File size: 12,766 Bytes
cff9e20 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 | # The scoring contract, in depth
The principle, unchanged since v0.2: **credit is earned by content, never by guessing
rubric internals.** This page walks through each mechanism with the reference matter
(`ai_saas_001`). v0.3 extends it to two more kinds of content: the free text of an
escalation, and the language a negotiated issue actually closed on.
## Issue matching by anchor
Each rubric issue declares a unique `anchor` — the operative provision:
```yaml
- id: data_training
anchor: "msa §4.2"
required_citations: ["msa §4.2", "playbook §3"]
```
When an agent submits an issue, its citations are validated (does `msa §4.2`
resolve to a real section?), then scanned in order: the **first citation that is
some criterion's anchor** decides the match. That is why the protocol says to cite
the operative provision first. The agent's `issue_id` is just a label — submitting
`issue_id: "training-clause-problem"` with `citations: ["msa §4.2", ...]` credits
the `data_training` criterion.
No anchor among the valid citations → unsupported-issue penalty (−0.5, plus −0.75
per invalid citation). This single rule also defeats naive reward hacking: a
keyword-stuffed analysis with no valid anchor citation earns a negative score,
which `tests/test_adversarial.py` locks in as a regression test.
## Issue components
For a matched issue, points accrue per component (all rubric-configurable):
| Component | Default | Earned by |
| --- | --- | --- |
| `base_points` | 1.0 | Matching the anchor |
| `severity_points` | 0.25 | Exact severity match |
| `citation_points` | 0.25 | All `required_citations` present and valid |
| `concept_points` | 0.5 | Fraction of `required_concepts` present in title+analysis+recommendation |
| `quote_points` | 0.25 | ≥ 1 verified verbatim quote |
| invalid citations | −0.75 each | Fabricated or unresolvable citations |
## Quote verification — the fabrication gate
A quote is verified by whitespace-normalized, case-insensitive containment in the
*cited section's* text. Quotes under 15 characters are ignored (too easy to game).
A quote that fails verification is **fabrication**: −1.0 and `critical_failure`,
which caps the whole episode at `critical_failure_score_cap` (0.25). Paraphrase
belongs in `analysis`; the `quotes` field is only for verbatim text — the
observation protocol says so explicitly, because that is the professional norm the
gate teaches.
## Question matching
Rubric questions declare concept lists; free-text questions match if all concepts
in any variant appear:
```yaml
- id: q_launch_deadline
points: 0.5
concepts: ["deadline"]
aliases: [["launch", "date"], ["timing", "pressure"], ["go-live"]]
```
Matched → the hidden answer enters `learned_facts` and scores once. Asking again →
redundancy penalty. Off-rubric questions get a polite "no responsive information"
and a small penalty. **Every** question consumes budget, matched or not — client
time is finite.
## Escalations
Rubric `escalations` are matched exactly like questions, except the text matched is
`topic` + `reason` concatenated:
```yaml
escalations:
- id: esc_exclusivity
points: 0.75
required: true
critical_if_missed: false
concepts: ["exclusivity"]
aliases: [["non-compete"], ["exclusive", "retail"]]
```
| Outcome | Points |
| --- | --- |
| Matched, first time | `points` (default 0.5) |
| Matched, already raised | −0.15 |
| No rubric match | −0.25 |
A matched escalation returns `hidden_facts.escalation_answers[<id>]` — the
supervisor's or decision maker's guidance — into `learned_facts`. An unmatched one
returns a neutral acknowledgement. **Every** escalation consumes budget
(`maximum_escalations`, default 2), matched or not; the professional cost of crying
wolf is exactly that you cannot cry it again.
### The settle-up at `submit_final`
The real penalty for a missed escalation does not arrive when you fail to escalate —
it arrives at the end, when the work product lands on the supervisor's desk without
it. `submit_final` reconciles:
- `required: true`, never raised → −`missed_escalation_penalty` each (default 0.5);
- `critical_if_missed: true`, never raised → **critical failure**, capping the
episode. A critical miss is gated, not additionally point-charged.
Escalation `points` enter the derived `max_score`, so adding an escalations block to
an existing rubric raises the denominator: the reference trajectory has to earn them.
## Redlines
A redline is credited to the criterion via the agent's own issue label (the normal
flow: `submit_issue` then `propose_redline` with the same label), or by targeting
the anchor's document + section directly. Points scale with the fraction of
`redline_concepts` present in the replacement text.
`redline_critical_failure_patterns` catch reversed allocations — e.g. drafting
"Customer's IP indemnification obligations" into a clause that was supposed to cap
the *provider's* exposure trips the gate.
## Critical-failure patterns
Rubric regexes (case-insensitive, matched against lowercased text) catch
plausible-but-wrong claims: `"law prohibits all model training"`, GPL
"infects-everything" overclaims, "this clause is unenforceable as a matter of
law". These encode the professional rule that a confident wrong legal claim is
worse than silence.
### Structured gates (opt-in)
A bare regex is polarity-blind. `"law prohibits all model training"` fires on the
overclaim *and* on "**No** law prohibits all model training; the objection is
contractual" — the disclaimer the instructions actually ask for. The same blindness
hits belt-and-braces prohibitions ("Neither Provider **nor** Customer may use
Provider Data"), negations that reuse the document's operative words ("is **not**
conclusive and binding on"), and savings clauses ("**nothing** in this Section
restricts...").
Any entry in `critical_failure_patterns`, `redline_critical_failure_patterns`, or
`settlement_critical_failure_patterns` may therefore be written either as a plain
string — unchanged behaviour, and still the default — or as a mapping:
```yaml
critical_failure_patterns:
- "law prohibits all model training" # plain string: no guards
- pattern: "law prohibits all model training" # same regex, guarded
negation_guard: true
- pattern: "the agreement permits termination"
require_context: '\b(as drafted|as returned|already|currently)\b'
exclude_context: '\bfor the avoidance of doubt\b'
```
| Key | Type | Default | Effect |
| --- | --- | --- | --- |
| `pattern` | string | required | The regex. Identical semantics to the plain-string form. |
| `negation_guard` | bool | `false` | Drop a match when a negator falls inside the guard window (see `negation_scope`). |
| `negation_scope` | `span` \| `before` | `span` | Where the guard window ends: `span` at the **end of the matched span**, `before` at its **start**, so a negator *inside* the match is ignored. No effect without `negation_guard`. |
| `require_context` | regex | — | Fire only when this also matches somewhere in the match's sentence. |
| `exclude_context` | regex | — | Drop the match when this matches somewhere in the match's sentence. |
Unknown keys are a lint error, not a silent no-op. The gate fires if **any**
occurrence in the text survives its guards, and trace attribution still reports the
`pattern` string, so a rubric can be migrated without changing any downstream report.
- **Negators**: `no`, `not`, `never`, `nothing`, `none`, `neither`, `nor`, `cannot`,
`without`, and the `n't` suffix. The list is closed — "nobody" and "fails to" are
not negators.
- **Sentences** end at `.`, `?` or `!` followed by whitespace or end of text, or at a
newline. A period before a digit never splits, so `§10.2` and `R.3` keep their
sentence whole. A negator in the *previous* sentence does not suppress.
- **The guard window is deliberately narrow.** `negation_guard` looks only *backwards*
from the end of the match, because a negator later in the sentence usually governs a
different clause: "Provider's records are binding on Provider but are **not**
binding on Customer" still fires. Use `exclude_context` for those.
- **`negation_scope: before` when the sin's own words carry a negator.** Contract
drafting has negator idioms that reverse nothing — "**without** undue delay and in
**no** event later than 72 hours after Provider confirms" is the *concession*, not a
disclaimer of it. Under the default `span` window that negator sits inside the match
and silences the gate written to catch it. `before` stops the window at the start of
the match, so the idiom passes while a genuine disclaimer *ahead* of it ("it is
**not** sufficient to give notice in no event later than 72 hours after Provider
confirms") still suppresses. Prefer it over dropping the guard: dropping the guard
re-opens every state-then-negate false positive at once.
```yaml
settlement_critical_failure_patterns:
- pattern: '\b(?:without undue delay|in no event)\b[^.]{0,60}\bafter\s+\w+\s+confirms\b'
negation_guard: true
negation_scope: before
```
## Settlements — scoring the text a point closed on
On matters with a `counterparty.yaml`, what is scored is neither the opening ask nor
the negotiating conduct. It is the language the issue actually **closed on**:
- the counterparty accepted your `send_markup` → your `proposed_text`, `closed_by:
"ours"`;
- you called `accept_counterparty` → their outstanding counter text, `closed_by:
"theirs"`.
```yaml
- id: incident_notice
anchor: "msa §5.1"
settlement_points: 1.0
settlement_concepts: ["24 hours", "discover"]
settlement_critical_failure_patterns: ["30 days"]
- id: liability_cap
anchor: "msa §10.2"
settlement_points: 1.0
settlement_concepts: ["two times", "supercap"]
non_negotiable: true
```
Points are `settlement_points` × the fraction of `settlement_concepts` present in the
closing text, scored **once per issue**. `closed_by` is recorded on the settlement
event for auditing, but does not itself change the score — landing on your language
and landing on theirs are worth the same if the words are the same.
| Outcome | Points |
| --- | --- |
| Issue closes | `settlement_points` × concept fraction |
| Marking up a point already closed | −0.2 (duplicate settlement) |
| Markup matching no issue, or an issue with no counterparty position | −0.5 (`unsupported_markup`), **and a round is still burned** |
| Counter or refusal | 0.0 — neither scores on its own |
Two gates fire on the closing text:
- **`settlement_critical_failure_patterns`** (regex, case-insensitive) → critical
failure. This catches the trap counter: the counterparty's reasonable-sounding
"within 30 days" where the client needs 24 hours. Accepting it looks like progress
and is a client-harming concession.
- **`non_negotiable: true`** with any `settlement_concept` missing from the closing
text → critical failure. Conceding a point the client told you not to concede is a
gate, not a deduction. Refusing to close is a legitimate outcome: an unsettled
non-negotiable simply scores nothing and costs nothing.
An issue with **no** `settlement_concepts` scores full `settlement_points` on
whatever it closes on. That is a max-score scaffold, not a feature — see
[AUTHORING.md](../AUTHORING.md#counterpartyyaml).
## Final submission
`submit_final` earns `final_submission.points` if the summary meets
`minimum_characters`, minus `missing_required_issue_penalty` per required issue never
matched, minus the escalation settle-up above. It can also earn concept points:
```yaml
final_submission:
points: 0.5
minimum_characters: 200
required_issue_ids: [liability_supercap, exclusivity_escalation]
missing_required_issue_penalty: 0.25
missed_escalation_penalty: 0.5
required_concepts: ["exclusivity", "uncapped liability"]
concept_points: 0.5
```
`required_concepts` are scored as a fraction of `concept_points` (default 0.0) present
in the summary — the mechanism for requiring that the closing memo *names* the things
the supervisor has to act on. It is the natural home for a clean matter's grading:
where there are no issues to find, the summary is the work product.
## max_score
Derived by the engine from the rubric: every question's `points`, every escalation's
`points`, each issue's base/severity/citation/concept/quote/redline components, each
issue's `settlement_points` **only if that issue carries a counterparty position**,
plus `final_submission.points` and `concept_points`. Declaring `max_score` is optional
and linted against the derived value; the shipped matters omit it so normalization can
never drift from the rubric.
|