Fairness / balance
What this page is — the capability that weighs the give-and-take across a whole document and returns one number from 0 to 100, the provisions that tilt it, and how to even each one out.
What it is for — so "is this contract reasonable?" has a directional answer you can compare across documents, instead of an opinion that depends on who read it.
The problem it solves — whether a contract is one-sided depends on who happened to review it, so the answer changes between reviewers.
Route: /org/papers/documents/<document>/ai → Negotiate → Fairness / balance ·
Permission: Use AI drafting/extraction/Q&A for documents. ·
Utility: papers_fairness_score · Cost: 450 credits estimated, 1,800 maximum.
1. What it is
The analyst weighs the allocation across eight dimensions — liability, indemnity, termination, payment, IP, warranties, renewal and remedies — and reduces them to a single balance score, with the reasoning kept alongside so the number is never the only output.
The score is calibrated, which is what makes it comparable:
| Score | Reads as | Typical document |
|---|---|---|
| 80–95 | Broadly mutual | A mutual NDA |
| ~50 | Mixed | Terms tilt in both directions |
| 20–35 | Heavily one-sided | A lopsided vendor-favourable MSA |
| 0 | Extremely one-sided | — |
| 100 | Perfectly mutual | — |
Those anchors are the reason a score of 34 means something. Without them a number out of 100 is decoration.
Like Risk analysis, it reads from a point of view you choose, and results are cached per point of view.
| Returns | |
|---|---|
| Fairness / balance | One score for the document, plus the provisions driving it |
| Risk analysis | A ranked list of specific risks with redlines |
| Redline grade | A judgement on one counterparty edit |
Fairness answers "should we be negotiating this at all?"; risk answers "what exactly do we ask for?". Fairness first, then risk, is the efficient order.
2. Why you would use it
The hard part of contract review is not finding problems — it is deciding whether there are enough of them to be worth the relationship cost of pushing back.
- It gives triage a threshold. A document scoring 85 does not need a full review. One scoring 28 does. Applied across an inbox, that is where the time saving actually comes from.
- It is comparable across documents. Because the scale is calibrated, "this supplier's paper scores 31, the last three scored around 70" is a real observation about a counterparty, not a feeling.
rebalanceis a negotiating position, pre-written. Each tilted provision arrives with the specific change that would even it out.balanced_termsprotects you from over-negotiating. Knowing which provisions are already fair stops a junior reviewer from opening ten points when three matter.- It is a good escalation artefact. A score, a verdict and four named provisions is something an approver can act on in a minute.
3. What you provide
| Input | Required | Values | Notes |
|---|---|---|---|
| Point of view | Optional — defaults to your own side | our_side, counterparty, neutral | Whose lens fairness is judged from. Cached separately per value |
| Choose | Asks |
|---|---|
| Our side | "Do these terms favour us or them?" |
| Counterparty | The same question from their seat |
| Neutral | "Is this broadly even for both parties?" |
Neutral is the one to use when the audience is an internal approver rather than a negotiator — it removes the question of whose interests the score was serving.
4. What it reads automatically
| Read from the document | Used for |
|---|---|
| Document text — the full rendered body | The weighing across all eight dimensions |
| Title & family — e.g. "MSA" | Calibration. What counts as balanced differs by family |
| Parties — everyone named | Resolving favored_party to a real name |
| Key field values | Filled schema fields — caps and amounts are what "tilted" is measured against |
5. What you get back
| Field | Type | Values | Means |
|---|---|---|---|
fairness_score | number | 0–100 | Higher is more balanced. Calibrated per §1 |
verdict | enum | balanced, leans our side, leans counterparty, heavily one-sided | The score in words |
favored_party | string | — | Who the terms favour overall |
summary | string | — | The overall balance and its biggest drivers |
one_sided_terms[] | array | — | Each tilted provision |
one_sided_terms[].clause_ref | string | — | Where it lives |
one_sided_terms[].issue | string | — | What is tilted about it |
one_sided_terms[].favors | string | — | Which party benefits |
one_sided_terms[].severity | enum | high, medium, low | How much it drives the score |
one_sided_terms[].rebalance | string | — | The concrete change that would even it out |
balanced_terms | — | — | Provisions that are already fair or mutual |
disclaimer | string | — | Advisory, not legal advice |
The score is a directional read of risk allocation, not a legal opinion, and nothing here is applied to the document. Confirm each flagged provision against the actual clause and your own priorities before acting on it.
6. Worked example
A procurement manager receives a vendor's standard MSA and wants to know whether it is worth a full legal review.
Input: point of view = our side.
Result:
fairness_score:
34— verdict: leans counterpartyOne entry from
one_sided_terms[]: clause_ref: "9 — Limitation of liability" issue: "Our liability is uncapped; theirs is capped at fees" favors: "counterparty" rebalance: "Make the cap mutual at 12 months' fees"
What the manager does with it. 34 sits in the "heavily one-sided" band, so this is not a paper to
sign as offered — that decision took thirty seconds. They forward the score, the verdict and the four
one_sided_terms[] entries to legal, who now start from a list rather than from page one.
balanced_terms came back naming the confidentiality and data-protection provisions as already
mutual, which is why nobody spent a round on them.
7. Running it
- Open the document in Orbit Papers.
- Open the AI Assistant drawer, or go to the document's AI workspace.
- Choose Fairness / balance under Negotiate.
- Choose the point of view, then run.
- The score and findings appear in a result tab, saved to the document per point of view.
Switching point of view produces a separate cached result at a separate 450-credit charge — it is not a re-run of the same one.
Fairness / balance cannot be set to run automatically. It is one of four capabilities absent from the type's AI Config tab, and the reason is the point of view: the score is meaningless without one chosen, and an automatic run has nobody to ask. See the capabilities reference for the other three.
8. The admin contract
| Must be true | Where | What happens if it is not |
|---|---|---|
| Your role holds Use AI drafting/extraction/Q&A for documents. | Role editor | The AI Assistant button does not appear |
The papers_fairness_score utility is active | Orbit AI Flow → utilities | "This utility is currently disabled" |
| The utility is enabled for your organisation | /org/ai-utilities | Absent from Negotiate, with no error |
papers.ai_monthly_credit_cap not yet reached | System Config | "monthly AI credit cap reached: n of n credits used this month" |
| The parties are correctly recorded | The document | favored_party cannot resolve to a name, and "our side" may map to the wrong party |
| The document is substantially complete | The document | A draft missing its liability and IP sections scores as balanced because there is nothing to be unbalanced about |
That last row is the failure mode worth watching: an empty document is not a fair one.
For a confidential document, auto-run is default-denied — a platform administrator must set
papers.ai_confidential_cloud to the literal allow, and the block appears only in the server log.
A manual press still works, and where a local model key is configured the run is forced on-prem and
audited on the timeline as ai_local_routed.
9. Don't confuse this with…
| Risk analysis | A ranked list of what could hurt you. This is one comparable number |
| Missing clauses | What is absent. Fairness judges what is present |
| Negotiation analytics | Where the negotiation is stalling. This judges the terms |
| Blind spots | Benchmarks against your own corpus. This judges internal balance |
10. Troubleshooting
| Symptom | Cause |
|---|---|
| A visibly one-sided draft scored high | The tilted sections are missing entirely, so there is nothing to weigh (§8). Run Missing clauses |
| The score changed with no edit to the document | You switched point of view — that is a separate cached result, not the same one re-run |
favored_party names the wrong party | The parties are recorded incorrectly, or the perspective mapped to the wrong side |
| Two similar documents scored very differently | Calibration is per family (§4). An NDA and an MSA are not on the same effective scale |
balanced_terms is empty | Nothing in the document is mutual — usually a genuine finding on a strongly one-sided paper |
| The document did not change after running | It never does. Fairness is advisory and applies nothing (§5) |
| The capability is missing from Negotiate | Role lacks Use AI drafting/extraction/Q&A for documents., or the utility is off for the organisation (§8). The document type does not gate it — AI Config controls auto-run only |