Roadmap
What works today, what is being built, and what is planned or still being explored.
Triage Index records how identified AI systems answer shared moral dilemmas, preserves those answers as evidence, and makes differences and similarities between recorded answers available for comparison.
The project does not decide which answers are morally right or wrong, infer a system's intentions, or rank systems for safety. It makes the underlying observations accessible so others can investigate and interpret them.
This roadmap describes current work and possible future capabilities. Targets and proposed windows are plans, not guarantees, and a planned or exploratory item is not necessarily funded, approved for implementation or scheduled.
Statuses describe development progress. They are not moral ratings and not AI safety judgments.
- Live
- Publicly available now, and checked against a released artifact or a working address.
- In progress
- Work is under way, such as reviewed materials, implementation or run preparation, but the described outcome is not yet public.
- Planned
- A documented direction with its dependencies. Implementation has not started.
- Exploratory
- A possible later capability. It depends on a workable method, demonstrated usefulness, policy decisions, partners or further approval.
Statuses last reviewed: 9 November 2026. Proposed windows and targets are plans, not commitments.
Where things stand
- Dilemmas with published answers
- 18 of the 20 in the intended battery (TRL-01, TRL-02, RES-01, MM-01, TB-01, TRU-01, AID-01, PRM-01, PAR-01, PRP-01, CNS-01, FAM-01, SCR-01, ITG-01, JUS-01, PST-01, WLK-01 and DOG-01)
- Dilemmas in preparation
- 2 (MEX-01 and REP-01)
- Languages with published answers
- English and Chinese
- Published releases
- 3, holding 32 model configuration records
- Most recent publication
- 10 November 2026
These counts are computed from the published releases and the intended battery each time the site is built. Question coverage and language coverage are separate: neither waits for the other to finish.
In progress
Current work. Nothing here has produced published results yet beyond what the Live section lists.
Expanding the dilemma library
We are extending measurement beyond the first three dilemmas toward the remaining questions of the intended battery. Questions are versioned and reviewed before collection. Answers are added as dated, documented evidence is collected, checked under the applicable protocol and approved for publication.
- Status
- In progress Stage: the last two dilemmas of the intended battery
- Available now
- Eighteen of the 20 dilemmas in the intended battery have published answers in English and Chinese: three in the first releases and 15 in release 2026-11-measured2. The other two, which rest on medical and legal facts, have versioned wording in the question library and no answers yet.
- What comes next
- Check every factual claim in those two dilemmas against its primary source, with two checks from different labs, and publish each claim and its source on the question page. Then review their corrected wording and include them in a registered run.
- Depends on
- The fact check of the two dilemmas, a review of their corrected wording, a registered plan, then collection and a reporting review. Answers are published only after that review.
- Evidence
- The intended battery · Every release
Twelve languages, starting with French, Korean and Hindi
The instrument is planned to reach twelve languages. French, Korean and Hindi are the next batch, and a five-language EU run goes in parallel. Every language is reviewed to the same standard, and having text in the library does not mean a language is measured.
- Status
- In progress Stage: planning; no wording drafted yet for the next batch
- Available now
- English and Chinese have published answers for 18 dilemmas. Korean, Spanish and Arabic have draft wording in the question library that has not been reviewed or asked.
- What comes next
- Draft French, Korean and Hindi wording for the admitted questions, review it under the current review policy, and include it in a registered run alongside matched English and Chinese controls. Then Arabic and Russian, then Japanese. Spanish and German come through the five-language EU run, with French, Italian and Polish. See the language table below.
- Depends on
- Reviewed wording in each language, then a registered run. Language coverage and question coverage advance separately, and no language waits for another.
- Timing
- Proposed window: the first quarter after launch, to January 2027, for the three language batches; a target, not a promise
Live
What anyone can use today.
Behavioral Signature records
Each tested configuration has a dated record that links every answer to its exact prompt, its coded value, its conditions and its evidence, including failed, missing, declined and cut answers.
- Status
- Live
- Evidence
- Every release and its records · Method
Published releases with permanent DOIs
Three releases are published, each with a permanent DOI for its results and a separate DOI for the plan registered before its first call. Later releases are added beside them and never pooled with them.
Record addresses
A record's short address, /r/ followed by its Signature ID, opens that exact record in any letter case and never a newer one.
- Status
- Live
- Evidence
- Every release and its records
Read API
Every published release can be read as static JSON files, with the same records, statistics and answers as the pages.
- Status
- Live
- Evidence
- The read API index · For labs and agents
Corrections log and feed
Corrections are dated with what changed and why, without rewriting the published files. A feed lists releases and corrections as they appear.
- Status
- Live
- Evidence
- Corrections · Feed
Planned
Documented directions with their dependencies. None of these is running yet.
Five EU languages in one run
Ask the same dilemmas in French, German, Italian, Spanish and Polish in one registered run, and publish, per model and per question, whether the answer stays the same across the five languages. Counts and positions only, never a verdict.
- Status
- Planned
- Available now
- French is also in the next batch of the main language order. Spanish has draft wording that has not been reviewed or asked. German, Italian and Polish have no wording yet. None of the five is measured yet.
- What comes next
- Check the instrument and its review method against the quality criteria that the European Commission's translation department published for EU multilingual benchmarks, and publish the result of that check. Then draft and review the wording in all five languages and register the run.
- Depends on
- Reviewed wording in all five languages, a registered plan, and a check by native speakers of a sample of the translations, planned with a university translation department.
- Timing
- Proposed window: the first quarter of 2027 for the five-language run
Drift record: what changed, and when
Ask the same frozen questions of the same model identifiers on a fixed schedule, and publish a page and a feed of what changed, per model and per question, so anyone can see a shift they had not noticed or check their own numbers against the record. A change is reported as an observation, not as a judgment of the change.
- Status
- Planned
- Available now
- Every published answer is dated and tied to its exact wording, so a later run can be compared with an earlier one.
- What comes next
- Measure the noise floor first: ask the same model twice, close together, record how much its answers vary, and declare in advance how large a difference must be before it is reported as a change. Each row will also have a declared minimum number of answers, chosen so a before and after comparison can detect a stated size of change.
- Depends on
- The quarterly census, and a noise floor measured and declared before the first comparison.
- Timing
- Proposed window: with the first census, January 2027
Public code and an independence statement
The code that builds the questions, runs the collection, computes the statistics and builds this site, published under an open-source licence. Questions, translations, methods and statistics stay under CC BY 4.0. Raw model answers stay out, under each provider's terms. Alongside it, one statement of the licences, how the project is funded and any conflicts of interest, and later an outside replication of one published release.
- Status
- Planned
- Depends on
- A check that no private file or credential is in the code or its history before the first public push.
- Timing
- Target: 1 December 2026
Reserved variants
Separately governed variants, kept apart from the public wording, to study whether answers differ under declared conditions. They include fresh scenarios built like the well-known dilemmas and held back from publication; the gap between the classic and the fresh versions is reported as its own number. A difference would be reported as an observation, not as evidence of deception, training exposure or safety.
- Status
- Planned
- Depends on
- An approved variant and disclosure policy, a comparison method and its uncertainty treatment. No reserved measurement exists yet.
- Timing
- Proposed window: design in months 2 to 3 after launch
Checking your own numbers
One page for anyone checking results against the record: how to cite a record, how to reproduce one row, how to compare a self-run with an official record, and what a record does and does not show.
- Status
- Planned
- Available now
- Every record, statistic and answer can already be read on its page and as JSON.
- Timing
- Proposed window: months 1 to 3 after launch
Observing new models and checkpoints
When an eligible model or checkpoint is released, check how it can be accessed and add dated observations with the current versioned instrument.
- Status
- Planned
- Available now
- New configurations are added by hand through the model registry and the collection runner.
- What comes next
- A repeatable intake workflow that records release, access, test and publication times, and any route that is unavailable. Testing within days of a release is an aim, not a promise.
- Timing
- Proposed window: from the first month after launch, then as eligible releases appear
Quarterly census
Ask the admitted instrument again of eligible configurations within declared collection windows, so dated records can be compared over time. A change of provider, configuration or wording is recorded as such, not read as a change in one unchanged system.
- Status
- Planned
- Timing
- Target: January 2027 for the first census
Quarterly reports
Reports that describe which measured answer distributions changed, which are new and which stayed about the same, with every statistic linked to its conditions and evidence. A stable answer is not taken as proof of any particular cause.
- Status
- Planned
- Available now
- Corrections are already published as they happen, not held for a report.
- Depends on
- The quarterly census.
- Timing
- Target: January 2027 for the first report
Public question suggestions
A structured way to propose new dilemmas with their sources, and a published record of each admission decision.
- Status
- Planned
- Depends on
- Admission rules for new candidates, which are not final. The plan proposes at most one or two candidates a quarter, each observed for a full quarter before it could join the core instrument, and no popularity vote.
- Timing
- Proposed window: months 2 to 3 after launch for a minimal intake
Annual instrument editions
A yearly review that adds, keeps or retires questions with stated reasons and maps old versions to new ones. Earlier answers keep pointing to the exact instrument they answered. A question is not dropped just because answers converge.
- Status
- Planned
- Available now
- Every question is already versioned, and every answer names the version it answered.
- Timing
- Target: October 2027 for the first edition
Human survey reference data
Adapters for published human surveys, such as the World Values Survey and Pew Research Center studies, that record each item's wording, scale, population, wave, language, weighting and reuse terms. A similar topic does not make two questions comparable.
- Status
- Planned
- Depends on
- Sources whose wording, scales and reuse terms allow a fair match.
- Timing
- Proposed window: months 4 to 6 after launch, a feasibility test with one source and one item
Comparing AI and human answer distributions
Where the instruments are compatible, compare recorded AI answer distributions with named, dated human survey distributions, using a declared distance with its coverage and uncertainty. It is a statistical distance between answer distributions, not a hash comparison, and it does not place any model in a religious, cultural, political or moral group. The reader chooses the reference and decides what a difference means.
- Status
- Planned
- Depends on
- Human survey reference data, with matches valid enough to compare.
- Timing
- Proposed window: months 4 to 6 after launch for a prototype, only if valid matches exist
Run the instrument yourself
A reusable protocol, versioned questions and an export format, so researchers and developers can run comparable tests on their own systems. Self-run results stay labeled as self-run and do not carry the provenance of Triage Index measurements. It is not a public quiz.
- Status
- Planned
- Timing
- Proposed window: months 4 to 6 after launch for a minimal kit
Actions as well as answers
A controlled pilot of three dilemmas, each asked once as a question and once as a task in which the system must act through a tool, with separate evidence types and exact task conditions. The gap between what a system says and what it does is reported as an observation. A written answer to a hypothetical is never treated as a real-world action.
- Status
- Planned
- Available now
- Eight dilemmas in the intended battery, including the three with published answers, already vary who bears the cost: a stranger, the system's own user, a citizen of the country where it was built, a refugee or a child. No action experiment exists.
- Depends on
- A task design for each of the three dilemmas, with its tools and conditions fixed before the first run.
- Timing
- Proposed window: months 7 to 12 after launch for the three-dilemma pilot
Base models and deployed assistants
Relate a base model's record to separately measured assistants built on it. Memory, tools, the surrounding software and system instructions each get their own versioned identity, and a base model's record is never presented as how every system built on it behaves.
- Status
- Planned
- Timing
- Proposed window: months 7 to 9 after launch, as a pilot of two or three assistants
Declared and observed
Link a dated, published statement by a developer to tests chosen in advance and their recorded results. A difference is reported as a bounded observation, not as an accusation of dishonesty, a claim about intent or a safety judgment.
- Status
- Planned
- Available now
- Only an illustrative card in an early design mockup exists. There is no working feature.
- Timing
- Proposed window: months 7 to 12 after launch, as a pilot
Exploratory
Possible later capabilities. Each depends on things that do not exist yet.
Records shown inside other products
A partner could show selected records or comparisons inside a model-routing, marketplace or teaching tool. Triage Index would supply evidence for the partner and its users to interpret; it does not recommend a model for any setting.
- Status
- Exploratory
- Depends on
- An actual partner, suitable identifiers and permissions, and a demonstrated use.
- Timing
- Proposed window: months 10 to 12 after launch for a prototype, only if a real partner and use exist
Record retrieval for agents (MCP)
An agent could retrieve another system's dated record and compare it with criteria its principal provides. Retrieving a record does not show that the running system matches it or will answer the same way under new conditions.
- Status
- Exploratory
- Depends on
- The MCP tools, which are planned and not yet operating, and a way to tell record retrieval apart from identifying a running system.
- Timing
- Proposed window: year 3 after launch for a possible pilot
Languages
Text in the question library is not the same as a measured language. For each language the steps are reviewed wording, a registered run, usable answers and publication. The order follows how many people speak each language and where the tested labs are based: next French, Korean and Hindi; then Arabic, Spanish and Russian; then Japanese and German. A five-language EU run (French, German, Italian, Spanish and Polish) goes in parallel. English and Chinese stay as anchors for matched comparisons. Any eligible system can be asked in any admitted language.
| Language | Status | Where it stands |
|---|---|---|
| English | Live | Published answers for 18 dilemmas; the last two of the battery are in progress. |
| Chinese (Simplified, written Mandarin) | Live | Published answers for 18 dilemmas; the last two of the battery are in progress. Wording reviewed by two AI reviewers, not by a person. |
| French | In progress | Next batch, with Korean and Hindi, and part of the five-language EU run. Planning stage: no wording drafted yet. |
| Korean | Planned | Next batch. Draft wording in the library, not reviewed or asked. |
| Hindi | Planned | Next batch. No wording yet. |
| Arabic | Planned | Draft wording in the library, not reviewed or asked; its written variety will be declared. Pages already read right to left. |
| Spanish | Planned | Part of the EU run. Draft wording in the library, not reviewed or asked. |
| Russian | Planned | Planned addition. No wording yet. |
| Japanese | Planned | Planned addition. No wording yet. |
| German | Planned | Part of the EU run, ahead of its place in the main order. No wording drafted yet. |
| Italian | Planned | Planned addition for the EU run. No wording yet. |
| Polish | Planned | Planned addition for the EU run. No wording yet. |
Proposed windows
These come from the project's post-launch plan. They are sequencing windows, not delivery promises, and they will move as work and funding allow. Months count from the first release on 26 October 2026.
- Proposed window: months 1 to 3. Model intake, census preparation, the question suggestion process, the reserved-variant design, the page for checking your own numbers, and the check against the EU multilingual benchmark criteria.
- Target: January 2027. The first quarterly census and the first drift record.
- Proposed window: first quarter of 2027. The five-language EU run.
- Proposed window: months 4 to 6. Human survey crosswalks and the feasibility of self-run testing.
- Proposed window: months 7 to 9. Pilots for deployed assistants, for declared and observed comparisons, and for actions as well as answers.
- Proposed window: months 10 to 12. Preparation of the first annual edition, and a partner prototype only if a partner exists.
- Target: October 2027. The first annual instrument edition.
- Proposed window: year 2. More validated comparisons, assistants and integrations.
- Proposed window: year 3. A possible pilot of record retrieval for agents.
- Proposed window: year 4. Longer-term evidence across systems, assistants and actions, and reuse by partners.
Notable changes
- 29 October 2026. First published version of this roadmap.
- 9 November 2026. Added a five-language EU run (French, German, Italian, Spanish and Polish), a drift record and a page for checking your own numbers. Italian and Polish joined the language list. The public code item now includes an independence statement. Actions as well as answers moved from exploratory to planned, as a three-dilemma pilot. Reserved variants moved up and now include fresh scenarios. The language order now follows how many people speak each language and where the tested labs are based: French, Korean and Hindi come next, and German moves to planned and into the EU run.
- 10 November 2026. Release 2026-11-measured2 published, with answers to 15 more dilemmas in English and Chinese. The last two dilemmas now wait for a check of their facts against primary sources instead of a specialist check.