Independent reproduction will determine whether MirrorCode’s detailed, checkable specifications enable frontier coding agents to autonomously reimplement substantial real-world software over multi-week-equivalent horizons.

state: expiredheat: lowuncertainty: highknownscott: highcoding-agents long-horizon-agents agent-evalsEpoch AI METR Anthropic

What is this?

MirrorCode is a benchmark co-developed by Epoch AI and METR for testing whether coding agents can reimplement substantial existing software without seeing its source code, using detailed specifications and canonical outputs as checks. Preliminary results report that Anthropic’s Claude Opus 4.6 autonomously rebuilt gotree, a Go bioinformatics toolkit of roughly 16,000 lines and more than 40 commands, suggesting that some tightly specified coding projects can be sustained over task horizons associated with weeks of human work. The supplied material emphasizes that this is a particular, specification-rich setup and that broader implications remain uncertain; it does not establish that the result has yet been independently reproduced.

Why it matters to Scott

The radar already tracks this same development in `radar:mirrorcode-autonomous-project-scope`. Its eventual reproduction or failure would directly test Scott’s claims that substantial software can be regenerated from specification and behavioural oracles, while clarifying whether apparent weeks-long capability reflects the model alone or the complete agent system and evaluation setup.

ip:framework.ai-legacy-takeoverip:concept.specification-as-assetip:framework.long-running-agentsip:concept.benchmarking-the-wrong-unitradar:mirrorcode-autonomous-project-scoperadar:concept.software-reconstructionradar:concept.agent-evaluationradar:concept.long-horizon-agents

queries asked of Scott's wikis

long-horizon coding-agent harnesses and autonomy
executable specifications as agent verification oracles
coding-agent evaluation beyond issue-resolution benchmarks
agent task decomposition, recovery, and multi-session memory
autonomous software reimplementation and clean-room workflows
human-time estimates versus agent task-horizon measurement

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (3) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	MirrorCode: Evidence AI can already do some weeks-long coding tasks singularity	141_1337	71	17
🟧 echo.blog ⭐	Preliminary MirrorCode results report that Claude Opus 4.6 autonomously reimplemented gotree, a bioinformatics toolkit with roughly 16,000 l	Epoch AI and METR	—	—
🟧 hn	200B Tokens Later: A Month of Letting AI Agents Decompile MW2	Gander5739	2	0

Interpretation history

2026-08-27T12:26:59Z
No independent reproduction, failure report, artifact, or methodological disclosure arrived across the full monitoring ladder. This does not disprove MirrorCode, but the active episode has faded and should reopen only on substantive validation evidence.
2026-08-25T11:30:37Z
The MW2 item establishes only that an adjacent month-long, high-token decompilation experiment exists; without results, artifacts, verification details, or autonomy disclosures, it is not yet an independent reproduction or capability finding. MirrorCode therefore remains a consequential but unvalidated claim awaiting substantive external evidence.
2026-08-25T11:24:05Z
evidence attached: hn.story.49431593 — Independent month-long evidence of agents spending 200B tokens on substantial software decompilation materially informs the open question of long-horizon autonomous coding performance.
2026-08-24T11:22:34Z
Another staleness-only review adds no reproduction, failure report, or methodological disclosure, so the validation question remains unchanged. Keep the case open but suppress routine revisits until substantive external evidence appears.
2026-08-22T10:28:51Z
The staleness threshold adds no substantive evidence and does not resolve a validation question whose natural horizon is longer than a few days. Keep the case open but defer further review until an independent reproduction, failure report, or meaningful methodology disclosure appears.
2026-08-20T09:40:08Z
The Reddit velocity spike is repetitive amplification of the original preliminary result, not independent reproduction or new methodological evidence. MirrorCode remains a consequential but unvalidated test of specification-driven long-horizon coding, with no basis for promotion or renewed attention.
2026-08-19T14:35:23Z
No independent reproduction, failure report, or methodological disclosure has appeared; the elapsed time and repeated engagement-only refreshes do not change the case’s meaning. Keep it open as a long-horizon validation watch, but cool and revisit only when substantive external evidence arrives.
2026-08-17T14:08:00Z
Refreshed discussion adds skepticism about specification cost, public reference implementations, and verifier dependence, but no independent reproduction or methodological finding. The case remains a preliminary, testable claim whose meaning still hinges on external validation.
2026-08-17T06:27:57Z
No independent reproduction, implementation, or methodological evidence has arrived; the slight engagement increase does not change the preliminary claim. The case remains a high-relevance but unresolved test of specification-driven, long-horizon coding agents.
2026-08-17T06:26:04Z
grounded: known/high — The radar already tracks this same development in `radar:mirrorcode-autonomous-project-scope`. Its eventual reproduction or failure would directly test Scott’s
2026-08-17T06:23:01Z
case created — A first-party benchmark artifact makes a concrete, independently testable claim about coding-agent performance on unusually long-horizon software reconstruction.

Decision trace

08-27 22:27expireNo independent reproduction, failure report, artifact, or methodological disclosure arrived across the full monitoring ladder. This does not disprove MirrorCode, but the active episode has faded and s
08-27 22:26alert_silentThe only new delta is staleness; it adds no capability finding or actionable information and does not justify another briefing mention.
08-27 22:26alert_routeThe only new delta is staleness; it adds no capability finding or actionable information and does not justify another briefing mention.
08-27 10:29drop_targetsquiet through full ladder or over cap 8
08-25 21:30repriceThe MW2 item establishes only that an adjacent month-long, high-token decompilation experiment exists; without results, artifacts, verification details, or autonomy disclosures, it is not yet an indep
08-25 21:30alert_silentThe new title lacks enough evidence to establish what the agents achieved or how independently they operated, so alerting today would overstate an experiment as validation; revisit when the project re
08-25 21:30alert_routeThe new title lacks enough evidence to establish what the agents achieved or how independently they operated, so alerting today would overstate an experiment as validation; revisit when the project re
08-25 21:24alert_silentThe linked first-party project title suggests a potentially important month-long, 200B-token software reimplementation experiment, but the available evidence contains no results, verification method,
08-25 21:24surface_candidateThe linked first-party project title suggests a potentially important month-long, 200B-token software reimplementation experiment, but the available evidence contains no results, verification method,
08-25 21:24alert_routeThe linked first-party project title suggests a potentially important month-long, 200B-token software reimplementation experiment, but the available evidence contains no results, verification method,
08-25 21:24attachIndependent month-long evidence of agents spending 200B tokens on substantial software decompilation materially informs the open question of long-horizon autonomous coding performance.
08-25 21:23propose_attachIndependent month-long evidence of agents spending 200B tokens on substantial software decompilation materially informs the open question of long-horizon autonomous coding performance.
08-24 21:22repriceAnother staleness-only review adds no reproduction, failure report, or methodological disclosure, so the validation question remains unchanged. Keep the case open but suppress routine revisits until s
08-24 21:22alert_silentThe small engagement drift is not a consequential delta and does not address independent validation; another briefing would only repeat the preliminary claim.
08-24 21:22alert_routeThe small engagement drift is not a consequential delta and does not address independent validation; another briefing would only repeat the preliminary claim.
08-22 20:28repriceThe staleness threshold adds no substantive evidence and does not resolve a validation question whose natural horizon is longer than a few days. Keep the case open but defer further review until an in
08-22 20:28alert_silentNo consequential delta occurred; another alert would merely repeat the preliminary MirrorCode claim without addressing independent validation.
08-22 20:28alert_routeNo consequential delta occurred; another alert would merely repeat the preliminary MirrorCode claim without addressing independent validation.
08-20 19:40repriceThe Reddit velocity spike is repetitive amplification of the original preliminary result, not independent reproduction or new methodological evidence. MirrorCode remains a consequential but unvalidate
08-20 19:40alert_silentEngagement alone does not alter the central validation question, so alerting would recycle an already-known claim; wait for an independent reproduction, failure report, or substantive methodology disc
08-20 19:40alert_routeEngagement alone does not alter the central validation question, so alerting would recycle an already-known claim; wait for an independent reproduction, failure report, or substantive methodology disc
08-20 19:21sensor_dirtyengagement_update
08-20 17:21sensor_dirtyengagement_update
08-20 11:21sensor_dirtyengagement_update
08-20 00:35repriceNo independent reproduction, failure report, or methodological disclosure has appeared; the elapsed time and repeated engagement-only refreshes do not change the case’s meaning. Keep it open as a long
08-20 00:35alert_silentThere is no new consequential delta—only staleness—so alerting would repeat the already-known preliminary claim without resolving its central validation question.
08-20 00:35alert_routeThere is no new consequential delta—only staleness—so alerting would repeat the already-known preliminary claim without resolving its central validation question.
08-18 20:21sensor_dirtyengagement_update
08-18 08:21sensor_dirtyengagement_update
08-18 03:21sensor_dirtyengagement_update
08-18 01:21sensor_dirtyengagement_update
08-18 00:21sensor_dirtyengagement_update
08-18 00:08repriceRefreshed discussion adds skepticism about specification cost, public reference implementations, and verifier dependence, but no independent reproduction or methodological finding. The case remains a
08-18 00:08alert_silentThe new comments only rehearse known caveats and anecdotal impressions; they neither validate nor falsify MirrorCode’s result, so this can wait for an independent reproduction, failure report, or subs
08-18 00:08alert_routeThe new comments only rehearse known caveats and anecdotal impressions; they neither validate nor falsify MirrorCode’s result, so this can wait for an independent reproduction, failure report, or subs
08-17 23:21sensor_dirtycomment_update
08-17 22:21sensor_dirtyengagement_update
08-17 20:21sensor_dirtyengagement_update
08-17 17:21sensor_dirtyengagement_update
08-17 16:27repriceNo independent reproduction, implementation, or methodological evidence has arrived; the slight engagement increase does not change the preliminary claim. The case remains a high-relevance but unresol