A Proposal to Extend the Common Model of Cognition with Metacognition

Andrea Stocco; Christian Lebiere; John Laird; Paul Rosenbloom

REVIEW 3 major objections 4 minor 2 cited by

A Proposal to Extend the Common Model of Cognition with Metacognition

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Metacognition is ordinary cognition applied to working-memory representations of one's own mind, the paper argues.

desk verdict A useful conceptual synthesis of metacognition in the CMC, but the third extension quietly imports control mechanisms the paper claims not to need. read the letter →

arxiv 2506.07807 v2 pith:SWL456GB submitted 2025-06-09 cs.AI

John Laird , Christian Lebiere , Paul Rosenbloom , Andrea Stocco This is my paper · ORCID

classification cs.AI

keywords metacognitionCommonModelofCognitioncognitivearchitectureworkingmemoryepisodicmetareasoningprocess-statebuffersself-representation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the Common Model of Cognition, an abstract blueprint for flexible minds shared by many cognitive architectures, can account for self-awareness and self-regulation without adding a separate supervisory system. The central move is to treat metacognition as ordinary reasoning over explicit working-memory representations of the agent's own cognition: retrieval failures, confidence levels, traces of past reasoning, and imagined future states. To supply those representations, the paper proposes three minimal structural extensions to the model: per-module process-state buffers, a distinct episodic memory for the agent's own reasoning history, and a working-memory mechanism for hypothetical states that are marked as unreal. If this proposal is right, metacognition is not a distinct faculty but a mode of the same cognitive cycle, which means existing CMC-based architectures can be extended to monitor and control their own thinking without new reasoning modules or direct access to long-term memories.

What carries the argument

The load-bearing mechanism is the CMC's cognitive cycle: each cycle, procedural memory inspects working memory and selects a single action that rewrites part of working memory. The paper's innovation is to let that same cycle inspect self-referential content. Three machinery pieces carry the argument. First, process-state buffers attached to every module summarize module states—success or failure, confidence, partial results, feelings of knowing, surprise, desirability—into working memory, providing initiation signals and raw material for metareasoning. Second, a functionally separate episodic memory automatically records the agent's reasoning history, so sequences of past situations can be retrieved into working memory and analyzed retrospectively. Third, working memory supports hypothetical-state representations that satisfy two constraints at once: they are structured so existing procedural knowledge can reason over them as if they were real, and they are marked as non-real so the agent does not confuse them with the current situation. Existing CMC-compatible architectures already have partial versions of these mechanisms, which the paper reads as evidence that the extensions are natural rather than ad hoc.

What would settle it

Take a CMC-compatible architecture with exactly the three proposed extensions, no additional metacognitive module, and no direct long-term-memory readout, and run it on the paper's own examples—an uncommon-but-familiar word retrieval, a chess move after memorized openings are exhausted, and a robot that must notice it repeats a door-opening action. If reliable metacognitive behavior appears, the proposal's sufficiency is supported; if the agent needs a supervisory loop outside the single-action cycle, or needs to inspect long-term memories directly, the claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that metacognition in the Common Model of Cognition reduces to a question of information availability, not a question of additional processing power. An agent becomes metacognitive when working memory contains explicit, retrievable representations of its own cognitive processes—its failures, confidence, familiarities, past reasoning traces, and imagined alternative states—and when procedural knowledge can treat those representations as the subject matter of reasoning. To make this possible, the paper extends the CMC in exactly three places: each module receives a process-state buffer that reports into working memory, episodic memory is separated from semantic memory so past reasoning can be reconstructed, and working memory can hold hypothetical past or future states that are marked as distinct from reality. No new reasoning modules, no direct access to long-term memories, and no hierarchical control layer are required. The proposal is presented as a framework, not a complete recipe; it specifies the architectural structure that must exist beyond whatever metacognitive knowledge is encoded in long-term memory.

Load-bearing premise

The proposal assumes that the ordinary single-action-per-cycle cognitive loop, with everything reasoned over in working memory, is powerful enough to support self-monitoring and self-control of cognition without any dedicated supervisory mechanism.

Editorial extensions

If this is right

Metacognitive judgments become inferences drawn from ordinary working-memory contents—confidence flags, retrieval failures, familiarities—rather than direct readouts of hidden cognitive states.
An agent can switch between task-level reasoning and metareasoning within the same cycle, so metacognition does not require a separate executive layer or parallel monitoring process.
Because episodic memory makes the agent's own reasoning history retrievable, the agent can retrospectively analyze past behavior and use existing learning mechanisms to turn those insights into improved procedural knowledge.
The same long-term knowledge and learning capabilities used for ordinary cognition are available for metacognition, so metareasoning can be practiced, learned, and eventually automated like any other skill.
By restricting metareasoning to what is in working memory, the proposal rules out omniscient introspection of long-term memories; metacognition is efficient and neurally plausible but limited to what has been brought to mind.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial extension: if the proposal is correct, self-reports of confidence and 'feeling of knowing' should be predictable from just the process-state information and episodic traces available in working memory—a quantitative prediction that could be tested against human metacognitive judgments.
Editorial extension: the restriction to working memory implies that metacognitive blind spots are not bugs but architectural features; an agent cannot introspect what it has not represented, which may explain why humans are often poor at introspecting automatic processes.
Editorial extension: a minimal implementation test would be to equip a single CMC-compatible architecture with only these three extensions and run it on the paper's Wordle, chess, and robot examples; success would demonstrate the sufficiency claim, while failure would push toward the hierarchical designs the paper rejects.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Thanks for the report. My take: this is a real synthesis, not a revolution. The three extensions—process-state buffers, separating episodic from semantic memory, and hypothetical states in working memory—are concrete and map honestly onto ACT-R, Sigma, and Soar. The authors know what their architectures do. The paper also defines its scope: it is a framework proposal, and the conclusion says so. That counts.

The weak point is exactly where the stress test lands. The third extension does not stay representational. To let base-level knowledge apply to a hypothetical state without executing the imagined action, you need something that makes that state inert. In Sigma and Soar, that something is the substate machinery, which suspends normal operator execution and creates a nested context. That is an architectural control mechanism, not just a new source of information in working memory. So the claim that metacognition uses the same cognitive cycle and needs no new modules is oversold for exactly the case that matters. ACT-R, by the authors' own admission, falls back on knowledge-based conventions, which is a different kind of answer. Either widen what counts as an extension or weaken the same-cycle claim.

Other soft spots are real but minor. Minimality and sufficiency are asserted, not argued; no formal specification, no implementation; the examples are illustrations, not evidence. The relation to Kralik et al. 2018 is acknowledged but the delta is not sharply drawn. None of this is disqualifying for a position paper, but the central load-bearing claim needs another pass.

The audience is the cognitive architecture community and AI safety people interested in metacognition. It is not an empirical paper. I would send it to review: it is a clear, useful proposal from people who know the territory, and the control-mechanism issue is exactly what a good referee should push on. I would cite it if I worked on metacognition in architectures, and I'd bring it to reading group for the argument.

Referee Report

3 major / 4 minor

Summary. This paper proposes extending the Common Model of Cognition (CMC) with three architectural additions: module process-state buffers in working memory, a separation of episodic from semantic memory, and the ability to represent hypothetical past or future states in working memory. The authors argue that metacognition is reasoning over explicit representations of an agent's own cognitive processes in working memory, and that this can be supported with the same cognitive cycle and no new reasoning modules. They map the proposal onto ACT-R, Sigma, and Soar, and illustrate it with three narrative examples (Wordle, chess, and a robot chore).

Significance. If the central claim is borne out, the proposal would be a valuable unifying framework for metacognition in cognitive architectures, extending the CMC's consensus-model approach to a domain the original CMC explicitly set aside. The authors are the originators of the CMC and of two of the three architectures discussed, which gives the paper unusual authority on the compatibility question. The paper does not provide formal specifications or implementations, and the examples are narrative rather than validated, but as a conceptual proposal it could serve as a useful roadmap. Its main weakness is that the third extension silently relies on a control mechanism that goes beyond the stated representational additions, which undermines the 'same cycle, no new modules' claim unless addressed.

major comments (3)

[Section 4.1, Hypothetical State Representations] This extension is load-bearing for the paper's 'same cycle, no new modules' claim (Sections 3 and 4), but as stated it cannot work. Under the CMC (Section 2), procedural memory tests the full contents of working memory every cycle and selects a single action that is actually executed, changing working memory and potentially initiating motor, perceptual, or retrieval operations. If a hypothetical state is encoded in working memory in the same format so that base-level knowledge applies to it, nothing in the cycle prevents that knowledge from selecting an action that is executed. The 'seemingly contradictory properties' named in Section 3.2 are left unresolved: the representation must support base-level reasoning about the hypothetical state while also being distinguished so that it does not interfere with reality, but the distinguishing mark alone does not block execution. Soar and Sigma solve this with substates, which are an architectural control mechanism that suspends base-level operator execution and creates a nested context; ACT-R, as the paper notes, has no such mechanism and falls back on knowledge-based conventions. The paper therefore either silently imports a control mechanism beyond the stated representational extension, or it delegates the burden of non-execution to procedural knowledge, which contradicts the claim that the same knowledge can be used in the hypothetical state. Please specify the control mechanism explicitly, or show how the representational distinction alone prevents execution, and then reassess whether the proposal has three extensions or four.
[Sections 3.1 and 4.1, Module process-state buffers] The proposal does not specify how process-state buffers are updated. Each module has a process-state buffer in working memory (Section 4.1), and these buffers are described as a direct source of information about module state. But in the CMC, working memory is modified by the single selected action of procedural memory or by module results posted to their buffers. If process-state buffers are updated automatically at the end of every module operation, that is a new architectural mechanism not listed among the three extensions. If they are updated by deliberate procedural actions, then metacognitive monitoring is itself an action, and the paper's claim that metacognition can be triggered as a 'side-effect of other long-term memory retrievals or perception' (Section 3.1) is unsupported. Please specify the update semantics for process-state buffers and how this preserves the 'no new modules' claim.
[Section 4 and Section 5] The paper states that it attempts to identify 'the minimum architectural information required to support general metacognition' (Section 4), but it offers no systematic argument that the three proposed extensions are sufficient or minimal. The three examples in Section 5 are narrative and do not formally trace how the proposed extensions combine to produce the claimed metacognitive behavior. For instance, the Wordle example (Section 5.1) relies on a familiarity signal generated by episodic memory, but the paper does not explain why this requires a separation of episodic from semantic memory rather than a confidence metadata on a single declarative memory. Without a derivation from the requirements in Section 3 to the extensions, the minimality claim is an assertion. I am not asking for a formal proof, but the paper would be substantially stronger if it argued more explicitly why these three extensions are necessary and jointly sufficient.

minor comments (4)

[Section 4.1, Figure 3] The third extension (hypothetical state representations) is explicitly 'not shown' in Figure 3; the figure should be updated to represent it, since it is one of the three central proposals.
[Section 3.3] The termination condition is stated only informally as 'reasoning is no longer sensitive to representations of the agent's processing.' A more precise criterion, such as a condition on working memory contents or on the absence of impasses, would make the proposal more testable.
[Section 5] The examples would be more convincing if each were accompanied by a step-by-step trace through the CMC cycle, showing which buffer contents trigger which procedural actions and where the proposed extensions enter. As written, the examples read as plausible stories rather than as demonstrations of the architecture's operation.
[Keywords] The keywords appear as a single run-on string ('CommonModelofCognition,CognitiveArchitecture,Metacognition'); please format them as separate items.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a design proposal with no fitted predictions or equation-level derivation.

full rationale

The paper is a design proposal, not a derivation: it makes no quantitative predictions and fits no parameters. Its central claim is that metacognition can be realized by adding three working-memory extensions to the CMC, justified by reference to existing architectures (ACT-R, Sigma, Soar) and by illustrative examples, not by equations. The closest to a self-referential move is that the CMC and its example architectures are largely defined in the authors' prior work, but those citations are background definitions and architectural evidence, not circular inputs that force the conclusion. The hypothetical-state extension is underspecified and possibly inconsistent with ACT-R's lack of substates, but that is a correctness or feasibility concern, not a circularity. No step reduces to its own definition or to a fitted quantity, so the proposal is self-contained in the relevant sense.

Assumptions & free parameters 0 free parameters · 3 assumptions · 1 invented entities

The proposal contains no fitted numerical parameters. It rests on domain assumptions about working memory as the locus of metacognition, the sufficiency of the standard cognitive cycle, and the separability of episodic memory. One new architectural component, process-state buffers, is introduced without independent evidence.

assumptions (3)

domain assumption Metacognitive reasoning can be carried out by the standard CMC cognitive cycle using working memory contents, with no new control module.
Section 3 states metacognition "employs the same cognitive cycle process, utilizing the same modules, and often the same knowledge." The paper does not prove that this cycle is sufficient for self-referential monitoring and control.
domain assumption Information about an agent's own cognition can be represented symbolically in working memory.
Section 3.2 requires process-state information, episodic traces, and hypothetical states to be available in working memory; the representability assumption is inherited from CMC but unproven.
domain assumption Episodic memory functionality can be separated from semantic memory without disrupting existing CMC learning mechanisms.
Section 4.1 proposes the separation; Section 2's automatic learning mechanisms (procedural compilation, reinforcement, declarative learning) are not analyzed for interactions.

invented entities (1)

Module process-state buffers in working memory
purpose: Provide each module's current status (success, failure, confidence, partial results) as symbolic content for the agent to reason over.
This is a newly proposed CMC component. The paper describes the kinds of information it would carry and cites partial analogs in Sigma and Soar, but provides no formal specification, implementation, or empirical validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Proposal to Extend the Common Model of Cognition with Metacognition." pith.science (2026). https://pith.science/paper/SWL456GB

@misc{pith2026250607807,
  author       = {Pith},
  title        = {Pith review of: A Proposal to Extend the Common Model of Cognition with Metacognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SWL456GB}},
  note         = {Machine review of arXiv:2506.07807}
}

read the original abstract

The Common Model of Cognition (CMC) provides an abstract characterization of the structure and processing required by a cognitive architecture for human-like minds. We propose a unified approach to integrating metacognition within the CMC. We propose that metacognition involves reasoning over explicit representations of an agent's cognitive capabilities and processes in working memory. Our proposal exploits the existing cognitive capabilities of the CMC, making minimal extensions in the structure and information available within working memory. We provide examples of metacognition within our proposal.

Figures

Figures reproduced from arXiv: 2506.07807 by the authors.

**Figure 1.** Alternative Metacognitive Architectures. state of cognitive processing, the histories of processing, and representations of the agent’s procedural knowledge that drives cognition, with reasoning and metareasoning operating in parallel without intermixing. In the CMC and cognitive architectures more generally, a capability is realized through architectural structures and knowledge. Therefore, we propose a unified ap… view at source ↗

**Figure 2.** The Common Model of Cognition. Below, we review the Common Model of Cognition and present our proposal for extending it to include human-like metacognition. Our proposal focuses on adding new representational distinctions and sources of information about cognition that become available to an agent to initiate, reason, and terminate metacognition. We attempt to identify the minimum architectural information required… view at source ↗

**Figure 3.** Structural extensions to the CMC to support Metacognition. 4 Proposal for Metacognition in the CMC The crux of our proposal is to expand the sources and representations of an agent’s available information to meet the requirements described above. We present our proposal in two stages. First, we describe structural modifications to the CMC that make new representations and direct sources of information available. In … view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Nested Episodic State Topology (NEST): A Graph-Theoretic Architecture of Cognitive States
cs.HC 2026-07 conditional novelty 6.0 of 10

NEST defines a nested, typed graph ontology for cognitive states and maps existing cognitive architectures as constrained regions of this single graph-theoretic language.
Mapping Neural Theories of Consciousness onto the Common Model of Cognition
q-bio.NC 2025-06 conditional novelty 6.0 of 10

The paper shows that GNWT, IIT, RPT, and PP/NREP can each be interpreted in terms of the Common Model of Cognition's working memory and cognitive cycle, suggesting a shared computational substrate.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · cited by 2 Pith papers

[1]

Psychological Review111(4), 1036–1060 (2004)

Anderson, J.R., Bothell, D., Byrne, M.D., Douglass, S.A., Lebiere, C., Qin, Y.: An integrated theory of the mind. Psychological Review111(4), 1036–1060 (2004)

work page 2004
[2]

Proceedings of the AAAI Conference on Artificial Intelligence30(1) (Mar 2016)

Cox, M., Alavi, Z., Dannenhauer, D., Eyorokon, V., Munoz-Avila, H., Perlis, D.: MIDCA: A Metacognitive, Integrated Dual-Cycle Architecture for Self-Regulated Autonomy. Proceedings of the AAAI Conference on Artificial Intelligence30(1) (Mar 2016)

work page 2016
[3]

Cox, M., Raja, A.: Metareasoning: An Introduction, pp. 3–14. MIT Press (2011)

work page 2011
[4]

American Psychologist34(10), 906–911 (1979)

Flavell, J.H.: Metacognition and cognitive monitoring: A new area of cogni- tive–developmental inquiry. American Psychologist34(10), 906–911 (1979)

work page 1979
[5]

Johnson, S.G.B., Karimi, A.H., Bengio, Y., Chater, N., Gerstenberg, T., Larson, K., Levine, S., Mitchell, M., Rahwan, I., Schölkopf, B., Grossmann, I.: Imag- ining and building wise machines: The centrality of AI metacognition (2025), https://arxiv.org/abs/2411.02478

work page arXiv 2025
[6]

Artificial Intelligence Review53(1), 17–94 (Jan 2020)

Kotseruba, I., Tsotsos, J.K.: 40 years of cognitive architectures: Core cognitive abilities and practical applications. Artificial Intelligence Review53(1), 17–94 (Jan 2020)

work page 2020
[7]

Procedia Computer Science145, 730–739 (Jan 2018)

Kralik, J.D., Lee, J.H., Rosenbloom, P.S., Jackson, P.C., Epstein, S.L., Romero, O.J., Sanz, R., Larue, O., Schmidtke, H.R., Lee, S.W., McGreggor, K.: Metacogni- tion for a Common Model of Cognition. Procedia Computer Science145, 730–739 (Jan 2018)

work page 2018
[8]

MIT Press, Cambridge, MA (2012)

Laird, J.E.: The Soar Cognitive Architecture. MIT Press, Cambridge, MA (2012)

work page 2012

Show all 17 references

[9]

AI Magazine38(4), 13–26 (2017)

Laird, J.E., Lebiere, C., Rosenbloom, P.S.: A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Sci- ence, Neuroscience, and Robotics. AI Magazine38(4), 13–26 (2017)

2017
[10]

In: Proc

Laird, J.E., Mohan, S.: Learning Fast and Slow. In: Proc. of the 32nd AAAI Con- ference on Artificial Intelligence. p. 5. AAAI Press, New Orleans (2018)

2018
[11]

In: Bower, G

Nelson, T.O., Narens, L.: Metamemory: A theoretical framework and new find- ings. In: Bower, G. (ed.) The Psychology of Learning and Motivation: Advances in Research and Theory, pp. 125–173. Academic Press, New York (1990)

1990
[12]

Nolte, R., Pomarlan, M., Janssen, A., Beßler, D., Javanmardi, K., Jonge- bloed, S., Porzel, R., Bateman, J., Beetz, M., Malaka, R.: How metacogni- tive architectures remember their own thoughts: A systematic review (2025), https://arxiv.org/abs/2503.13467

2025 arXiv
[13]

Perspectives on Psychological Science17(6), 1746–1765 (2022)

Rahnev, D., Balsdon, T., Charles, L., de Gardelle, V., Denison, R., Desender, K., Faivre, N., Filevich, E., Fleming, S.M., Jehee, J., Lau, H., Lee, A.L.F., Locke, S.M., Mamassian, P., Odegaard, B., Peters, M., Reyes, G., Rouault, M., Sackur, J., Samaha, J., Sergent, C., Sherma...

2022
[14]

Journal of Artificial General Intelligence 7(1), 1–103 (Dec 2016)

Rosenbloom, P.S., Demski, A., Ustun, V.: The Sigma Cognitive Architecture and System: Towards Functionally Elegant Grand Unification. Journal of Artificial General Intelligence 7(1), 1–103 (Dec 2016)

2016
[15]

In: 22nd International Conference on Cognitive Modeling, ICCM 2024

Rosenbloom, P.S., Laird, J.E., Lebiere, C., Stocco, A., Granger, R., Huyck, C.: A proposal for extending the Common Model of Cognition to emotion. In: 22nd International Conference on Cognitive Modeling, ICCM 2024. Tilburg University, the Netherlands (Apr 2024)

2024
[16]

Journal of Experimental & Theoretical Artificial Intelligence19(2), 159–193 (2007) 12 Laird, Lebiere, Rosenbloom, Stocco

Sun, R.: The importance of cognitive architectures: an analysis based on CLAR- ION. Journal of Experimental & Theoretical Artificial Intelligence19(2), 159–193 (2007) 12 Laird, Lebiere, Rosenbloom, Stocco

2007
[17]

Technologies13 (2025)

Walker,P.,Haase,J.,Mehalick,M.,Steele,C.,Russell,D.,Davidson,I.:Harnessing metacognition for safe and responsible ai. Technologies13 (2025)

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Psychological Review111(4), 1036–1060 (2004)

Anderson, J.R., Bothell, D., Byrne, M.D., Douglass, S.A., Lebiere, C., Qin, Y.: An integrated theory of the mind. Psychological Review111(4), 1036–1060 (2004)

work page 2004

[2] [2]

Proceedings of the AAAI Conference on Artificial Intelligence30(1) (Mar 2016)

Cox, M., Alavi, Z., Dannenhauer, D., Eyorokon, V., Munoz-Avila, H., Perlis, D.: MIDCA: A Metacognitive, Integrated Dual-Cycle Architecture for Self-Regulated Autonomy. Proceedings of the AAAI Conference on Artificial Intelligence30(1) (Mar 2016)

work page 2016

[3] [3]

Cox, M., Raja, A.: Metareasoning: An Introduction, pp. 3–14. MIT Press (2011)

work page 2011

[4] [4]

American Psychologist34(10), 906–911 (1979)

Flavell, J.H.: Metacognition and cognitive monitoring: A new area of cogni- tive–developmental inquiry. American Psychologist34(10), 906–911 (1979)

work page 1979

[5] [5]

Johnson, S.G.B., Karimi, A.H., Bengio, Y., Chater, N., Gerstenberg, T., Larson, K., Levine, S., Mitchell, M., Rahwan, I., Schölkopf, B., Grossmann, I.: Imag- ining and building wise machines: The centrality of AI metacognition (2025), https://arxiv.org/abs/2411.02478

work page arXiv 2025

[6] [6]

Artificial Intelligence Review53(1), 17–94 (Jan 2020)

Kotseruba, I., Tsotsos, J.K.: 40 years of cognitive architectures: Core cognitive abilities and practical applications. Artificial Intelligence Review53(1), 17–94 (Jan 2020)

work page 2020

[7] [7]

Procedia Computer Science145, 730–739 (Jan 2018)

Kralik, J.D., Lee, J.H., Rosenbloom, P.S., Jackson, P.C., Epstein, S.L., Romero, O.J., Sanz, R., Larue, O., Schmidtke, H.R., Lee, S.W., McGreggor, K.: Metacogni- tion for a Common Model of Cognition. Procedia Computer Science145, 730–739 (Jan 2018)

work page 2018

[8] [8]

MIT Press, Cambridge, MA (2012)

Laird, J.E.: The Soar Cognitive Architecture. MIT Press, Cambridge, MA (2012)

work page 2012