Password required

This draft is private. Enter the password to continue.

Incorrect password.
Tranche 2 menu
Internal draft · not yet shared with CHIRP

Insight Engine Tranche 2: costed menu

Every item priced in days at £500 a day. Each build day comes with 0.7 days of contingency, project management and ZigZag product management (investigation bodies: 0.4 days, no project management). Pick items, or take a bundle.

Basis

Item Value
Day rate £500
Build 1 day
Contingency 0.2 days per build day
Project management 0.3 days per build day
ZigZag product management 0.2 days per build day
Total days 1.7 days per build day (£850)
Investigation bodies 1.4 days per build day (£700): build, contingency and ZigZag product management, no project management
Excluded Running costs (hosting and AI usage), paid by CHIRP

Bundles

Baseline is 10 more official investigation bodies at full ingestion, chosen from section F. Bronze adds every other Must, Silver adds every Should, Gold adds every Could. Each includes the one above it.

What the costs include

Every build day comes with 0.2 days of contingency, 0.3 days of project management and 0.2 days of ZigZag product management: 1.7 days in total, at £500 a day. The investigation bodies are the exception: no project management, so 1.4 days in total.

Product Manager
  • User requirements capture and validation
  • Product research
  • UI/UX design and research
  • Usability testing
  • Product marketing
  • AI Foundations Training
  • Maritime Safety Stakeholder Workshop
  • Project management software subscriptions
Bundle and contents Build days Total days Cost Running total
BaselineOfficial investigation bodies 50 70 £35,000 £35,000
Ingest reports from 10 more national investigation bodies, chosen from section F, full ingestion at 5 days each contingency, no PM 50 70 £35,000
BronzeBaseline plus every other Must 121.5 206.55 £103,275 £138,275
Chat Musts (section A) 49 83.3 £41,650
Public launch platform (section C) 29.5 50.15 £25,075
Capture more metadata, with one shared vocabulary across jurisdictions (section D) 14 23.8 £11,900
Analyst tool: find what keeps recurring across all reports, with the supporting cases, and a first-draft outline of a write-up (section B) 8 13.6 £6,800
Analyst tool: proposed coding and similar cases for each report CHIRP receives (section B) 6 10.2 £5,100
Ingest MAIB incident webpages, with or without a report (section D) 4 6.8 £3,400
Automatically ingest new reports from every source (section D) 4 6.8 £3,400
Ingest CHIRP Maritime’s own published reports, a new report format (section D) 3 5.1 £2,550
Ingest MAIB Safety Digests, a new report format (section D) 2 3.4 £1,700
Link reports that describe the same accident, across countries and digests (section D) 2 3.4 £1,700
SilverBronze plus every Should 75 127.5 £63,750 £202,025
Ingest six short-form sources: IMO lessons, MARS, SAFETY4SEA, IMCA, USCG, Marine Safety Forum (section D) 15 25.5 £12,750
Dashboards filtered by vessel type, date, outcome and country (section B) 12 20.4 £10,200
Link findings to the laws they relate to: UK/MCA, MLC, IMO conventions (section D) 12 20.4 £10,200
Chat Shoulds (section A) 20 34 £17,000
Design how short-form sources are stored, and build their ingestion system (section D) 5 8.5 £4,250
Ingest free-text accounts, such as a seafarer’s own description (section D) 4 6.8 £3,400
Ingest investigation reports submitted to IMO GISIS, filling gaps for countries without their own site (section D) 4 6.8 £3,400
Link short-form retellings to the accident report they came from (section D) 3 5.1 £2,550
GoldSilver plus every Could 248 421.6 £210,800 £412,825
Causal AI (section E) 119 202.3 £101,150
Maritime Narrative Analytics (section B(a)) RAi funded 53 90.1 £45,050
Chat Coulds (section A) 34 57.8 £28,900
Estimate accident costs, refined by factors like location and date (section D) 25 42.5 £21,250
Add historical weather and sea state at each accident’s time and place, from Copernicus (section D) 5 8.5 £4,250
Estimate accident costs where reports don’t publish one, simple version (section D) 5 8.5 £4,250
Ingest P&I club loss prevention publications (section D) 4 6.8 £3,400
Ingest scanned pre-2005 reports (section D) 3 5.1 £2,550

A. Chat (P7)

Priority Feature Build days Total days Cost
Must Learning from user questions: features extracted from each question and reused when others ask similar ones 12 20.4 £10,200
Must Ask a question and get an answer drawn from verified reports (the tools the model can call, plus retrieval) (Scope Doc) 11 18.7 £9,350
Must Quality test set, re-run after every change 7 11.9 £5,950
Must Every claim cited and checked, with a link to the report at the right PDF page (Scope Doc) 4 6.8 £3,400
Must Replies stream in; chat screen works on mobile and is accessible 4 6.8 £3,400
Must Answers in plain maritime English, tuned for task briefings and descriptions of a near miss (Scope Doc: plain English, briefings, near miss) 4 6.8 £3,400
Must Refuses when there is no evidence or the question is off-topic; prompt-injection defence; safety disclaimer 3 5.1 £2,550
Must Follow-up questions within a conversation 2 3.4 £1,700
Must Tracing and cost per conversation 2 3.4 £1,700
Must total 49 83.3 £41,650
Should Shared memory across all chats 6 10.2 £5,100
Should Saved conversation history (needs sign-in) 3 5.1 £2,550
Should Toolbox talk format (Scope Doc) 3 5.1 £2,550
Should Chat searches past conversations to keep up with the user 3 5.1 £2,550
Should Tables in the reply 2 3.4 £1,700
Should Suggested follow-up and sample questions 1 1.7 £850
Should Long conversations: older turns summarised 1 1.7 £850
Should Copy, print and export an answer 1 1.7 £850
Should total 20 34 £17,000
Could Charts in the reply 10 17 £8,500
Could Upload your own document and discuss it, deleted at session end, pattern record kept (Scope Doc) 10 17 £8,500
Could Maps and heat maps 8 13.6 £6,800
Could Remembers your role and vessel type across chats 3 5.1 £2,550
Could Share link to a conversation 2 3.4 £1,700
Could Regenerate an answer, or edit and resend a question 1 1.7 £850
Could total 34 57.8 £28,900
Won't Web search (breaks the verified-data-only rule) n/a n/a
Won't Voice, image generation, native app n/a n/a

B. Analyst tools and dashboards

Priority Feature Build days Total days Cost
Must Pattern finding for CHIRP analysts: ask what keeps recurring across all reports (for example, what keeps causing enclosed-space deaths), see the supporting cases, and get a first-draft outline of a CHIRP write-up to edit (Scope Doc) 8 13.6 £6,800
Must Proposed coding and similar cases for incoming CHIRP reports, inside the review tool (Scope Doc) 6 10.2 £5,100
Should Dashboards filtered by vessel type, date, outcome and country (P6) (Scope Doc) 12 20.4 £10,200

B(a). Maritime Narrative Analytics

Responsible AI Institute funding

Breaks each report into a narrative structure (orientation, complicating action, harm, resolution), profiles of the vessels, people, places and equipment in it, and a chain of events, all published as an open ontology. Adapts Steve’s Hostile Narrative Analysis (Anning, 2023), and every extracted fact traces back to its sentence and the grammar rule that produced it. Feeds pattern finding and proposed coding (section B), and overlaps with event sequences in section E.

Priority Feature Build days Total days Cost
CouldMust if taken Semantic relation mapping: the grammar of each sentence turned into who did what to what, where and when, as linked data 12 20.4 £10,200
CouldMust if taken Event analysis: events in each report classified (navigation, mechanical failure, human error, environmental) and linked into a causal chain 10 17 £8,500
CouldMust if taken Entity profiling: vessels, people, places and equipment extracted and matched to IMO or Wikidata records 9 15.3 £7,650
CouldMust if taken Maritime Narrative Ontology v1.0: schema, entities and events brought together and published 7 11.9 £5,950
CouldMust if taken Narrative schema: the parts of an incident narrative, checked against MAIB and IMO reports 6 10.2 £5,100
CouldMust if taken Narrative structure and events stored against each report, and used for similar cases and pattern finding (section B) 5 8.5 £4,250
CouldMust if taken Validation against held-out reports, with precision and coverage figures 4 6.8 £3,400
Could total 53 90.1 £45,050

C. Public launch platform (rest of P5)

Priority Item Build days Total days Cost
Must Closed beta with real seafarers, then a round of fixes 6 10.2 £5,100
Must Landing page, first-time walkthrough, sample questions, accessibility (WCAG) 5 8.5 £4,250
Must Rate limits, a spending cap that degrades gracefully, abuse and scraping protection (Scope Doc) 4 6.8 £3,400
Must Watching answers, a “report a wrong answer” route, withdrawing an answer or record (Scope Doc) 4 6.8 £3,400
Must Deployment, load testing, backups, cost per conversation 4 6.8 £3,400
Must User and developer documentation 3 5.1 £2,550
Must Public access kept separate from staff tools; anonymous sessions (Scope Doc) 2 3.4 £1,700
Must Terms, privacy notice and disclaimers built in (CHIRP writes the wording) 1.5 2.55 £1,275
Total 29.5 50.15 £25,075

D. Sources

Priority Source Build days Total days Cost
Must Official investigation bodies (section F) contingency, no PM 5 each, 2.5 fast 7 £3,500 each, £1,750 fast
Must Capture more metadata, with a shared vocabulary for each field across jurisdictions 14 23.8 £11,900
Must MAIB webpage ingestion, alongside the report or on its own where there is no report 4 6.8 £3,400
Must Automatically ingest new content from every source as it is published 4 6.8 £3,400
Must CHIRP Maritime published reports 3 5.1 £2,550
Must MAIB Safety Digests 2 3.4 £1,700
Must Accident linking, official reports: a report on an accident we already hold is kept and linked to that accident, never rejected (several countries’ reports on one accident, digests to the report they summarise; duplicate PDFs are already caught) (Scope Doc) 2 3.4 £1,700
Should Legislation: UK/MCA, MLC, IMO references, AI-proposed links, recurring findings listed by law (Scope Doc) 12 20.4 £10,200
Should Short-form schema design, implementation and ingestion system: short-form sources may not store like-for-like with a document report and may need their own category 5 8.5 £4,250
Should Generic free-text ingestion 4 6.8 £3,400
Should IMO GISIS casualty module 4 6.8 £3,400
Should Accident linking, derivative sources: MARS, SAFETY4SEA, GISIS and IMO lessons matched to the accident they retell, on looser clues (Scope Doc) 3 5.1 £2,550
Should IMO Lessons Learned 2.5 4.25 £2,125
Should Nautical Institute MARS 2.5 4.25 £2,125
Should SAFETY4SEA 2.5 4.25 £2,125
Should IMCA safety flashes 2.5 4.25 £2,125
Should USCG Marine Safety Alerts 2.5 4.25 £2,125
Should Marine Safety Forum alerts 2.5 4.25 £2,125
Could Refined, fairly accurate cost estimate based on factors like location and date 25 42.5 £21,250
Could Copernicus metocean data 5 8.5 £4,250
Could Simple cost estimate for reports that don’t publish one 5 8.5 £4,250
Could P&I club loss prevention publications 4 6.8 £3,400
Could Pre-2005 scanned reports (the pipeline already reads page images) 3 5.1 £2,550
Won't Paris and Tokyo MoU detentions n/a n/a
Won't Other news and trade press n/a n/a
Won't Non-English reports via translation n/a n/a

E. Causal AI

Moves the corpus from factors that appear together to factors that cause outcomes, using hypotheses from CHIRP’s experts, analysts and the data. From Steve’s user journeys, not the Scope Doc. We have not built causal models before, so the spike comes first and confirms or revises the estimates below it.

Priority Feature Build days Total days Cost
CouldMust if taken Test a hypothesis: effect size, robustness checks and a plain verdict 10 17 £8,500
CouldMust if taken Analysis-ready dataset built from verified records and the MAIB spreadsheets 6 10.2 £5,100
CouldMust if taken Causes stated in each report extracted and cited 6 10.2 £5,100
CouldMust if taken SHIELD codes proposed across the corpus and confirmed in the review tool (shares a model with proposed coding, section B) 6 10.2 £5,100
CouldMust if taken Cause-and-effect maps for priority accident types, drawn up with CHIRP’s experts and fed into the system 5 8.5 £4,250
CouldMust if taken Chat checks causal claims before replying, and words untested ones as associations (section A) 5 8.5 £4,250
CouldMust if taken Analysts raise and test hypotheses through the chat (section A) 5 8.5 £4,250
CouldMust if taken Hypothesis register in the review tool, signed off before results are cited 4 6.8 £3,400
CouldMust if taken Feasibility spike: test the approach on our data before committing the rest 3 5.1 £2,550
CouldMust if taken Quality test set of known cause-and-effect relationships, re-run after every change 3 5.1 £2,550
CouldShould if taken Exposure data (fleet sizes, time at sea), so findings are rates rather than counts 6 10.2 £5,100
CouldShould if taken Did the rule work: outcomes before and after a regulation change, and between countries (section D) 6 10.2 £5,100
CouldShould if taken Causal pathway view: the maps on screen, with the evidence behind every link 6 10.2 £5,100
CouldShould if taken Where a factor matters most: effects broken down by vessel type, region and conditions 5 8.5 £4,250
CouldShould if taken Control measures ranked by estimated lives saved 5 8.5 £4,250
CouldShould if taken AI-proposed links for experts to accept or reject into the maps 5 8.5 £4,250
CouldShould if taken Chat suggests causal lines of enquiry and offers to test them (section A) 4 6.8 £3,400
CouldShould if taken Evidence tables for regulators, exported as PDF and CSV 2 3.4 £1,700
CouldCould if taken Sequence of events extracted from each report, to find recurring chains 8 13.6 £6,800
CouldCould if taken Did the recommendation work: outcomes before and after implementation 5 8.5 £4,250
CouldCould if taken Weather and sea state as tested factors (Copernicus, section D) 3 5.1 £2,550
CouldCould if taken Incoming CHIRP reports set against the maps, with follow-up questions for the reporter (section B) 3 5.1 £2,550
CouldCould if taken Hypotheses re-tested as new reports are verified 3 5.1 £2,550
CouldCould if taken Methods note for regulators and external reviewers 3 5.1 £2,550
CouldCould if taken Dashboard figures labelled as association or tested cause (section B) 2 3.4 £1,700
Could total 119 202.3 £101,150
Won't Causal findings shown to the public without human sign-off n/a n/a
Won't Claims about how often accidents happen without exposure data n/a n/a
Won't Real-time risk scoring for vessels or voyages n/a n/a

F. Official investigation bodies

Each build day comes with 0.2 days of contingency and 0.2 days of ZigZag product management, with no project management: 1.4 days in total, at £500 a day. Full ingestion (5 days) encodes each jurisdiction’s rules. Fast (2.5 days) is quicker with less reliable data: halfway between full ingestion and ingesting the reports with no jurisdiction rules encoded. The Baseline bundle includes any 10 of these at full ingestion.

Country Body Status Full: 5 build days, 7 in total Fast: 2.5 build days, 3.5 in total
United Kingdom MAIB Done (Tranche 1)
United States NTSB Done (Tranche 1)
Australia ATSB To survey £3,500 £1,750
Bahamas Bahamas Maritime Authority To survey £3,500 £1,750
Belgium FPS Mobility, Shipping To survey, check English £3,500 £1,750
Canada TSB To survey £3,500 £1,750
Cyprus MAIC To survey £3,500 £1,750
Denmark DMAIB To survey £3,500 £1,750
Estonia Safety Investigation Bureau To survey £3,500 £1,750
Finland Safety Investigation Authority To survey £3,500 £1,750
France BEAmer To survey, check English £3,500 £1,750
Germany BSU To survey, check English £3,500 £1,750
Greece HBMCI To survey £3,500 £1,750
Hong Kong MAIB HK To survey, summaries only £3,500 £1,750
Isle of Man Ship Registry To survey, annual casualty reports £3,500 £1,750
Japan JTSB To survey, check English £3,500 £1,750
Malta MSIU To survey £3,500 £1,750
Marshall Islands IRI To survey £3,500 £1,750
Netherlands Dutch Safety Board To survey, check English £3,500 £1,750
New Zealand TAIC To survey £3,500 £1,750
Norway NSIA To survey £3,500 £1,750
Singapore Ministry of Transport To survey £3,500 £1,750
South Africa SAMSA To survey £3,500 £1,750
Sweden SHK To survey £3,500 £1,750
Turkey TSIC To survey, check English £3,500 £1,750
Latvia TAIIB Won't (published in Latvian) n/a n/a
Total, 23 bodies £80,500115 days £40,25057.5 days

Calibration from Tranche 1

Phase Estimate (days) Actual (days) Variance
MVP 19.5 19.34 on estimate
P1 Foundation 35 50.5 +44%
P1.5 Foundation extension 3 4.1 +37%
US (NTSB) jurisdiction, first non-UK, generalised the pipeline n/a ~12 working days sets the per-country seam
NTSB corpus AI processing, 587 reports, including narrative-entity extraction n/a ~$40-50 AI processing cost is negligible

Appendix: Scope Doc

Direct quotes from the Insight Engine Tranche 2 Scope Doc, next to the menu item each one supports. Priorities are the Scope Doc’s.

Chat

Menu item Priority Scope Doc
Answers drawn from verified reports, tuned for task briefings Must As a seafarer or supervisor about to start a task, I want to know what has gone wrong during similar work, so that I recognise the critical hazards before I start. … A member of the public asks a task question and receives a briefing in which every claim is cited to a verified report.
Tuned for descriptions of a near miss Must As a seafarer who has just had a near miss, I want to describe what happened, so that I can see whether others have experienced the same thing and what they learned. … A typed description returns related verified incidents, their shared factors and a link to CHIRP’s reporting form.
Every claim cited and checked Must Every claim is cited to source. … Every public answer cites verified records and links to the source report.
Plain maritime English Fixed constraint Outputs are in English only, written in plain maritime English so that users can rely on browser translation.
Toolbox talk format Should As an onboard supervisor, I want the briefing produced as a toolbox talk, so that I can brief the crew from real incidents.
Upload your own document and discuss it Could As an onboard supervisor, I want to upload a draft risk assessment, permit or plan, so that it is challenged against previous incidents before authorisation. … As an organisation or onboard investigator, I want to upload our own report or event narrative, so that I can find comparable cases and lines of enquiry, knowing the document itself is not kept.

Analyst tools and dashboards

Menu item Priority Scope Doc
Proposed coding and similar cases for incoming CHIRP reports Must As a CHIRP analyst, I want each report we receive coded and set against similar cases, so that I can assess it quickly and consistently and respond with context. … A submitted report returns proposed coding and similar cases, and nothing enters the database until verified in the review tool.
Pattern finding for CHIRP analysts Must As a CHIRP analyst, I want to ask questions across all reports and datasets, so that I can unearth patterns and publish insight that raises awareness and encourages improvement. … An analyst query returns pathways with supporting cases, and drafting support produces a reviewable outline.
Dashboards Should As a shore-side manager or commercial operator, I want to explore trends across the evidence base, so that I can find vulnerabilities and decide where to invest. … Dashboards filter by vessel type, date, outcome and country, and every figure traces to verified records.

Public launch platform

Menu item Priority Scope Doc
Public access kept separate from staff tools Must Logins and role-based access for CHIRP staff and partners; anonymous public access kept separate from the review tool and drafting features.
Rate limits, spending cap, abuse protection Must Rate limiting per user, a spending cap on question-answering costs, and protection against abuse, scraping and prompt injection.
Watching answers, withdrawing an answer or record Must Usage and error monitoring with alerts, and a route to correct or withdraw a published answer or record.

Sources

Menu item Priority Scope Doc
Accident linking, official reports and derivative sources Must Many sources republish or edit the same event (MARS, SAFETY4SEA, IMO GISIS and national bodies). Each derivative account links to its source report, and pattern analysis counts the event once.
Legislation Should As CHIRP, I want to see where incidents recur because regulation is absent, unclear or poorly implemented, so that I can put evidence to flag states and encourage change. … Findings carry AI-proposed legislation links, clearly caveated for the user to verify, and an analyst can list recurring findings by instrument.

Appendix: Steel (£50k)

A £50k option that sits below Bronze. It combines chat first and data first: a public chat that answers from verified reports, a minimal platform to run it safely in public, and six more investigation bodies at fast ingestion.

What Steel trades away
  • No landing page or walkthrough. Users reach the chat through a link from CHIRP’s site.
  • The six bodies use fast ingestion, so their data is less reliable than full ingestion.
  • A slim quality test set (core questions only), and no closed beta before release.
  • No analyst tools, no shared vocabulary and no learning from user questions.

Included

Item Build days Total days Cost
SteelChat first plus data first 55.5 89.85 £44,925
Chat (section A)
Ask a question and get an answer drawn from verified reports (the tools the model can call, plus retrieval)1118.7£9,350
Every claim cited and checked, with a link to the report at the right PDF page46.8£3,400
Answers in plain maritime English, tuned for task briefings and descriptions of a near miss46.8£3,400
Replies stream in; chat screen works on mobile and is accessible46.8£3,400
Refuses when there is no evidence or the question is off-topic; prompt-injection defence; safety disclaimer35.1£2,550
Follow-up questions within a conversation23.4£1,700
Quality test set, slimmed to core questions23.4£1,700
Platform (section C)
Rate limits, a spending cap that degrades gracefully, abuse and scraping protection46.8£3,400
Deployment and backups, without load testing35.1£2,550
Public access kept separate from staff tools; anonymous sessions23.4£1,700
Terms, privacy notice and disclaimers built in (CHIRP writes the wording)1.52.55£1,275
Data (section F)
Ingest reports from 6 more national investigation bodies, fast ingestion at 2.5 days each contingency, no PM1521£10,500
Product management
Maritime Safety Senior Stakeholder Workshop, to decide the future direction of CHIRPdb in ZigZag product management

Where the days go

Work Days Cost
Build55.5£27,750
Contingency11.1£5,550
Project management (not on investigation bodies)12.15£6,075
ZigZag product management11.1£5,550
Total89.85£44,925

Stretch goals

Taken on if time allows within the build, or funded on top. Each is priced on its own; all of them take Steel to £56,825. Without the shared vocabulary, pattern finding compares each jurisdiction in its own terms.

Item Build days Total days Cost
Pattern finding for CHIRP analysts: ask what keeps recurring across all reports, counted across every report rather than read from the top matches, see the supporting cases, and get a first-draft outline of a CHIRP write-up (section B)813.6£6,800
SHIELD code matching: SHIELD codes proposed across the corpus and confirmed in the review tool, so reports can be counted by causal factor (section E)610.2£5,100
Stretch goals total1423.8£11,900

Notable exclusions

  • Learning from user questions. The engine does not get better as people use it.
  • Analyst tools for CHIRP. No proposed coding for incoming reports, and pattern finding only as the stretch goal. Without it, questions about what keeps recurring are refused rather than answered from the top matches.
  • A shared vocabulary across jurisdictions. Each body’s reports keep their own terms for accident type, vessel type and so on.
  • Reporting and withdrawing a wrong answer. No route for users to flag an answer, or for CHIRP to withdraw one.
  • CHIRP Maritime’s own reports and MAIB Safety Digests. The chat answers from investigation reports only.
  • Automatic ingestion of new reports. New reports are added by hand.
  • Full ingestion of investigation bodies. Six bodies at fast ingestion, rather than ten at full.
  • A closed beta and a landing page. No round of testing with real seafarers, and no walkthrough for first-time users.
  • SHIELD code matching and Causal AI. SHIELD code matching only as a stretch goal, and no cause-and-effect analysis.
  • Toolbox talks and dashboards. No printable toolbox talk format, and no dashboards to explore trends.

Excluded

Steel plus the Bronze Musts below adds up to Bronze (£138,275). Every Should and Could stays out too.

Item Build days Total days Cost
Bronze Musts left outThe gap from Steel to Bronze 116 186.7 £93,350
Chat (section A)
Learning from user questions: features extracted from each question and reused when others ask similar ones1220.4£10,200
Rest of the quality test set58.5£4,250
Tracing and cost per conversation23.4£1,700
Platform (section C)
Closed beta with real seafarers, then a round of fixes610.2£5,100
Landing page, first-time walkthrough, sample questions, accessibility (WCAG)58.5£4,250
Watching answers, a “report a wrong answer” route, withdrawing an answer or record46.8£3,400
User and developer documentation35.1£2,550
Load testing11.7£850
Analyst tools (section B)
Pattern finding for CHIRP analysts, with supporting cases and a first-draft outline813.6£6,800
Proposed coding and similar cases for incoming CHIRP reports610.2£5,100
Sources (section D)
Capture more metadata, with a shared vocabulary for each field across jurisdictions1423.8£11,900
MAIB webpage ingestion, alongside the report or on its own46.8£3,400
Automatically ingest new content from every source as it is published46.8£3,400
CHIRP Maritime published reports35.1£2,550
MAIB Safety Digests23.4£1,700
Accident linking, official reports23.4£1,700
Investigation bodies (section F)
4 more bodies to reach Baseline’s 10, at full ingestion contingency, no PM2028£14,000
Upgrade Steel’s 6 bodies from fast to full ingestion contingency, no PM1521£10,500
Every ShouldAs in Silver 75 127.5 £63,750
Chat Shoulds: shared memory, saved history, toolbox talks, searching past chats, tables, suggested questions, summarising, export (section A)2034£17,000
Dashboards filtered by vessel type, date, outcome and country (section B)1220.4£10,200
Source Shoulds: legislation, short-form schema, free-text, IMO GISIS, derivative accident linking, six short-form sources (section D)4373.1£36,550
Every CouldAs in Gold 76 129.2 £64,600
Chat Coulds: charts, document upload, maps, remembering role and vessel, share links, regenerate (section A)3457.8£28,900
Source Coulds: refined and simple cost estimates, Copernicus, P&I clubs, pre-2005 scans (section D)4271.4£35,700
Causal AIAll of section E 119 202.3 £101,150
Maritime Narrative AnalyticsFunded separately by the Responsible AI Institute (section B(a)) 53 90.1 £45,050
Total excluded 439 735.8 £367,900

Appendix: Copper (£50k)

An alternative to Steel. Instead of building our own chat, we build a CHIRPdb MCP server: a connector that users add to Claude. Claude provides the chat screen, memory, history, charts, uploads and mobile apps; our server provides the verified reports, search and counting across every report. Most of section A comes with Claude, so the budget goes into the backend.

What Copper trades away
  • Users need a Claude account and must add the connector. Which Claude plans and apps support custom connectors needs confirming before we commit.
  • Claude writes the answers, not us. Our tools return only verified records with citations, and our instructions tell Claude to cite every claim and refuse off-topic questions, but we cannot stop it drawing on its own knowledge or web search.
  • Documents users upload go to Anthropic under the user’s Claude terms, not deleted by CHIRP at session end.
  • The experience is Claude’s, not CHIRP’s branded product.
  • Question-answering costs move to the user’s Claude plan. CHIRP pays only for hosting and search.

Included

Item Build days Total days Cost
CopperCHIRPdb connector for Claude, plus backend 59.5 96.65 £48,325
MCP server
Tools Claude can call: search verified reports, find incidents similar to a description, count and compare across every report by causal factor, vessel type, date, outcome and country (pattern finding), and fetch a report with a link to the right PDF page813.6£6,800
Quality test set, run through Claude and re-run after every change35.1£2,550
Rate limits and abuse and scraping protection on the public server35.1£2,550
Deployment, backups and monitoring35.1£2,550
Instructions and prompts: plain maritime English, cite every claim, refuse off-topic questions, task briefings, near misses, toolbox talks and draft CHIRP write-ups23.4£1,700
Logging and tracing of every question asked, the starting point for learning from user questions23.4£1,700
Setup page: how to add the CHIRPdb connector to Claude, with sample questions23.4£1,700
Terms, privacy notice and disclaimers (CHIRP writes the wording)1.52.55£1,275
Backend (section B, section D)
Capture more metadata, with a shared vocabulary for each field across jurisdictions, so counts across countries compare like with like1423.8£11,900
Proposed coding and similar cases for incoming CHIRP reports, inside the review tool610.2£5,100
Data (section F)
Ingest reports from 6 more national investigation bodies, fast ingestion at 2.5 days each contingency, no PM1521£10,500
Product management
Maritime Safety Senior Stakeholder Workshop, to decide the future direction of CHIRPdb in ZigZag product management

Where the days go

Work Days Cost
Build59.5£29,750
Contingency11.9£5,950
Project management (not on investigation bodies)13.35£6,675
ZigZag product management11.9£5,950
Total96.65£48,325

Comes with Claude

Chat features from section A that we would not build: 56 build days, £47,600 if we built them ourselves.

Streaming, mobile and accessible chatFollow-up questionsMemory across chatsSaved historySearching past chatsTablesLong conversations summarisedCopy, print and exportChartsDocument uploadMaps and heat mapsRemembers role and vessel typeShare linksRegenerate and edit

Stretch goals

Taken on if time allows within the build, or funded on top. It takes Copper to £53,425. Copper’s counting by causal factor depends on SHIELD code matching.

Item Build days Total days Cost
SHIELD code matching: SHIELD codes proposed across the corpus and confirmed in the review tool, so reports can be counted by causal factor (section E)610.2£5,100

Notable exclusions

  • Learning from user questions. Questions are logged, but not yet reused to improve answers.
  • Reporting and withdrawing a wrong answer. Claude’s answers are not stored by us, so a wrong record can be withdrawn but an answer cannot.
  • CHIRP Maritime’s own reports and MAIB Safety Digests. Claude answers from investigation reports only.
  • Automatic ingestion of new reports. New reports are added by hand.
  • Full ingestion of investigation bodies. Six bodies at fast ingestion, rather than ten at full.
  • A closed beta. No round of testing with real seafarers.
  • Causal AI, legislation and dashboards. SHIELD code matching is a stretch goal.

Appendix: Iron (£50k)

Steel’s chat, for CHIRP staff only. There is no public launch, so the money that Steel spends on public access goes into a chat that learns: from the questions staff ask, and from the wrong answers they report. It is still hardened, tested and deployed to production, and it lives inside the existing review tool, behind its sign-in and roles.

What Iron trades away
  • No one outside CHIRP can use it. A public launch later needs Steel’s public platform items at least: rate limits and a spending cap, public access kept separate from staff tools, and terms (£6,375).
  • Two more investigation bodies at fast ingestion, rather than Steel’s six.
  • Replies arrive whole rather than streaming in, and the chat is built for desktop, not mobile.

Included

Item Build days Total days Cost
IronA chat that learns, for CHIRP staff only 59 98.8 £49,400
Chat (section A)
Ask a question and get an answer drawn from verified reports (the tools the model can call, plus retrieval)1118.7£9,350
Every claim cited and checked, with a link to the report at the right PDF page46.8£3,400
Answers in plain maritime English, tuned for task briefings and descriptions of a near miss46.8£3,400
Refuses when there is no evidence or the question is off-topic; prompt-injection defence; safety disclaimer35.1£2,550
Follow-up questions within a conversation23.4£1,700
Learning (section A, section C)
Learning from user questions: features extracted from each question and reused when others ask similar ones1220.4£10,200
Watching answers, a “report a wrong answer” route, withdrawing an answer or record; reported answers join the quality test set46.8£3,400
Hardening and testing (section A, section C)
Quality test set, re-run after every change711.9£5,950
Deployment, load testing, backups46.8£3,400
Tracing and cost per conversation23.4£1,700
Chat inside the existing review tool, behind CHIRP staff sign-in and roles11.7£850
Data (section F)
Ingest reports from 2 more national investigation bodies, fast ingestion at 2.5 days each contingency, no PM57£3,500
Product management
Maritime Safety Senior Stakeholder Workshop, to decide the future direction of CHIRPdb in ZigZag product management

Where the days go

Work Days Cost
Build59£29,500
Contingency11.8£5,900
Project management (not on investigation bodies)16.2£8,100
ZigZag product management11.8£5,900
Total98.8£49,400

Stretch goals

Taken on if time allows within the build, or funded on top. Each is priced on its own; all of them take Iron to £61,300. Without the shared vocabulary, pattern finding compares each jurisdiction in its own terms.

Item Build days Total days Cost
Pattern finding for CHIRP analysts: ask what keeps recurring across all reports, counted across every report rather than read from the top matches, see the supporting cases, and get a first-draft outline of a CHIRP write-up (section B)813.6£6,800
SHIELD code matching: SHIELD codes proposed across the corpus and confirmed in the review tool, so reports can be counted by causal factor (section E)610.2£5,100
Stretch goals total1423.8£11,900

Notable exclusions

  • A public launch. No public access, landing page, terms, public rate limits or closed beta with seafarers.
  • Analyst tools for CHIRP. No proposed coding for incoming reports, and pattern finding only as the stretch goal. Without it, questions about what keeps recurring are refused rather than answered from the top matches.
  • A shared vocabulary across jurisdictions. Each body’s reports keep their own terms for accident type, vessel type and so on.
  • Memory across chats and saved conversation history. The chat learns from questions across all users, but does not remember an individual user.
  • CHIRP Maritime’s own reports and MAIB Safety Digests. The chat answers from investigation reports only.
  • Automatic ingestion of new reports. New reports are added by hand.
  • Full ingestion of investigation bodies. Two bodies at fast ingestion, rather than ten at full.
  • SHIELD code matching and Causal AI. SHIELD code matching only as a stretch goal, and no cause-and-effect analysis.