half of it is the pace of AI, which no measurement removes.

Three rules.

• A figure quoted without its range is not an ORC figure.
• Declared inputs widen the range. They never move the median in the owner's favour.
• A sealed measurement replaces a declared range with a measured one. It can raise the median or lower it.

The range is itself tested. When a scored product is later rebuilt in public with disclosed costs, the register records where the outcome fell. An 80% range should contain eight outcomes in ten. The Keeper publishes that hit rate every year, so the score earns trust the way a forecaster does: by a record.

The Value Score

The Value Score is the discount to replication price that a subscription still offers 36 months from now, at the cautious end of the range. It is issued only when that discount is positive in at least nine outcomes out of ten.

ORC36 is the replication price of a build that starts 36 months from now, on the Index trend. P10 is its 10th percentile. A score of 40 says: after three more years of AI progress, and even if replication turns out as cheap as the range allows, subscribing still costs 40% less.

Why this test. A buyer can switch to a replica in any month, so the long-term question is whether some future build date beats staying subscribed. For a need that continues, that happens when forward P/R passes 1. The month it does is the runway of version 0.1. The score asks whether the runway is longer than the horizon, with 90% confidence.

Four results. The advice still runs in one direction only.

What every label prints. Replication price with its range. P/R at the three dates. The hold probability: the share of simulated outcomes in which subscribing is still cheaper at 12, 24 and 36 months. The month the value holds until. The finesse grade, the sealed share, the assurance level and the version stamps. Then the Value Score, or the plain statement that none is issued.

What the score leaves out. It counts only what can be priced. Content, trust and anything the seal marks as not replicable appear beside it as the sealed share and are never folded in. That keeps the number conservative, and it means that adding one unreplicable feature does not lift it.

Three ways to earn it.

1. Price at or below the durable price: the highest price the floor supports.
2. Prove under seal a component whose quality takes calendar time to verify, against a public measure.
3. Supply something a replica must keep paying for: licensed data, live operations, continuous access to the original.

The gate is strict on purpose. P/R describes every plan. The Value Score is a seal for the plans whose value survives the pace of AI.

Worked example: COTBase Premium under version 0.2

On public evidence COTBase Premium costs about a tenth of its replication price today and keeps that advantage for 27 months at 90% confidence. It does not yet earn a Value Score.

An illustration, not a certified score. The assessment is at level A1. Nobody has inspected the COTBase code. Unit costs, hidden work and finesse are declared ranges, and every result is provisional.

Inputs. The low and high cases of version 0.1 become the 10th and 90th percentiles. Three inputs are new.

The label.

P/R climbs towards parity because the part of the cost that AI prices keeps halving. The grey mark is the terminal reading: the same outlay against the floor that remains once that part has gone.

What the result says.

1. Subscribe. A personal copy costs about ten times the plan today.
2. The advantage is real but not permanent. Most of today's hurdle is priced in tokens and oversight, and both are falling.
3. What remains at the floor is the user's own attention: an hour to commission a copy and a few minutes a month to mind it. That supports about $9 a month, against a price of $14.
4. Tokens are still the small part, at about a fifth of the cost.

What a sealed measurement could change. The findings below are invented to show the mechanics.

• A seal can cut both ways. Scenario A lowers the owner's number.
• Visible finesse buys months. Scenario B adds three months to the 90% runway and earns no score.
• Proof time changes the result. In scenario C the floor rises from about $335 to about $620, above the $504 outlay, and a score is issued.
• Measurement compounds. A measured Index cuts the spread of today's price from 2.6 times to 1.7 times. A sealed measurement on top cuts it to under 1.5 times, and scenario B then passes narrowly.

The rest of the product. The group and entrant tests of version 0.1 are unchanged: a shared living replica matches the price at 35 to 58 members. iCOT stays outside the score because its algorithm is hidden. If each week's published report can serve as a public check on an intraday estimate, iCOT has a quality measure and a verification lag, which is the profile of scenario C. That is an inference from the public description, not a finding.

The upkeep principle: built value expires

Anything built is a decaying asset. The price of rebuilding it falls every month, so a price that is justified today stops being justified unless its creator keeps renewing the reason. This section is proposed for version 0.3.

The precedent. Value has been made to expire on purpose before. In Wörgl in 1932 a local note stayed valid only if its holder bought a stamp for it each month, at 1% of its value. In post-war Germany a food ration stamp was valid for one month. A built service is a note of the first kind. It keeps its value only while someone keeps paying for the stamps.

The law of decay. The replication price has two parts: the floor, which AI does not make cheaper, and the hurdle above it, which AI prices. The hurdle halves every h months, whether or not anyone copies the service.

The score, dated. The PayOrBuild score is the headroom above the line, at the cautious end of the range, for a build that starts in month t.

Above zero the service is above the line: subscribing is cheaper in at least nine outcomes out of ten. The Value Score is this score read 36 months ahead. A score is always quoted with its date, because it is falling.

The upkeep condition. A service stays above the line only if what it adds each month makes up for what decay removes. Three things count: new replication value m, a rising floor and a falling price.

The right-hand side is the stamp: the rate of decay times the part of the outlay that the floor does not cover. A service whose floor is above its outlay owes no stamp.

Three ways to pay the stamp.

The creator's advantage. The creator pays for an addition once. Every replicator would have to pay for it again. That makes the stamp cheap for a service with many subscribers. It works only for additions the standard buyer uses, because capability nobody uses does not count towards size.

COTBase Premium with no upkeep. On public evidence the score falls from 83 today to the line at month 27.

The line as a price is the highest monthly price at which the plan stays above the line. It falls towards $6, the cautious floor. The plan's $14 is under it until month 27.

What the stamp costs here.

In dollars the stamp is small: one to two months of the plan's price in new replication value each month. In function points it doubles every halving time, and by the fourth year it asks for close to half of today's core every month. Running buys time. Anchoring and following are the upkeep that lasts.

What the label adds.

• Valid until: the month in which the score reaches the line with no upkeep.
• The stamp: the upkeep a month that holds the line.
• The line as a price, today and 36 months ahead.

A score already lapses after 90 days, so a creator who displays the mark renews it four times a year.

The rule. A service that asks for money should keep its score above the line. The standard does not enforce that. It dates every score and prints the stamp, and buyers and their agents do the rest.

Figures in this section come from upkeep.py, a short script on the same model and seed.

The business model

The standard gives away what tokens can copy and sells what they cannot. The method, the instrument and the Index are free. The paid products are independence, a signature, a live register and a track record.

The standard applied to itself. Its own function and data are layers 1 and 2, which the framework says fall towards zero. So nothing is charged for them. Revenue comes from layers 3 to 6: operations, the stream of updates, judgement and trust. Each year the Keeper publishes the P/R of its own paid products.

Five roles.

Nine products.

Who sells what. The Keeper sells the mark licence, the register feed and accreditation. Assessors sell attestation and moat reports. Registers sell buyer tools and run the licence exchange inside their own business.

Four pricing rules.

1. Never sell the direction of a number. Fees are fixed and paid before the result exists.
2. The Keeper assigns the assessor. The owner does not choose.
3. Aim for at least half of revenue from readers of the score, not from those it rates.
4. Publish the standard's own P/R every year.

Why these rules. Credit rating agencies began in 1909 on a subscriber-pays model and moved to issuer-pays in the 1970s, which an SEC commissioner describes as a fundamental conflict of interest. After the financial crisis the SEC was asked to study assigning the rater instead of letting the issuer choose. Consumer Reports has rated products since 1936, accepts no advertising and pays for what it tests. Rules 1 to 3 take the second model and guard against the first.

Why the conflict is smaller here. A rating is an opinion. An ORC assessor witnesses a deterministic measurement that any other assessor would reproduce. Shopping for a kinder assessor changes nothing.

What it costs to run. A public check sizes and prices a product without building anything, so it costs tokens and minutes. An Index reference run costs about what the replication it measures costs: $200 to $1,800 in tokens for a service of 100 to 300 function points, at the declared $2 to $6 a point.

The launch needs no new company. Phase 1 is a feature of businesses that already exist: one consenting product as the first subject and one marketplace as the first register. Where the register's owner also owns a listed product, that product's label says so, shows only A1 or A3 results with every input, and carries no mark until an independent assessor exists.

Incentives, conflicts and gaming

The design works only if the people it measures gain by telling the truth. Four incentives carry it, and every new feature brings an attack to block.

Four incentives.

The same review warns that certifiers do not always report accurately or without bias. The rules below assume they will be tempted.

Ten new attacks. These add to the nine of version 0.1.

Becoming the standard unit

Units of comparison win in three ways: a law requires them, a dominant intermediary adopts them, or a free tool makes them the default. The third path is open now, and buyers' agents are its engine.

How other units won.

Why agents decide it. A person rarely computes build-or-buy. An agent does it every time it is asked to renew a subscription. It needs a number it can read and recompute, and it will prefer vendors that publish one. The orc.json statement of version 0.1 is that number's address.

The adoption loop.

1. The free checker scores products at level A1.
2. Registers show P/R and sort by it.
3. Owners with real finesse seal their code to narrow the range.
4. Sealed scores build the calibration record.
5. The record makes the unit safe to write into procurement rules and contracts.

Four phases.

The counts in phases 2 and 4 are proposals, not forecasts.

Governance, legal exposure and limits

The standard's weak points are its founders' conflicts, the law on comparative statements, and the honesty of a three-year forecast.

Governance rules added in version 0.2.

Legal exposure. This is a map, not legal advice, and it needs a lawyer before any label names a product without its owner's consent.

Limits.

• The forward test is a forecast. A halving time of 6 to 12 months is declared, not measured, and the hold probability depends on it more than on anything else.
• Decision points are a new measure. Two instruments may count differently until a reference suite fixes the rules.
• The seal needs a quality harness. A product with no measurable quality cannot use it.
• The floor rests on what a user's attention is worth. For a buyer who values that time at zero the floor is lower.
• Hidden finesse can be re-invented. An AI that independently builds something as good is a competitor, not a replicator, and no score protects against it.
• A strict gate will issue few scores. That is the price of a seal that means something. P/R still describes every plan.
• The founders are judge and party until the Keeper is independent.

Open questions.

1. Should the forward date be 36 months in every category, or follow each category's pace of change?
2. What tolerance defines quality parity, and who sets it for each category?
3. How coarse must the public bands be to stop leakage?
4. Should a personal label use the buyer's own wage in place of the reference wage?

Roadmap

Six steps take version 0.2 to a first attested score, and the first three need only what the founders already have.

Sources

Every figure in the worked example comes from the value model, a short Python script delivered with this document as orc_value_model.py. It runs with a fixed seed, so the same inputs give the same numbers. The theory it extends is in PayOrBuild.

Units and standards

• Truth in Lending, Comptroller's Handbook, US Office of the Comptroller of the Currency
• Directive 98/6/EC on the indication of the prices of products offered to consumers
• History of the MICHELIN Guide, Michelin
• Principles for Financial Benchmarks, IOSCO, July 2013

Private evidence, public result

• Common Criteria security evaluation, Secura factsheet
• SOC 2 and SOC 3 reports: what is the difference?, Linford and Co
• Guide to the processes of technology appraisal, section 3, NICE
• Confidential Attestation: Efficient in-Enclave Verification of Privacy Policy Compliance, arXiv

Certification, disclosure and conflicts of interest

• Quality Disclosure and Certification: Theory and Practice, Dranove and Jin, Journal of Economic Literature, 2010
• Statement on the removal of references to credit ratings from Regulation M, SEC Commissioner Crenshaw, June 2023
• Agenda for the Credit Ratings Roundtable, SEC, April 2013
• Consumers Union acquires The Consumerist, Consumer Reports, for its funding policy

Imitation and tacit knowledge

• Capturing Value from Technological Innovation, Teece, a later version of his 1986 paper in Research Policy
• 'Sticky Information' and the Locus of Problem Solving, von Hippel, Management Science, 1994
• Uncertain Imitability, Lippman and Rumelt, Bell Journal of Economics, 1982

The quality gap in AI rebuilds

• GSO: Challenging Software Optimization Tasks for Evaluating SWE-Agents, NeurIPS 2025
• PaperBench, OpenAI, April 2025
• Building a C compiler with a team of parallel Claudes, Anthropic

Law

• Directive 2006/114/EC, Article 4, on misleading and comparative advertising
• DOJ's RealPage settlement: a blueprint for 'safer' algorithmic pricing?, Fenwick

Expiring value

• Good ideas that didn't pan out: free money, The Long and Short, on the stamp scrip of Wörgl
• The Wörgl experiment in Tyrol, Cité de l'Économie
• Lebensmittelmarken, transcript, EDU Westfalen, on post-war ration stamps valid for one month