The Open Record Architecture

01The brief

The Federal Court Entity can take the opportunity to design, in the wake of an emergency transition, the long-term system that produces the official record for the Federal Court of Australia, both divisions of the FCFCOA, and the National Native Title Tribunal. The courts have already set out the elements that matter. The interim arrangements use a panel of onshore transcript service providers and a panel of transcriptionists managed in-house. Under those arrangements the courts have access to and control of all data and recordings. The stated focus for the longer term is a low-cost model for all litigants, including those most disadvantaged or at risk.¹

Recording and transcription in these courts was outsourced to a private contractor for several decades.² In 2026 that model failed three times over, and each failure teaches a separate design lesson.

1. Commercial failure. The Australian subsidiary of a foreign-owned supplier entered voluntary administration in March 2026 and was wound down. The courts' contract ceased on 30 June 2026.³ Record production for the federal courts was put at risk, and reporting indicates courts in New South Wales, Queensland and Victoria were also affected, with an impact of around 600 staff and transcripts for around 200 courtrooms.⁴
2. Sovereignty failure. The supplier disclosed a data incident. Media reported that thousands of court files were accessed without authorisation by a contractor in India.⁵ A procurement risk became a justice-system integrity incident.
3. Structural failure. After decades of outsourcing, the courts owned neither the platform, the workflow, the pricing, the formats, nor the workforce relationships. When the supplier fell, the courts had to rebuild that institutional capability under emergency conditions.

A design that fixes only the first failure, by picking a healthier vendor or three of them, repeats the third. The design below aims higher: no future supplier failure, whether commercial, security or performance, should be capable of threatening the record itself. At the same time it has to deliver the courts' access-to-justice mandate and stay current through two decades of change in AI.

02Five layers, sorted by rate of change

Court transcription systems become obsolete or captured for a structural reason. Procurement bundles together things that change at very different speeds, then hands the bundle to one company. The audio of a hearing must be valid in fifty years. The best speech-recognition model changes every eighteen months. If the two are tied together, the courts must either freeze the technology or perpetually re-tender the record. This design separates the system into five layers by rate of change and assigns ownership accordingly.

Layer 0. The record itself (changes: never). The audio, the certified transcripts, and the registers of what happened to them. Owned by the courts absolutely, held in open published formats, forever. No supplier holds ownership of this layer under any future model.

Layer 1. The data spine (changes: over decades). The schemas, identifiers, access-control model and the published record format. Owned by the courts and published as open standards, so any conforming supplier or tool can participate. Vendor lock-in lives in formats and schemas, and this is the layer where it is removed.

Layer 2. The platform (changes: five to ten years). Ordering portal, workflow, allocation, secure production workspace, delivery, billing and registers. Court-owned, or licensed with full source escrow and perpetual rights. Built under contracts that assign intellectual property. Replaceable component by component, because it speaks only the Layer 1 standards.

Layer 3. The engine bay (changes: about eighteen months). Speech recognition, alignment, flagging, redaction assistance and checker models, treated as swappable modules behind a fixed interface, promoted and retired by a permanent evaluation harness. No engine is ever load-bearing for the validity of the record, and every engine is one evaluation cycle from replacement. The architecture avoids betting on any model by institutionalising the process of always using the current best one, on sovereign infrastructure.

Layer 4. Human capacity (elastic). A national network of accredited transcribers, editors, captioners and realtime stenographers, both firms and individuals, working at regulated rates through the platform, together with a permanent in-house core. Capacity scales with demand. The failure of any one firm costs the system nothing.

One procurement rule does most of the work: no single supplier may span more than one layer. The company that supplies capture hardware does not run the platform. The platform integrator does not supply engines. Engine suppliers do not employ the transcribers. The layer rule removes concentration risk structurally, before any contract clause is drafted.

03The record as structured data

The most consequential design decision, and the one that future-proofs everything downstream, is this: the authoritative record is a structured, signed data object.

The courts should define and publish an open Court Record Format. For each hearing session, one canonical object contains the speaker-turn sequence with word-level timestamps into the source audio; speaker attribution bound to the courtroom channel map; provenance for each segment (which engine drafted it, which human produced it, which human certified it, what the checker queried); confidence scores at draft stage; the assurance tier of the artefact (section 5); the applicable access-control policy; and a cryptographic signature chain from capture to certification.

Everything a court, litigant or citizen receives is then a rendering of that object: the traditional court-style PDF; an accessible HTML transcript meeting WCAG 2.2 AA that works with screen readers as delivered; caption files for streamed hearings; braille-ready and large-print exports; appeal-book pagination; and a machine-readable feed for the digital court file. Accessibility becomes a property of the format itself. It is produced once and rendered many ways at near-zero marginal cost, which is how the courts' obligations to court users with disability⁶ can be met structurally.

Corrections become controlled data operations. A correction is a new signed entry in an append-only Transcript Register. It supersedes a segment, carries full provenance, and requires dual-key authorisation: the person proposing a post-certification change can never be the person approving it. The history of the record is part of the record. Fifty years from now, anyone with the published specification can read every version and verify every signature, regardless of which vendors existed along the way.

The specification should be published openly, versioned conservatively, and governed by the courts. State jurisdictions can adopt it at will. Courts in New South Wales, Queensland and Victoria were reportedly served by the same collapsed supplier,⁷ so the appetite is likely to exist. Over time this gives Australia a shared national record standard, portable accreditation, and pooled engine evaluation.

04Capture

Everything downstream inherits the quality of capture, so this layer allows the least compromise.

Court-owned recording in every courtroom, on commodity hardware and open formats. Broadcast WAV with embedded timing metadata, plus a sidecar manifest carrying the case identifier from the case-management system, the courtroom, the date, and the channel map. No proprietary container formats. Recording systems that require a vendor’s own player to open a court recording are one of the dependencies being eliminated.

Per-channel capture as the default. Each microphone (bench, witness, each counsel position, interpreter) is a discrete recorded channel. This single engineering choice does more for transcription accuracy and speaker attribution than any AI advance. Attribution becomes a fact of the channel map, and diarisation is demoted to confirming who spoke on a shared channel. Interpreted proceedings capture the source language and the interpretation as separate labelled channels, so the record can show both what the witness said and what the court heard.

Integrity at the point of creation. Rolling cryptographic hashes during recording, and a signed session manifest at close. Synthetic audio is now cheap, and the authenticity of the court’s audio will be challenged. A signed hash chain answers that challenge conclusively, and chain of custody becomes automatic and machine-verifiable.

Hybrid hearings, same standard. Remote and hybrid proceedings are captured by platform-side integration into the identical pipeline: same formats, same manifests, same integrity wrapper, and per-participant channels where the conferencing platform allows. A hearing’s record should not vary in quality with its geography.

Resilience and monitoring. Local buffering with store-and-forward, so recording never depends on the network. Live channel-health monitoring visible to the courtroom officer, with automatic alerts on failure. A lost recording is a potential retrial, and the capture layer should be monitored accordingly. Mediation and conciliation events are confidential and remain unrecorded, as now.⁸

Native title and cultural protocols. On-country hearings receive portable capture kits meeting the same standard. The access-control model (section 7) natively supports culturally restricted evidence at segment level, including gender-restricted and group-restricted material. The Court’s Native Title Practice Note already contemplates orders restricting who may hear particular evidence or access the transcripts, and gender-restricted evidence has been managed by suppression orders with the transcript and audio separated from the general record and marked restricted.⁹ In this design, that restriction travels as metadata the system enforces.

05Assurance tiers

The debate about whether AI should transcribe the courts becomes tractable once two things are accepted. A court needs several different text products with different integrity requirements. And the real danger throughout is mislabelling. This design therefore makes assurance level an explicit, machine-enforced property of every text artefact.

Tier L (Live). Streaming recognition with seconds of latency, rendered to bench displays, party devices and caption surfaces. Visibly watermarked as an uncertified live draft on every screen and page. It serves as the judge’s and parties' working text during the hearing, the accessibility captioning baseline in any courtroom on request, and the caption source for publicly streamed matters. India’s Supreme Court has run live AI transcription in Constitution Bench hearings since early 2023, with transcripts vetted by counsel before publication,¹⁰ so the concept is proven at the top of a major common-law judiciary.

Tier D (Draft). A complete machine-produced transcript of a session, available within hours, watermarked and metadata-flagged as uncertified. Its marginal cost is close to nothing, so it can be priced at close to nothing, including free to parties for their own proceedings. This tier is the direct answer to the courts' stated mandate. A self-represented litigant in a family or migration matter, who today may simply go without, gets same-day access to what was said in their own case. The UK is moving in this direction by policy, with free transcripts of judges' sentencing remarks for Crown Court victims, on request, from spring 2027.¹¹ This design gets there structurally, for everyone, for every hearing.

Tier C (Certified). The official record. Produced by an accredited human transcriber working from the Tier D draft in a secure workspace, verified under the quality regime in section 8, certified over the named transcriber’s accreditation, committed to the append-only register, and rendered into every delivery format. It carries a regulated fee and tiered turnarounds from standard through same-day and progressive same-day, tiers the courts already offer.¹² Only Tier C is filable, citable in judgments, or usable on appeal. That rule is enforced both by practice note and by the e-filing system refusing lower-tier artefacts.

Tier C+ (Certified Realtime). Machine-shorthand realtime with a certified daily transcript, for the matters that justify it: long commercial trials, major appeals, high-profile inquiries. A party-funded premium tier that preserves and funds a craft capability no speech engine replaces at the top end.

The certified record keeps a human author. Everything else is labelled for what it is.

Two properties make the tier system work. First, the upgrade path. A party holding a Tier D draft pays only the verification and certification delta to obtain Tier C. This is the economic engine that lowers certified costs, because human effort is spent confirming and correcting a draft the machine has already produced. Second, tier enforcement. Watermarks, distinct visual design, and metadata that filing systems check. The predictable risk of the whole model, uncertified drafts leaking into use as if they were the record, is met with engineering and rules together.

Within Tier C the machine’s role has firm limits. Engines draft, flag doubt, and inform. Humans author the record. An independent checker model, deliberately from a different model family than the drafting engine so that their failures do not correlate, compares certified text against audio and raises dissents for human resolution. It can say that a passage does not match the audio. It can never fix anything. Machine drafting, independent machine dissent, human authorship and named human certification are four different failure modes, and all four would have to align for a fabrication to survive into the record.

06The engine bay

Four rules govern the AI layer. Each one does anti-obsolescence or sovereignty work.

Sovereign inference only. All engines run inside a court-controlled environment assessed under the Australian Government’s IRAP scheme against the Information Security Manual, with a separated high-side partition for closed-court and national-security material. Audio never leaves the boundary for processing. This prevents the offshore-access incident from recurring at the level of architecture. The data cannot go where the policy forbids, because there is no path.

Open-weight preference. Engines the courts can run, fine-tune and keep are preferred. Commercial engines are admissible only if fully deployable within the boundary. The courts hold a unique national asset for this purpose: decades of certified transcripts aligned to audio, across Australian accents and federal legal terminology. Used lawfully and privately, that corpus fine-tunes engines into the best Australian court recogniser in existence, owned by the public.

A fixed engine contract. Audio plus channel map in; a time-aligned, confidence-scored, speaker-attributed token stream out. Every engine conforms, so every engine is replaceable without touching anything else.

A permanent evaluation harness. A curated ground-truth corpus, privacy-cleared and spanning matter types, registries and audio conditions. Standing metrics for word error, attribution error and legal-terminology fidelity. Challenger engines evaluated continuously and promoted when they beat the incumbent, with results published annually. The UK Ministry of Justice is testing its in-house engine, Justice Transcribe, against the accuracy of contracted human transcribers as a one-off study.¹³ Here the same idea becomes the permanent metabolism of the AI layer. Whatever engine exists in 2031 or 2041 is adopted the year it wins the evaluation. That is what “not obsolete before it starts” means in practice.

Assistance beyond drafting stays bounded. Terminology packs per jurisdiction and list, maintained by the standards office and applied at draft time. Redaction and anonymisation candidate detection for the heavy statutory workloads, with a human confirming every candidate. Flagging of crosstalk, low confidence and inaudible passages. Explicitly outside the record chain: summarisation as the record, silent correction, and gap-filling. An engine that is confidently fluent where the audio is unclear is manufacturing the record, and the checker exists to catch exactly that. Judge-assist summarisation tools can exist, as they now do in UK pilots,¹⁴ but as labelled chambers tooling, architecturally separate from the record pipeline.

07Ordering, access and distribution

The courts' emergency arrangement, with a single order form and a central Recording and Transcript Team triaging and distributing work to a provider panel,¹⁵ is the correct skeleton. The long-term model grows that skeleton into a platform.

One front door. A single portal, integrated with the case-management build, where any party, practitioner, media representative or eligible member of the public orders against a specific hearing. It provides an instant quote from regulated rates, tier and turnaround selection, fee-waiver and hardship logic, live status, and delivery to the ordering party and the digital court file at the same time. Same-day and progressive tiers carry their booking cut-offs in the workflow. No party needs to know or care which production firm exists.

Allocation as a system. Orders are matched to accredited capacity by specialisation (family, migration, native title, commercial, appellate), current load, conflict checks, security endorsement where required, and rolling performance data. The in-house core takes sensitive-matter work, surge overflow, and quality-assurance duty. Geography stops mattering. An accredited transcriber in Townsville serves a Perth registry identically, because production happens in the secure workspace.

Audio stays put. By default, audio never leaves the platform. Producers, checkers and reviewers work on streamed, watermarked audio in a secure workspace, with no local copies and every access logged to an Access Register. External release of audio remains exceptional and approval-gated, as it is now, where access requires approval of the relevant Registry Manager,¹⁶ but the release itself becomes digital, watermarked, expiring and fully audited. The class of breach that occurred becomes physically unavailable.

Orders as executable policy. Suppression orders, family-law publication restrictions, protection-visa anonymisation, national-security classifications and native-title cultural restrictions¹⁷ are captured as machine-readable access policies attached to sessions or segments when they are made. The risk is not hypothetical. In 2020 the Federal Court commissioned an independent review after names of protection visa applicants became searchable online contrary to s 91X, in proceedings involving tens of thousands of parties.¹⁸ The platform then enforces them. Restricted audio cannot be allocated to an uncleared producer. A transcript subject to anonymisation generates its redaction workflow automatically. A suppressed segment cannot be delivered while the order stands, and releases automatically when it lapses. Compliance stops depending on every human in the chain remembering every order.

Appeals close the loop. An appellate registry pulls the certified record and its underlying audio directly, with no re-ordering and no re-transcribing of what already exists. Transcript references resolve to timestamps, so the Full Court can hear a disputed answer instead of arguing about it.

08The human system

National accreditation. An Accredited Court Transcriber scheme covering individuals and firms: a competency examination against the published style and format standards, security vetting, and undertakings on confidentiality and access. Endorsements by capability: standard production, same-day, realtime machine shorthand, captioning, native-title protocols, and a security-cleared subset for closed material. Certification is personal. A named human signs the record, and the register knows who. Ontario has run its Authorized Court Transcriptionist model on this trust structure since 2014, with court recordings released to authorised transcriptionists only under a signed undertaking.¹⁹

Regulated fees. The courts set the fee schedule per tier and turnaround. The US federal courts have regulated page rates and delivery schedules through Judicial Conference guidelines for decades, with an original-and-copy rate structure,²⁰ and Ontario regulates transcript fees by regulation.²¹ The schedule should adopt the same original-and-copy logic: the first purchaser pays the certification rate, later purchasers of an existing certified transcript pay a lower copy rate, and the court’s own copy attaches to the digital court file automatically. Regulation ends both price gouging of litigants and the race to the bottom on the humans who certify the record. The interim position, in which each provider sets its own fees,²² should be understood as temporary scaffolding. A modest platform levy inside the regulated fee funds the standards office and, deliberately, traineeships.

An in-house core, established permanently. The interim arrangements already include a panel of transcriptionists managed in-house.²³ This design goes one step further and converts that managed panel into a small, directly engaged core: guaranteed minimum capacity that no commercial event can remove, the natural home for the most sensitive matters, the quality-assurance bench, and the training ground. It can stay small, provided it is permanent and excellent.

Succession as an obligation of the system. Machine-shorthand realtime is irreplaceable at the premium tier, and the workforce is in structural decline. In the United States the average age of National Court Reporters Association members is about 56, stenography school enrolments have collapsed, and official Californian judicial-branch data records a persistent shortage.²⁴ Comparable Australian statistics are not published, but the training pipeline here is thinner still. The levy funds scholarships and supervised traineeships across the accredited network, and the platform gives trainees graded, supervised work. A record system that consumes a profession without renewing it is running down capital and calling it efficiency.

Quality as telemetry. Risk-weighted sampling of certified output against audio by independent reviewers. Every checker-model dissent resolved by a human and logged. Error patterns feeding back into terminology packs, training and allocation weightings. Dual-key control on all post-certification corrections. Quality becomes an observable property of a running system, published annually.

09Accessibility and open justice

Because capture is universal, recognition is streaming, and the record is structured data, services that are traditionally expensive become cheap consequences of the design.

Live captioning in any courtroom, on request, as a standing service. Tier L captions are the baseline for any participant or observer who needs them, upgraded to human-in-the-loop delivery (stenocaptioner) where a deaf or hard-of-hearing party or witness makes accuracy rights-critical. Auslan interpretation integrates as a captured video channel, with interpretation events noted in the record.

Every delivered transcript accessible as delivered. Screen-reader-clean HTML as a primary format alongside the court PDF. Large print and braille-ready as renderings. Plain-language cover sheets for self-represented parties explaining what the document is, what its tier means, and how to seek correction.

Streamed hearings captioned as a matter of course, since the caption track is a by-product of Tier L. And because the marginal cost of the draft tier is near zero, open-justice policy becomes a dial the courts control directly. What the courts choose to make freely available, starting with parties' own drafts, is a policy decision, no longer a commercial negotiation.

10Governance and delivery

A Court Record Office within the Federal Court Entity: small, permanent and senior. It is custodian of the record format specification, the style and terminology standards, and the fee schedule. It operates the accreditation scheme and the evaluation harness. It is product owner of the platform. It publishes an annual public report on cost, turnaround, quality metrics, accessibility delivery and engine performance. Fifteen to twenty-five people is enough, because production belongs to the network and the machines. The seed already exists in the Recording and Transcript Team, and the Entity’s in-house Digital Court Program, which built the CourtPath case-management platform, in operation in the FCFCOA since October 2022 and extended to all Federal Court judges and registrars in March 2024,²⁵ demonstrates that a build-and-own posture is realistic.

Procurement by layer, never across layers. Capture hardware and fit-out from AV integrators under supply contracts. The platform built by systems integrators under contracts that assign intellectual property, or grant a perpetual licence with full source escrow, with an annual exit test in which the courts export all data in open formats and rebuild the platform’s core functions from escrow, so the exit plan stays rehearsed and current. Engines from research bodies, sovereign vendors and open-weight lineages, admitted and retired by the evaluation harness. Capacity from the accredited network at regulated rates. Four markets, four kinds of supplier, no supplier spanning two. No single entity exists whose collapse could take the record with it.

Phasing that respects the emergency. Phase 0, now through 2027: entrench the interim model’s correct instincts (central triage, the in-house-managed panel, the courts holding the data); stand up the Court Record Office; publish the record format v1 and accreditation v1; regulate interim fees; begin the per-channel capture retrofit; and pilot the free draft tier in one high-volume, high-need list with prominent labelling. Phase 1, 2027 to 2028: platform v1 (portal, registers, allocation, secure workspace); engine bay and evaluation harness v1; live captioning on request nationally; and the draft-tier access policy system-wide. Phase 2, 2028 to 2029: the realtime premium tier formalised; the checker in quality assurance at full coverage; deep case-management and appeal-loop integration; and the standards opened to state jurisdictions. After that there is no next big procurement. There is continuous engine refresh, continuous accreditation intake, and annual metrics. The end state is permanent infrastructure with a maintenance budget.

11Cost

Responsible dollar figures require the courts' volumetrics, but the cost logic is structural. Human hours per certified page fall substantially, because paid effort shifts from creating pages to verifying drafts. The draft tier costs little more than compute. Captioning moves from a procured service to a rendered by-product. Regulated fees pass efficiency to litigants instead of vendor margin. The levy funds the standards office and workforce renewal from inside the system. Against that sit real capital in the capture retrofit and platform build, and a permanent Court Record Office operating cost.

The model also ends specific things. The single point of commercial failure. The offshore access path. Format hostage-taking. The assumption that quality can be assured by a contract schedule. And the cycle in which, every several years, the courts must bet the record on somebody’s balance sheet.

12Assumptions

Confidence without stated assumptions is bluster, so here are the load-bearing ones. The Entity’s in-house technology capability can carry a build-and-own posture (evidenced by the in-house CourtPath build,²⁶ though scaling that capability is a real dependency). Regulated fees can be set at levels that sustain a healthy accredited network (the US and Ontario precedents say yes, but calibration is genuine work). Tier labelling plus filing-system enforcement will contain draft-tier leakage (the highest-risk cultural assumption; mitigated, but not eliminable). Open-weight engine quality on Australian courtroom audio continues its current trajectory (the evidence says yes, and the evaluation harness is the hedge if not). And government will fund capital for infrastructure it owns rather than operating expenditure for a service it rents (a Treasury argument the events of 2026 have made easier than it has ever been).

Every one of these is testable in Phase 0, before the majority of capital commits. That is what the phasing is for.