Valeur.A system that writes only what it can source
Writing a risk grade on a contract clause is not the same as being able to show the source that justifies it. Valeur keeps the two apart: which sentence has a reviewed source behind it and which does not is written out separately on every output. This page sets out how that separation is built.
Every behaviour below corresponds to a named rule in the product. Only a sentence that can name its source goes on this page.
FOUR PARTS · THE STANDARD FIRST, THEN THE MECHANISM
A risk score comes only from a reviewed source
There are two ways to put a risk grade on a clause: ask a model and print what comes back, or hold the source that gives the grade first. Valeur requires the second. A finding without a source is still produced — it shows the inventory, the differences between versions, where a phrase occurs — but it arrives without a grade, and the output says so.
The review programme
Contract form packs go through legal review one at a time. The programme is a closed list, and no record out of a pack enters production before that pack’s review is finished. Where the programme stands is not guessed; the output carries it as it is.
Two states, and a human moves between them
A pack has either been reviewed or it has not, and not reviewed is where it starts. The one thing that moves a pack across is a legal review completed outside the software. What advances the state is a human decision, and the software learns of it only from the result.
What a score requires
A finding’s grade — low, medium, high, critical — is assigned only when a sourced signature matches. With no match the finding comes back without a grade, and says as much. A heuristic, a keyword table and a language model help with the reading; none of them may set the grade.
The gate holds shut on a fault
If the step that reads the gate state hits a fault, the gate is treated as shut. On a fault the safe side is always the shut side: an unknown state behaves like a closed gate, never like an open one.
The arrow between them is a human step. What moves a pack from the left box to the right one is a legal review that finishes outside the software; the code learns the move from its result.
- What the programme covers
- Contract form packs, a closed list placed in review order. Packs can be added to the list later; none of them skips the queue unreviewed.
- Where the programme stands
- Every output carries on its face whether the review it rests on is finished. Where it is not, the output does not hide that; it names what is missing.
- Why print this at all
- Because it is the only claim here you can test. “We care about accuracy” is a posture, and nothing tests it. A system that marks, on every output, what is sourced and what is not can be tested: if it is wrong, the error is visible.
LIMIT: No source, no grade. The output still arrives — inventory, differences, where a clause appears — it simply arrives ungraded.
A match is literal text, word for word
There are two ways to find a time-bar clause in a contract: ask a model, or write down the wording you are looking for in advance and show where it appears. Valeur does the second. The plainness of the method is not a shortcoming here; it is the price paid for repeatability.
What a signature is
A search phrase taken out of a pack whose review is finished. Beside it stand the clause family it belongs to, its rule type and its default weight. A record missing any one of those fields is not admitted to the library, and every field arrives from the source.
How a match happens
The clause text and the search phrase are compared with case folded away. If the phrase occurs in the text there is a match, and the place it occurs is shown. The answer is two-valued: a plain text test in place of a probability, a similarity score or a threshold.
Why so plain
So that it repeats. The same contract against the same library gives the same result, and when you dispute a match there is a concrete phrase we can show you. This step runs on the text alone, start to finish.
What it costs
The reach of the method is written down openly: a match is exactly as wide as the judgement of whoever wrote the phrase. A clause that creates the same obligation in different words is caught once those words are written down too. What widens the reach is a growing library.
- What enters the library
- Only a record out of a pack whose review is finished. The library is born empty and grows by no other route — not on a model’s suggestion, and not by hand.
- What happens while empty
- The match count comes back as zero and the library is reported as empty. An empty library is not a fault; it is part of the answer.
- A known gap
- Matching runs on every upload and writes its result into the analysis output. Showing that result on a screen of its own is a named next step in the file.
LIMIT: A clause that creates the same obligation in different words does not match until those words are written down. The method does not hide this; the reach of the library is always something you can measure.
We assume a single counterparty
Any model that works out whether a dispute reaches arbitration has to assume how the other side will behave. The difference here is that the assumption is written down: the model has exactly one counterparty, and every behavioural parameter is stamped ASSUMED on the output.
What it does
It solves the escalation ladder — direct negotiation, DAAB, arbitration — backwards from the end, modelling the other side’s choice between settling and continuing at each rung as a probability curve.
What it assumes
Settlement propensity, rationality, risk attitude and tolerance for dispute cost are collected into one counterparty profile. The solver works with exactly one profile.
A reasoning aid
The solver is deterministic: the same inputs give the same result every time. The output is the arithmetic of the assumptions you supplied.
The amount comes from you
The model takes the disputed amount from your input or from a known case on file. Given one of those it runs; otherwise it returns an explicit error and leaves the calculation there.
- Two separate buckets
- Sourced findings and assumption-driven calculations are written to separate places. What this screen produces stays in its own bucket; verified indicators live in the other.
- The default values
- The default values are illustrative; they do not come from a measured base. The output’s own header says so: they have to be calibrated against your own dispute data before they are trusted.
- Its reach
- One counterparty. The model reads the other side as a single profile, so a multi-party line-up — employer and engineer together — arrives in the same box. Where that distinction matters, read the result alongside human judgement.
LIMIT: The output is the arithmetic of the assumptions you supply. Change an assumption and the result changes; this screen is a reasoning aid, not a prediction.
The same clause reads differently in each legal tradition
The notice period in FIDIC 20.1 may be treated as a hard time bar in one common-law forum and softened by a good-faith rule in another. Valeur keeps that difference intact; it says what happens in each legal tradition separately, and shows the source for each.
- Scope
- FIDIC, 1999 and 2017. Thirteen clause groups: notice and time bar, extension of time and concurrent delay, variations, payment, force majeure, dispute resolution, unforeseeable ground conditions, termination and suspension, completion and defects, contractor design, materials and workmanship, programme, insurance.
- Legal traditions
- Four traditions — common law, civil law, Gulf civil law and mixed systems — with 46 countries mapped onto them. A separate country-level note exists today for exactly one country.
- Sources
- Every tradition cell carries its own source by name and link; the source list holds twelve publishers. Only a sourced interpretation enters the knowledge base.
- Clause with no match
- A clause the knowledge base covers gets an analysis. For any other clause it returns an empty answer and names the reason.
- Known gap
- Delay damages (8.7 and 8.8) are handled today inside the programme group, and the consequences of delay are read from there. A separate group in the same format stands in the file as a named next step.
DRAFT — LEGAL REVIEW REQUIRED
This output is a sourced assessment framework; a definitive legal opinion and a prediction of the outcome belong to qualified counsel. A legal outcome depends on the specific facts, the governing law and the court or tribunal seised, and it is that forum which settles it. Consult qualified counsel before acting.
The stamp is attached automatically to every output. Whether or not an analysis is produced, both branches carry the same warning; an empty answer carries exactly the same stamp.
LIMIT: An answer comes at the level of the legal tradition your project’s country falls into. A country-level note comes into play only once that note has been written with its source.
What buyers actually ask
Can I upload my existing contracts today and get a risk score?
You can upload them. What you get today is a clause inventory and the differences between versions; findings come back without a grade, because a grade comes from a sourced signature. Grading opens when the first pack review finishes — and when it opens, it names the source it came from.
Why show a finding at all if it has no source?
Because a finding without a source is still information; what is missing is the grade. The clause inventory, the differences between versions and where a phrase occurs in the text are true independently of any source, and useful on their own. What would not be true is saying how dangerous that finding is with nothing behind it — so we do not say it.
Why not just answer with a language model?
We do, but labelled. Language model output sits in a separate bucket, stamped unverified, while payment and time calculations feed only on sourced data. What loses a claim file is a sentence that looks sourced while its source stays unnamed.
Can I show the output to my legal counsel?
Yes — it was built for that. The jurisdiction analysis carries its source by name and link in every cell, and the output opens with a draft stamp and a legal review warning. It gives your counsel a framework they can verify, and leaves the decision with them.
Is it useful on projects outside my home country?
The jurisdiction knowledge base covers FIDIC 1999 and 2017 across four legal traditions, with 46 countries mapped onto them. Your answer comes at the level of the tradition your project’s country falls into, with its source named; a separate country-level note stands today for one country, and each new note enters under the same sourcing rule.
Thirty minutes with your own contract
Bring one of your FIDIC contracts. We will go through which clauses we can read and why, where each answer comes from, and exactly what the empty screen looks like.