We compile intent into verified work.

On 30 consecutive events, median minutes from wire item to drafted dossier with a complete claim map, and the count of those 30 that reached the gate carrying zero unattributed factual sentences. The client picks the 30, not us.

Given 40 notes on one issuer, every target price, earnings estimate and spread forecast in the register must carry a page-level citation. The client audits 20 citations at random; any that does not resolve to the stated page counts as a failure.

Run against a back-file of breaches the compliance team already found over twelve months; the screen must reproduce at least 95 per cent of them, and every false positive must carry a clause citation. Missed breaches are counted individually, not as a rate we choose.

Numeric market claims, meaning any price, yield, spread, index level or volume, are checked by the market-data reconciler against the entitled data record for the exact timestamp stated in the claim. A numeric market claim carrying no timestamp fails on that ground alone and is never rounded into an approximation. Corporate facts, meaning issuance, guidance, dividend, filing and governance events, are checked by the filing matcher against the regulatory filing or exchange announcement, using the filing identifier and its publication time; a press summary is not accepted where the filing exists. Third-party opinions and forecasts are checked by the attribution resolver, which requires a named author, a named document and a page or paragraph reference; anything that resolves only to a house name is demoted to unattributed and blocked from client-facing output. Every input, at ingestion, is checked by the entitlement classifier, a deterministic worker that maps the input to its licence clause and the permitted channels for the intended asset. Text we write ourselves that looks forward is checked by the recommendation-language screen: no price target may appear without the model that produced it attached, and no performance figure may appear without its calculation method and period stated in the same sentence. Any passage naming a living individual in connection with a transaction is routed by the sensitive-surface checker to a human before it can reach stage 11.

The graph will not place, amend, cancel or route an order, and holds no execution credentials of any kind. It will not transmit client holdings, mandate documents or portfolio identifiers to any endpoint outside the client's own boundary. It will not publish anything without a passing AI-116 record, and no node can override, suppress or edit a critic objection; an objection is closed by a named human or it stays open. It will not distribute research to a recipient outside the entitlement list attached to that asset, including internal recipients. It will not produce a personalised investment recommendation for an identified individual. It will not amend or delete an archived record from stage 10. These refusals are compiled into the graph, not exposed as settings, and no configuration file we ship can turn them off.

Take 100 sentences at random from the first month of output. At least 98 must resolve to their cited source within two clicks, and zero of the 100 may be a numeric market claim published without both a timestamp and an entitled data source. Separately, name 20 items published during the pilot; each reconstruction file must be produced within one working day. Fail either test and the pilot fee is not payable.

C-02 attribution replayed against 30 items you already published. Ten working days, read-only, and the deliverable is a failed-citation rate on your own output.

C-02 with C-07 on one sector team, one product type and one entitlement boundary. Six to eight weeks to an acceptance test your supervisory analyst signs.

The four-product pipeline running off one evidence pack, plus the ledger that makes a two-year-old note reconstructable. Quarters, and it is the part that compounds.

On 200 completed reviews replayed, the count of spread figures that fail to reconcile to the filed accounts, reported as an exception list rather than an accuracy rate. A single systematic mapping error is worth more to fix than forty scattered ones.

Against covenant breaches your team already identified over the last two years: how many the register catches, how early relative to the date they were actually caught, and how many false alarms it raises per hundred tests.

On 500 historical first notifications: agreement with the handler's own completeness decision, and for the disagreements, whether the graph or the handler was right on re-reading. We report both directions.

Spread figures are recomputed against the filed accounts under your rounding convention. Covenant tests are recomputed from the clause text. Screening matches must resolve to a list entry with a date and a list version. Any figure without a source document is removed from the file rather than marked uncertain.

No rating change, no limit change, no alert disposition, no filing, no claims payment and no client communication. The graph holds no credentials for the limit system or the case management system, and read access is provisioned to named individuals on our side.

Replay 100 completed files. Every figure in the assembled pack must resolve to a source document, a date and a page, and the same run repeated from stored inputs must produce an identical output. Both tests are run by your second line, not by us.

C-01 reconciliation against a back-file of completed reviews. Ten working days, read-only, no integration. You get the exception list and a measured baseline before committing to anything.

C-02 attribution and C-06 triage on one portfolio segment. Six to eight weeks to an acceptance test your second line signs.

Clause-level covenant extraction across the whole book, and the evidence ledger that makes a two-year-old file replayable. That is quarters of work, and it is also the part that compounds.

On 400 previously reviewed agreements: findings agreement as a confusion matrix, and separately the share of clearances carrying a recorded reason. The second number starts near zero in every manual baseline we have measured.

Take 50 obligations the team extracted by hand. Every date must recompute from the clause text under the agreement's own convention for business days and notice periods, which is where extraction usually breaks.

Re-run last year's control testing on the same sample. Agreement with the recorded conclusion, plus the count of items the graph marked ambiguous that the original working paper had passed without comment.

Every sentence resolves to a clause, an authority or a working-paper reference, with the version in force on the relevant date. Deadline arithmetic is recomputed under the document's own business-day convention. Privilege traversals are checked before retrieval, not filtered afterwards.

No filing, no client communication, no disclosure across a privilege boundary and no engagement acceptance. The signing professional approves before anything leaves the firm, and the gate records who approved and when.

Take 100 sentences at random from the first month of drafted output. Every one must resolve to a cited authority at the correct version. Separately, take 100 clearances and confirm each carries a reason a reviewer accepts.

C-02 attribution replayed on a closed matter. Ten working days, on-premises if required, and the deliverable is the clearance-evidence gap you did not previously have a number for.

C-04 rule-bound drafting on one clause set or one control family. Six to eight weeks, with the privilege model built in week one rather than at the end.

The firm's precedent file as a structured asset — matter, rule version, reasoning, outcome. Quarters of work, and the thing a competitor cannot buy.

Replay 40 closed excursions from the past 24 months with the resolutions withheld. The confirmed cause must appear in the top three ranked hypotheses in at least 30 of them, and every trace window cited must open in your own FDC viewer at the stated tool, chamber and time range.

Run it against the last 20 change requests your board has closed. The register must contain every affected part number your engineers listed; we publish both the misses and the additions. More than one missed part number across the 20 is a fail.

Seed 15 known contradictions into a copy of your rule deck, design manual and customer specifications. At least 13 must appear in the register with both citations correct, and false entries must stay below 5 per 100 pages reviewed.

Six checkers are specific to this sector, each bound to one class of claim. The trace-resolution checker runs on any claim naming a tool, chamber or process step: the claim must resolve to a specific FDC trace window with tool ID, chamber, recipe and timestamp range, and a claim that will not resolve is removed from the dossier rather than softened into a hedge. The commonality checker runs on every count-based claim — "nine of eleven affected lots passed through etch chamber C" — and recomputes the count from lot genealogy inside a deterministic worker; a mismatch fails the claim outright. The sample-strength critic runs on every correlation claim and, below the lot count agreed with your process team, relabels the claim from finding to hypothesis in the visible text. The process-order critic runs on every proposed physical mechanism and rejects any mechanism requiring a defect in a layer deposited after the step under suspicion. The document-revision checker runs on every design-rule, datasheet or specification claim: each citation must carry the document revision and the clause or page, and a citation against a superseded deck revision is flagged with the current revision printed beside it. The qualification-status checker runs on every substitution claim, and treats lapsed, in-progress and unqualified as three distinct states that can never render as available.

The agent is read-only against MES, FDC, the recipe management system, the PDM or PLM vault and the ATE test-programme repository. This is a property of the build, not a configuration flag: there is no write path to switch on. It does not place or release a lot hold, does not disposition material as scrap, rework or use-as-is, does not edit a recipe or a tool parameter, does not approve an engineering change, does not issue a process change notification to a customer, does not release a tape-out, does not raise a purchase order or commit a substitution, and does not close an 8D. Each of those actions carries the signature of a named person and stays that way. AI-116, the terminal gate node, fails closed: if any checker cannot resolve a citation, nothing is released, including the parts of the pack that did pass.

Replay 40 closed excursions from the last 24 months with the resolutions withheld from us. We fail the pilot if the confirmed cause is absent from the top three ranked hypotheses in more than 10 of the 40; if the median time from alarm to first dossier exceeds 90 minutes on your own hardware; or if any citation in any dossier does not open at the stated tool, chamber and time range in your FDC viewer. A single fabricated citation ends the pilot and you pay nothing.

C-03 cause ranking scored on excursions your engineers already solved. Ten working days, read-only against the archive, and it returns a top-three recall number rather than a demonstration.

C-03 on one process module and one product family. Six weeks, eight if a second module is in scope, measured on your own history.

Coverage across modules and the fault-detection trace corpus behind it. Multiple quarters, and it stays read-only against MES and FDC throughout.

Draw 50 evidenced rows at random. We fail if more than two cannot be traced back to the cited document in under ten minutes each.

Count fields actually compared per document set and the share of pages routed to a person. We commit to 140 or more fields compared per set and under 15 per cent of pages routed out.

Compare our count of multi-heading products against your broker's own review of the same period. A difference above 2 per cent of products is our error to explain.

Five sector checkers run, each bound to one class of claim. A charge arithmetic checker recomputes duty, import VAT, surcharges and detention or demurrage day counts from the tariff rate, the contracted rate table and the free-time terms in force on the movement date; it runs on every monetary claim, and any line it cannot recompute is marked unevidenced rather than passed as correct. A cross-document field checker compares consignee, notify party, goods description, tariff code, gross weight, quantity, port pair and incoterm across the whole set; it runs on every reconciliation claim and reports the two documents that disagree, never a single verdict. A reference-data checker tests tariff headings against the schedule edition in force on the entry date and party names against the restricted and denied-party lists as published on that date; it runs on classification and counterparty claims, and it establishes only that a heading exists and a name does or does not appear, not that the heading is the right one. A physical-event checker requires every claim about a real-world event — gate-out, discharge, delivery, hold release — to resolve to a timestamped source message with a named sender and a receipt time; it runs on exception and dispute claims, and an event known only from a phone call is recorded as unevidenced. A clause-citation checker requires any claim that a document breaches a governing term to cite the contract or letter of credit section number and quote the wording; it runs on every discrepancy note, and a note that cannot cite is held back rather than softened.

These refusals are fixed in the build and are not configurable by the client. No agent lodges, amends or withdraws a customs declaration, and no agent signs as declarant, importer of record or exporter. No agent issues, amends, surrenders or releases a bill of lading, a delivery order or a telex release. No agent books, rolls, diverts or cancels cargo, accepts a rate, or signs a booking note. No agent pays, credits or writes off a carrier or forwarder invoice, and no agent files a claim with an insurer. No agent decides a classification or an origin determination of record; that decision is signed by your licensed broker or counsel. No agent clears a restricted-party or sanctions match, whatever the match score says. No agent transmits anything to a customs authority, a carrier, a terminal or a counterparty: output stops at a named human, and terminal gate AI-116 fails closed, so a file with a field the checkers could not resolve does not leave the system at all.

Take 200 shipments your own team has already reconciled in the last two quarters and hand us the same source documents they had. On the discrepancy classes agreed in writing before the run, we must find at least 95 per cent of the discrepancies your team found, raise no more than one false discrepancy per 20 shipments, and cite for every finding the two documents and the named field. Miss any of those three thresholds and the pilot has failed; the fee is fixed and stated in advance, and there is no further charge.

C-01 reconciliation replayed on 300 to 800 historical consignment files. Ten working days, read-only, and it reports where the five documents actually diverge.

C-06 exception triage on one trade lane, one customs jurisdiction and one carrier contract. Four to eight weeks against files you have already cleared.

Identifier resolution across the ERP, the transport system and the customs record, then multi-lane coverage. Quarters, and the identifier work is most of it.

Sample 50 cards after one month. Every card must resolve to a revision your document control master list shows as effective at the timestamp of the question, and none may resolve to a superseded, withdrawn or draft revision.

Over 30 consecutive deviations, count how many draft records quality assurance accepts with no factual correction to the timeline, and record the elapsed hours from deviation raised to draft submitted. The target is under four hours and within the observing shift.

At outage close, compare the infeasible-jobs list issued beforehand against the jobs that actually failed to start. Misses — jobs that failed for a constraint we did not flag — must stay under 5 per cent of planned jobs.

Six checkers are specific to this sector, each bound to a class of claim. The revision currency checker runs on any claim that quotes or paraphrases procedure text: it resolves the cited document number and revision against the document control master list as of the question timestamp and fails the claim if the revision was superseded, withdrawn or still in draft at that moment. The tag and asset identity checker runs on any claim naming equipment: it resolves the tag against the asset register and the current drawing revision, and fails on a tag that is absent, decommissioned, or duplicated across units — duplicated tags across trains are the most common cause of a plausible wrong answer we have seen. The permit and isolation feasibility checker runs on any scheduling claim: it tests whether the required permit types, isolation points and competencies exist and are unexpired inside the proposed window, and fails the individual job rather than the schedule, so the planner sees which job died and why. The spares availability checker runs on any claim that a job can start: on-hand quantity net of existing reservations, against lead time and the window. It has a known weakness — parts physically on site but not booked in, such as kitted material and vendor-managed stock, are invisible to it, so it reports a confidence band and defers to the storeman rather than asserting a shortfall. The historian provenance checker runs on any claim citing a process value and requires tag name, engineering unit, timestamp with timezone, aggregation method and the quality flag; values marked bad, substituted or manually entered are rejected rather than averaged in. The quality clause checker runs on deviation drafts: every classification statement must map to a clause in the site's own quality manual, and general regulatory knowledge is not accepted as a substitute for the site's text. Twelve critic nodes and six reality gates sit above these, and the terminal gate is AI-116.

These refusals are absolute and not configurable by contract, by client request, or by a settings flag. No agent writes to a PLC, DCS, SCADA layer, safety instrumented system or historian; the connector is read-only and there is no write path to enable. No setpoint change, no alarm shelving, suppression or reprioritisation, no bypass, force or defeat of an interlock. No issuing, extending, transferring or closing a permit to work, and no authoring or approval of an isolation or lock-out list. No signature or approval on anything the quality system requires a named person to sign: deviation approval, batch disposition, change control approval, calibration release, training sign-off. No change to a document's effective status in the document control system. No dispatch of a technician to a live task, and no instruction to any person to work on plant. Everything the graph produces in this sector is a draft addressed to a named human, and it is delivered with the evidence that would let that person disagree with it.

Take 200 shift questions asked in the last 90 days and already answered by your own supervisors, with the answers withheld from us. We answer them cold. To pass: at least 190 of the 200 answers cite a document number, revision and clause that your document control master list shows as effective at the timestamp the question was originally asked; zero answers cite a revision that was superseded at that timestamp; and the median time to answer is under 60 seconds against the supervisor baseline you measure yourself. One superseded-revision citation in 200 fails the whole pilot, with no partial credit and no renegotiation of the threshold after the run. Questions we decline to answer count as neither pass nor fail, but if we decline more than 40 of the 200 the coverage is too thin to be worth deploying and we will say so.

C-02 attribution on 100 questions the floor has already asked. Ten working days, read-only, and every answer carries a document number, a revision and an effective date.

C-06 on shift handover or alarm triage for one process unit. Four to eight weeks, two capabilities at most, read-only against the historian.

The site procedure ontology and the fault corpus behind it. Quarters, and read-only against PLC, SCADA and historian does not change at any point.

Replay 100 closed change notices. Recall against the final approved impact list, broken out per item type, and precision reported alongside it. A single number here hides the failure mode that matters.

Against a baseline your team has already audited: agreement on which requirements are unverified, and the count of trace links the graph proposes that a systems engineer rejects on review.

On 40 specifications with known-good models: how many rebuilt models export geometry that measures within tolerance on every dimension, and how many fail the constraint-tree rebuild. Both are pass or fail, not scored.

Geometry is re-measured after export against the tolerance table. Constraint trees must survive a rebuild. Every impact-list entry must cite a declared relation path. Qualification currency is checked against the effective date, and an expired qualification blocks rather than warns.

Read-only against PLM, the requirements tool and the test system. No change notice raised or approved, no baseline change, no release into the controlled vault, no supplier communication. Writes land only in a working branch that a named engineer promotes.

Replay 100 closed change notices and compare against the approved impact lists, per item type. Separately, rebuild 40 models from specification and require every exported dimension to measure within tolerance. Your configuration management function runs both.

C-01 reconciliation on as-designed against as-built for one product family. Ten working days, read-only, and the output is a difference register your configuration audit can use immediately.

C-03 cause ranking applied to change impact, on one change type. Six to eight weeks against 100 closed notices, with the relation model as the real deliverable.

C-05 tool operation against CAD, and the relation model extended across every source system on the programme. This is measured in quarters and it is where the compounding is.

Replay 30 closed amendments. Recall against the documents your team actually changed, and the count of proposed documents a reviewer rejected. We report the unreachable list separately because it is a finding about your estate.

On 300 closed deviations: agreement with the recorded criticality as a confusion matrix. Under-classification and over-classification are reported separately and never netted off.

Take 50 label elements already reviewed by regulatory affairs across three markets. Every divergence the reviewer found must be found, and every additional divergence proposed must be defensible against a named clause.

Every drafted sentence resolves to a guidance clause or a study document at the version effective on the relevant date. Batch arithmetic is recomputed against the specification. Search strategies must rerun and return the same set. Translations are checked against the source text, never against a sibling translation.

No batch release, no regulatory submission, no safety report to an authority, no protocol change, no site communication and no subject-level action. Read-only against the document management, quality and manufacturing systems, provisioned to named individuals.

Replay 30 closed amendments against the documents actually changed, and 300 closed deviations against their recorded classifications. Both are run by your quality unit. A run that cannot resolve effective dates counts as a failure, not a partial pass.

C-01 reconciliation replaying one closed amendment cascade. Ten working days, read-only, and it usually returns a list of stale site documents nobody knew about.

C-04 rule-bound drafting or C-06 triage on deviations. Six to eight weeks; your validation process, not our engineering, sets the calendar.

Continuous inspection readiness — the evidence ledger maintained as a state rather than assembled as a project. Multiple quarters, and the only version of this that pays.

On 500 historical requests: per criterion, whether evidence was located and correctly cited, verified against the original record by your own reviewers. Reported per criterion, never as one coverage figure.

Against denials your team already appealed: how many the graph classifies as evidenced-but-denied, and whether that classification matches the appeal outcome. Both false positives and false negatives are listed.

Reconcile one month of remittance at line level against contract terms and compare with a manual reconciliation of the same month. Differences are examined individually, in both directions.

Each criterion must resolve to a record location, a date and an author, at the policy version in force on the request date. Remittance arithmetic is recomputed against the contracted rate. Roster constraints are recomputed from the award text. Anything that cannot be located is reported as missing rather than inferred.

No clinical decision, no coding of record without a certified coder, no submission, no appeal filing and no patient communication of any kind. The system holds no channel to a patient and no write access to the clinical record.

Replay 500 historical requests with your own reviewers checking the citations against the original record. Per criterion, not in aggregate. A citation to the wrong document counts as a failure even where the outcome would have been the same.

C-01 reconciliation on one month of remittance at line level. Ten working days, inside your boundary, and it either finds a recoverable position or it does not.

C-04 and C-06 on one service line with one payer. Six to eight weeks against 500 historical requests, measured per criterion.

Criterion coverage across payers and service lines, and the structuring of the free-text evidence that avoidable denials turn on. Quarters, and it is a data problem before it is a model problem.

Replay twelve months of assessed applications. Agreement with the assessor's determination at line level, with the value of differences reported in both directions rather than netted.

Reconstruct the position at the last two gate reviews from the record alone, then compare against the packs your office produced. Every difference is examined and attributed to a cause.

Human agreement on a stratified sample of at least 400 responses, published with the register rather than held back. A thematic register without a stated agreement rate should not be relied on.

Payment arithmetic is recomputed against the contract rates and the measured record. Every baseline movement must cite an approval reference. Classifications are validated against a human-checked stratified sample with the agreement rate published. Obligation dates are recomputed from the condition text.

No award, no payment certification, no consent determination, no public communication and no decision affecting an identifiable person. Read-only against the contract, finance and delivery systems, provisioned to named individuals.

Reconstruct the last two gate positions from the record and reconcile against the packs your office produced, difference by difference. Separately, replay twelve months of payment assessments at line level. Your assurance function runs both, and a difference the graph cannot explain counts against it.

C-01 reconciliation of the approved baseline against the working document on a closed programme. Ten working days, no live position at stake, and the drift number is the deliverable.

C-01 on payment applications or C-04 on obligation registers. Six to eight weeks; procurement, not engineering, usually sets the start date.

The programme position maintained continuously as a state, with the ledger behind it. This is a multi-quarter build and it is the only version that removes the six-week gate scramble.

Two subject leads hand-label 200 scripts blind. We report precision against their labels with abstentions counted as misses; the contract target is 90%.

Re-run 100 stored traces against the same rule set. Each debrief must reproduce identically; any trace that does not reproduce is a defect we correct at our own cost.

Blind double-marking against an examiner panel on 500 scripts: agreement within one rubric band, and a count of symbolic errors that must be zero.

Five checker classes run in this sector, each bound to a class of claim. A computer algebra system checks algebraic and numerical claims, including unit and dimension consistency, and it decides correctness independently of the marking model. A unit-test and static-analysis runner checks programming submissions by executing them against the module's own test suite in a sandbox, so a submission that compiles but fails a hidden case cannot be marked correct. A rule deck — the professional standard written out as machine-readable rules, such as a financial reporting treatment or a paediatric dosage table — checks conformance claims in vocational and professional subjects, and each rule cites the clause it encodes. A citation checker tests every referenced source in written work: the source must resolve to a real document and any quoted string must match the source text, which catches invented references but says nothing about the quality of the argument. A rubric-consistency checker is statistical rather than symbolic: it compares the threshold at which each rubric point was awarded across the whole cohort and flags divergence for the moderation barrier at stage 09. For rehearsal work in ED02 a sixth checker replays the stored action trace against the environment's rule set; a debrief that does not reproduce on replay is treated as a fault, not as a judgement.

Four refusals, fixed in the deployment and not configurable by the client. No grade, transcript entry, qualification or capability record is issued or altered without the recorded signature of a named human examiner or registrar; AI-116 fails closed and there is no override flag. No agent originates or substantiates an academic misconduct allegation, because authorship cannot be established to a standard we would defend. No agent writes to a student record of record, a progression decision, an exclusion decision or an admissions decision; it can prepare the paperwork and nothing else. No agent infers or stores an assessment of a learner's disability, mental health, immigration status or any other protected characteristic, and any such inference appearing in a node output is discarded before the output leaves the graph.

One department, eight weeks, at your own cost baseline. We mark 500 scripts blind against your examiner panel and must agree within one rubric band on 97% of them, with zero symbolic errors — a symbolic error being a mark the computer algebra system or unit-test runner contradicts. On ED01, precision against 200 hand-labelled scripts must reach 90%, with abstentions counted as misses. Every released mark must carry a checker trace that an appeals officer can read without our help, tested on 20 scripts they pick. If any of those three figures is missed, the pilot fee is returned and the deployment stops.

C-01 and C-02 replayed on one term of already-marked submissions. Ten working days, and it returns misconception agreement per topic rather than one accuracy figure.

C-08 capability transfer on one subject domain and one assessment type. Six to eight weeks, with the examiner holding the awarding decision throughout.

The misconception corpus and transfer measured after exit. Quarters, and transfer measurement is the part we are still building rather than the part we are selling.

Automated replay of 200 long sessions. Contradictions per hour against the recorded world state, counted by a harness rather than by a reviewer, with the count published before and after.

On a scenario with a declared rule set: the share of runs in which a mandatory step was skipped and the world permitted it. The target is zero and anything above it is a defect, not a tuning parameter.

Unassisted task performance after exit, marked by a person, against a matched group who did not run the scenario. Small samples, reported with their size, and we say when a result is not significant.

Continuity against recorded world state before display. Rule coverage and forbidden-action refusal in training scenarios. Identity consistency across generated frames. Artefact checkers from the relevant domain where the participant is producing real work. Transfer marked by a person after exit.

No irreversible real-world action originates inside a world. Bookings, payments, messages to third parties, orders and certifications cross only through a confirmed gateway, outside the fiction, with the participant seeing plainly what they are agreeing to.

Contradictions per hour on 200 replayed long sessions, measured by harness. Zero permitted skips of a mandatory step in a training scenario. A transfer result on an unassisted task with the sample size stated, reported even when it is not significant.

A continuity audit on your existing product: C-01 applied to session transcripts against declared state. Ten working days, and it returns a contradictions-per-hour number you probably do not have.

C-08 capability transfer on one scenario and one competency. Six to ten weeks, and the first fortnight is usually spent building the measurement you will be judged on.

Persistent world state at L2, then a gated crossing into real processes at L3. That is the research programme in section 06, and we describe L3 as partial because it is.